SIPL Projects

In this project, we developed an automated system for detecting and localizing Ziziphus pollen grains in microscopic images of honey to determine their relative proportion among all grains in the image. This allows us to deduce the primary composition of each honey sample and decide on its botanical classification as monofloral or polyfloral. Following an extensive literature review, we selected the YOLOv8 architecture to implement our solution. Various experiments were conducted, primarily focused on data preprocessing and initial training phases, to enhance model performance.

This project, in cooperation with the Faculty of Mechanical Engineering, aims to create a device which can be thrown into a room and give 360 degrees of live camera feed, displayed on a tablet, to an operator outside. Possible uses of such a device include disaster response (being able to see into areas blocked off from easy access to find people who need assistance), fire rescue, and similar use-cases. Requirements for this device include long battery lifespan and reliability for repeat use, but also low cost because it is expected to occasionally be lost (for example in disaster response), or damaged (for example in a housefire).

This work focuses on developing a Vision-Language Model (VLM) for identifying the hyoid bone in CBCT images and analysing its position relative to the mandibular plane. CBCT scans are commonly used in dental clinics but have not been fully utilized for diagnosing Obstructive Sleep Apnea (OSA). The project's goal is to develop an AI-assisted diagnostic tool that provides automatic textual descriptions of the hyoid bone's position, assisting clinicians in early detection. We adapted the MyVLM (Alaluf et al., 2024) architecture for medical applications, replacing key components with Medical CLIP and LLaVA-Med, which are specifically trained for analysing anatomical structures.

Public spaces, particularly parks, are a vital and significant component of urban life. Due to population growth and urban development, especially in the wake of the COVID-19 pandemic, optimal planning and utilization of available public spaces have become critical for improving quality of life. Technological advancements in artificial intelligence and computer vision present an opportunity to optimize and enhance intelligent space planning through in-depth analysis of statistics such as the manner, frequency, type, and duration of public space usage, as well as population distribution and other characteristics.

Today, GPS and location information is widely used and essential for many means of transportation and varying tasks, creating the need for reliable alternatives for self-localization when such signals are obstructed or limited. The task of coarsely estimating the place of a photo (Visual Geo-Localization) is a rapid growing field of research, offering an alternative method for estimating location. It is addressed using image matching methods on a database of images of known locations. In our project, we explore geo-localizing helicopters using satellite images, by modifying state-of-the-art practices from ground-to-ground and aerial-to-ground matching problems.

In this work, we introduce EMD-Based Hyperbolic Diffusion Distance (EMD-HDD), a new method for constructing a meaningful distance metric for hierarchical data with latent hierarchical structure. Our method relies on hyperbolic geometry, diffusion geometry, and the Earth Mover’s Distance (EMD). Specifically, our method embeds data points into a product manifold of hyperbolic spaces, allowing us to recover the hidden hierarchical structure encoded by the mutual relationships between features.

This work outlines Team SIPL’s solution to the 2025 Signal Processing Cup challenge, which focuses on the task of distinguishing real images from Deepfakes. Recent Deepfake image detection techniques rely on deep neural networks but face challenges related to limited generalization capabilities. Our proposed solution addresses these challenges by employing an ensemble of state-of-the-art methods. At the core of our approach is a neural network trained to intelligently combine embeddings from these individual detectors. This fusion enables the ensemble to effectively integrate complementary features, enhancing its ability to identify subtle artifacts.

The goal of this work was to analyze images that were extracted from flexible laryngoscopy videos using classical image processing methods, in order to improve the diagnostic accuracy in vocal fold examinations.   Flexible laryngoscopy is a vital tool used by ENT (ear, nose, and throat) specialists to diagnose disorders and diseases of the vocal folds. However, interpreting the resulting videos is not a simple task. Unlike static medical images (such as X-rays), laryngoscopy videos present unstable visual conditions: Light reflections from the moist tissues, noise due to the camera quality, and rapidly changing tissue appearances, caused by motion and deformation of the vocal folds.

This project aimed to investigate the relationship between neural activity, as measured by functional magnetic resonance imaging (fMRI), and children’s life experiences. Specifically, the study examined whether there are correlates between children’s fMRI brain scans and their “Life Event Scores”, which measure the accumulation of stressful life events. The findings of the project may contribute to a deeper understanding of the neural mechanisms underlying the effects of childhood adversity on brain development and may help inform the design of more targeted interventions that support the mental health and wellbeing of children exposed to significant stressors.

In this project, an attempt was made to reconstruct the historical photographic technique known as the Sodium Vapor Process, using readily available components at a total cost of less than $2500. The method relies on simultaneous imaging of a scene illuminated by RGB LED lights, together with a background illuminated by a low-pressure sodium (LPS) lamp emitting at a single wavelength of 589.6 nm. The setup employs two cameras and a beam splitter: one positioned behind an optical band-pass (BP) filter, and the other behind an optical notch filter. A system prototype was successfully assembled.

The goal of the work is to analyze histopathology images of the thyroid gland using deep learning algorithms. The algorithm is designed to reduce the workload of pathologists in identifying regions of interest in the biopsys samples in order to minimize human errors in the identification process. We chose to perform the segmentation using the U-Net network. Following the classification, we performed feature detection to classify the subtype of thyroid cancer using the StarDist network. We conducted our tests using datasets from two hospitals, NKI in the Netherlands and VGH in Japan, for the U-Net algorithm, and the MoNuSeg dataset for the StarDist network.

The BMI index is a crucial index which gives a quantitative assessment of whether a person is in normal weight, underweight or overweight. The index is calculated using height and weight data. The purpose of our project is to estimate a person's BMI from a single 2-dimensional image. This is a complex task because visual inspection of the image is not sensitive to the distance of the object from the camera and the angle of the shot. To approach this task, we relied on previous works in the field, on their results, and on the dataset published with them.

Many studies in the field of child development have shown that the diagnosis of developmental problems in children at a young age significantly improves the ability to treat these problems (about 80-70 percent success rate). In addition, studies have shown that the nature of interaction between a child and their parents can indicate developmental problems. Building upon this premise, this project aims to classify joint activities between a child and parent through visual analysis of shared interaction videos. This system will ultimately contribute to diagnosing developmental problems by cross-referencing its output with additional data (EEG signals, etc.).

This work’s goal is to enable maneuvering abilities for a four-legged robot in an indoor environment by employing semantic segmentation. The segmentation is performed using deep learning, based on low-resolution grayscale and depth images captured by a Pico Flexx camera mounted atop the robot. While most existing semantic segmentation methods rely on RGB and depth images, there are no pre-trained models specifically designed for grayscale images. In the project, we adapted an architecture intended for semantic segmentation using RGB and depth images, leveraging transfer learning to tailor it to our specific requirements.

Table tennis is a fast game that requires quick reactions and great hand-eye coordination. Many of the points that are scored during the game are a direct result of a good serve which relies heavily on the strategic use of spin. This project aims to leverage smartphone slow-motion cameras and image processing techniques to analyze and quantify the spin of a table tennis ball during service. We use YOLO v7 for object detection on slow-motion videos in which the ball is uniquely patterned for effective feature detection. Additionally, we apply BRIEF feature descriptor and brute-force matching algorithm which we use to extract matching features across frames.

This work aims to address the pressing need for high-quality public open spaces in urban environments, with a focus on leveraging computer vision and deep learning techniques. The COVID-19 pandemic has emphasized the importance of public open spaces in enhancing the well-being and quality of life for city dwellers. It has become evident that these spaces serve as vital elements in urban landscapes and play a significant role in promoting physical and mental health, social interactions, and overall community resilience.

In recent years, AI-driven image editing has emerged as a promising field with numerous applications. This work explores the capabilities of the generative AI model, Diffusion, for image editing guided by reference images. We focus on leveraging a set of images that outline the desired editing features and applying them to a target image. By experimenting with various hyper-parameters, modifying the core components of the Diffusion model, and integrating the CLIP model, we demonstrate various improvements in image editing performance.

Joseph Yerushalmi, a librarian at the University of Haifa Library, created a catalogue with around 65,000 records on paper cards. The catalogue contains articles from the 1940s to the 1970s, focusing on individuals like artists, writers, philosophers, intellectuals, and historical figures. the collection also includes reviews on books and literary works.   To preserve this valuable catalogue, digitization is needed, the project is divided to two parts: The first part is to Detect text regions, which means classifying each region to its appropriate label: Title, Author, Text, and other.

Magnetic Resonance Imaging (MRI) is a critical imaging modality in modern medicine, providing high-resolution images of soft tissues without the use of ionizing radiation. However, the high cost and complexity of traditional high-field MRI scanners have driven interest in low-field MRI systems. While more affordable and portable, low-field MRI suffers from several drawbacks, including lower signal-to-noise ratio (SNR), reduced resolution, and poorer contrast, which limits its clinical utility.

Cancer is a leading cause of death worldwide, responsible for around ten million deaths in 2020. Cancer cells exhibit invasive behavior, which significantly contributes to tumor growth and metastasis. The invasiveness quality of cancer cells is a principal contributor to these deaths by influencing tumor growth and metastasis. Current methods for evaluating this include seeding cells on bio-gel, capturing images and semi-manually labeling the positions and depths of cells, which are both time-consuming and prone to human error. This study introduces an automated approach for estimating cell position and depth using Differential Interference Contrast (DIC) and Surface images.

The traditional image signal processing (ISP) pipeline involves a series of sequential image processing modules within a camera that transform raw sensor data into a high-quality sRGB image. Recently, methods have been developed to enhance traditional ISP performance using convolutional neural networks (CNNs). However, these approaches typically train a CNN to handle ISP tasks without adequately considering the interconnections among different ISP components. Consequently, the image quality in challenging scenarios, like low-light conditions, remains subpar.

In this work we explore the different noise mechanisms in event cameras and discuss known methods for filtering several noise sources. We focus on the threshold mismatch effect in event cameras and introduce a novel correction scheme to reduce the effect of threshold mismatch. To the best of our knowledge this is the first algorithmic solution to the threshold mismatch effect. We use a threshold estimation algorithm suggested by Ziwei Wang et al., and further explore it in simulations with knowledge of the thresholds to calculate estimation errors. Using the estimated thresholds, we apply a correction algorithm based on sampling theory.

This work presents a novel approach for monitoring 'beach scenes', utilizing existing beach webcams and computer vision algorithms based on deep neural networks to analyze the behavior of water bodies, humans, and the interactions between them. We used surf condition analysis as a study case for our method. The proposed approach detects and tracks both people and waves simultaneously, distinguishing between waiting surfers and riding surfers. The system uses a Faster-RCNN object detector for detection and classification and SORT, a fast multi-object tracking algorithm based on the Kalman filter, for tracking.

A new method to predict the development of metastases of cancer cells and accelerate treatment for patients was developed in Prof. Daphne Weihs' lab. The method works by placing cancer cells samples on gel and finding their penetrating level on it. According to the penetration level of the cell, one can predict the probability it will develop metastases. In a previous project, an algorithm was developed to automatically detect cells' locations on a DIC image (image of the cells on the gel's surface). Our first goal in this project was to improve the algorithm and reach higher accuracy. To manage this task, I added morphological operations and based on deep-learning ideas.

Spreading of cancer cells to distant organs is the main cause of all cancer-related mortalities, while early and accurate prediction of tumor cell fate can significantly impact treatment protocols and survival rates. In the Weihs laboratory a phenomenon was observed, where invasive cancer cells can infiltrate the surface of synthetic polyacrylamide gels, while normal cells have minimal to no indentation in the gels surface. Estimating such invasiveness of cells manually by lab expert is time and resource consuming. This project aims to predict the depth of each individual cell based on an image of the gels surface in order to estimate cell invasiveness.

Deep diffractive neural networks have emerged as a promising framework that combines the speed and energy efficiency of optical computing with the power of deep learning. This has opened new possibilities for optical computing suit for machine learning tasks and all-optical sensors. One proposed application of this framework is the design of a diffractive camera that preserves privacy by only imaging target classes while optically erasing all other unwanted classes. In this work, we investigated whether this camera design truly erases the information in unwanted class data. We used K-NN to achieve up to 94% accuracy in classifying optically erased images.

This works main goal is to create a demo system that estimates a subject's heart rate in real time by observing changes in light reflection from his skin. The system must be able to do so using a regular RGB camera. The heart rate estimation is based on a PPG (PhotoPlethysmoGraphy) signal, remotely extracted using the camera and digital signal processing. We have created such a system, which operates on a standard laptop using its built-in webcam or any other external camera connected to it. The median of the mean absolute error across the videos in UBFC-rPPG dataset that was achieved by our system is 3.8%.

Evaluating blood flow in intestinal tissue is important for several medical procedures, such as laparoscopic surgery. However, accurately assessing blood flow using traditional methods, such as Doppler ultrasound, can be challenging. In this project, we propose a new method for evaluating blood flow in intestinal tissue using real-time video from a laparoscopic camera. Our approach involves identifying the Region of Interest (ROI) in the video and then analyzing the video in the color, frequency, and texture spaces to detect temporary changes that may indicate changes in blood flow. We demonstrate the effectiveness of our method through experiments on a dataset of laparoscopic videos.

Classifying dairy cows is a critical operation for dairy farms. The primary goal of dairy farms is to maximize milk production, which is achieved by monitoring various aspects of each cow, including milk yield, health status, estrus time, and other characteristics. Therefore, the foremost objective of a dairy farm is to establish a reliable method for identifying each cow accurately. Currently, common methods for cow identification rely on permanent measures such as ear or back tattooing, as well as the use of ear tags equipped with radio frequency identification (RFID) technology. However, these methods have limitations as they can fade, fall off, or break over time.

This work examines the feasibility of incorporating models based on insect vision systems for motion detection in computer vision. In recent years, significant advancements have been made in both the research of visual processing in animal systems and in the field of artificial neural networks. Despite these advancements, and despite artificial neural networks already drawing inspiration from biology, there remains a disconnect between these two areas of research. Insects have the ability to quickly and accurately respond to visual input while in flight, using less hardware than comparable computing systems, making them a potential source of inspiration for computer vision systems.

Breast cancer is among the leading cause of mortality among women in developing as well as under developing countries. The detection and classification of breast cancer in early stages of its development may allow patients to have proper treatment. The work faced the problem of mass detection in mammograms, using CNN architecture. The initial architecture of the project was given to us by former student subject to Dr. Dror Lederman. Weve worked and streamlined this initial architecture. In this work we had to stabilize the existing model so that it would converge and not diverge. In addition, we had to improve the percentage accuracy of the model.

Melanoma skin cancer is common and deadly, and its diagnosis today requires a manual examination by a dermatologist. Any spot on the patients body marked by the doctor as suspicious should be examined using a dermoscope (polarized light photography), as this is the only way to determine whether the mole is malignant or not. Marpe Technologies develops a system that scans the patients body in visible light using a high-resolution camera, detects suspicious moles, and sends them to examination by a dermatologist with a dermoscope, thus saving the doctors manual examination time. Previous SIPL project aimed at improving this classification achieved a 99.

The goal of this work is to build an algorithm that analyse spatial parameters using simple and inexpensive sensors and a limited calculation cost. The current method of identifying objects in a given space and extracting its active properties requires a large number of advanced sensors and expensive processing power. We suggest the option of reducing the computing power and the number of sensors used while maintaining the output quality. The proposed solution is based on a visual system inspired by flying insects and the algorithm is implemented using signal processing tools.

Early detection of a cancerous tumor is important and may be the decisive factor in the patient's fate, since if the tumor is at the beginning - the treatment is simpler, and the chances of recovery are significantly higher. Due to the great importance of the tests, there is a goal to make these tests more accessible and more accurate so that many people can perform them and get accurate answers, which will save human life. HT Bioimaging is developing a system for detecting and testing cancerous tumors by heating the tumor and sampling the heat dissipation using a thermal camera.

In this work, we demonstrate the use and application of learning hierarchical data by embedding it into hyperbolic space. By applying the latest approaches in the field of machine learning and topological data analysis, specifically hyperbolic representations and persistence diagrams we showed the great potential of this method to classify data. We showed good results in classifying MNIST and Hyperspectral images of low resolution which may come in use for civilian and military applications alike.

Thermal image colorization is a topic that is gaining momentum in the world of artificial intelligence. In recent years, with a significant improvement in the tools, and with the algorithmic development of deep learning, the world of computer vision has managed to achieve impressive achievements in everything related to image processing and analysis. A significant development that has led to the rapid increase in achievements is the development of the CNN called GAN - Generative Adversarial Network. Networks of this type make it possible to produce a new set of information based on the characteristics of the existing information.

Nowadays, the metric used to calculate the statistic distance between different datasets is the FID (Frchet Inception Distance): it uses a pretrained inception network and a divergence very close to the W2 divergence (in this case) to approach the distance between them. We assume that the latent representation of each dataset has a Gaussian distribution. We also assume that the Gaussian distribution is not degenerate: we assume that the covariance matrix is a positive definite matrix: all the principal components are not zeros. To these assumptions, we can add the numerical instability and the impossibility to score a single image.

Prof. Daphne Weihs lab researches the probability of metastasis in cancer cells by checking the level of penetration that these cells exhibit on a gel substrate. This project is aimed at performing an automatic analysis of the gel, photographed using DIC microscopy. Using these images, and additional images of nucleus staining of the cells, we were tasked with creating an algorithm for morphological recognition of the cells by use of segmentation. With this algorithm, the researchers in the lab can learn the number of cells, their areas, and their liveliness automatically, promptly and accurately.

This project focuses on the issue of real image denoising cleaning noise that is created during the process of capturing the image. In the first part of the project, we designed a CNN, based on RIDNet architecture. In the second part of the project we wish to improve the result of the network, on a dataset it hasnt seen before. I.e., improve the generalization capability of the network. To achieve this goal, we have created datasets with synthetic noise, that aspired to be as similar as possible to real noise. In order to train this model, we have created 4 datasets, each of them constructed of different ratio of real noise image and synthetic noise image.

One of the problems in performing signal processing operations on sound clips stems from noise added to the measurement device. Noise can drastically damage the performance of accurate analysis of an audio signal. One method to deal with this problem is to use multimodal observations so that one of the modalities is not affected by the noise of the other. An example is a video source independent of noise added to the audio, thus unaffected by this noise. This way, one can try to extract information lost in the audio due to the noise, using the video. The purpose of this project is to perform spatial and temporal detection and recognition of speech, both in audio and video.

This work deals with the identification of pedestrians approaching a crosswalk when they use a mobile phone. As part of the research work of Michal Derhy from the Faculty of Architecture and Urban Planning, the option of creating a system that alerts distracted people approaching to cross the road was examined. In an experiment conducted by Michal, a system was set up that included an alarm that was activated manually when a person that use the phone approached the crosswalk. The results of the experiment showed that there is a direct effect between the warning to the pedestrians and their alertness to what is happening around the crosswalk.

The main goal of this work is to develop an algorithm which classifies depth of indentation for each shown cell in a given surface image. In the first step, we developed an algorithm that receives surface images and identifies which of the cells penetrated the gel substrate. In second step, we tried to calibrate the pushing depth of cells by a blur map. A close relationship can be found between the penetration depth of cancer cells and their metastatic nature. The deeper the cells penetrate, the more metastatic they are in nature, meaning more dangerous for the patient. Until now, analyzing such images in laboratories was partly manual and partly automatic and took a long time.

The goal of this work is to train a deep neural network so that it can receive an audio signal as input, and output a reconstructed image of the source from which that audio signal was produced. Under the assumption that an audio signal contains spatial properties of the object that produced it, we tried to use an audio classifier to extract these properties and transform them into a feature vector from which we can reconstruct the image from which the source was produced using a deep network with the GAN architecture. This project is a follow-up project with the same goal.

In this work we show a practical solution for image denoising using CNN Autoencoder Neural Network. The network we built is easy to implement and provide relatively high performance when compared to other classic methods like BM3D, and even compared to other, more complex networks. This network is also very flexible and can be adjusted to match different memory capacity of the graphic cards available for the training. We show how we use a relatively simple design and improve it by using custom performance metrics designed to evaluate images, replacing standard layers like MaxPool and UpSampling with convolutional layers, implementing custom loss functions and comparing between them.

In recent years, the field of artificial intelligence and deep learning is gaining a foothold in almost every field of our lives. In the field of medicine in particular, deep learning has a prominent place in everything related to image processing, information analysis and disease diagnosis. Data acquisition requires skilled and experienced personnel, who are required to perform tasks that are partly Sisyphean and tedious. Also, sometimes it is a matter of working with rare diseases, about which not enough information has been gathered to date. Many medical photographs contain various deformations which make it difficult for the algorithms to achieve optimal detection performance.

We suggest a new approach that enables spatial editing and manipulation of images using Generative Adversarial Networks (GANs). Though many tasks have been solved utilizing the powerful abilities of GANs, this is the first time that a spatial control is suggested. This ability is possible thanks to a test-time spatial normalization that uses the trained model as is and does not requires any fine tuning. Therefore our method is significantly fast and does not required further training. We demonstrate the new approach for the task of class hybridization and saliency manipulation.

The use of GAN has drastically affected low-level vision in graphics, particularly in tasks related to image creation and image-to-image translation. Today the training process, despite all the latest developments, is still unstable. Given a semantic segmentation map in which we can separate and look at each pixel in the image and tag it to the relevant class it represents, we can (with the help of GAN) produce images based on this map and hope to reach a more stable model. With the success of GANs we produced segmentation maps. With these maps and with the help of the generative model we can get a semantic understanding of the data set and even create completely new scenes.

...

Melanoma is a type of skin cancer that is both common and dangerous. Today, in order to diagnose skin cancer, a checkup must be performed by a dermatologist. For the doctor, such a procedure takes a long time. Marpe Technologies has developed an advanced scanning system, which includes a camera that scans the entire body of the patient and alerts the physician when it detects suspicious spots. A previous project on the subject was carried out in collaboration with Marpe Technologies in order to create a trained model that classifies images at technician-level as "suspicious moles" or "unsuspected areas".

The goal of this work is to research the capability of generating a completely new image with the same visual content of a single given natural image, by using unsupervised learning of a deep neural network without the use of a GAN. This project is based on the work presented in the paper: "SinGAN: Learning a Generative Model from a Single Natural Image" (Rot-Shahamet al.) Different papers published in the last couple of years have already established the connection between the deep features of classification networks and the semantic content of images, such that we can define the visual content of an image by the statistics of its deep features.

The target of this work is to dperformo automatic analysis to images from prof. Daphna Weihs lab's experiments. It is already known that we can evaluate the probability of development of metastases of cancer cells by their penetration level on gel surface and this evaluation ability is essential to fit the best treatment to sick patients. In this project we use only two images, the first is DIC a microscopic image of the cells and the second is surface that shows light of bids that are in the bottom of the gel. So, if we have a penetrating cell, we will see its region not in focus. Our first task was to do automatic analysis of DIC images in order to find cells locations and count them.

Image obtained through the air-water interface is determined by light refraction. This affects many parameters such as image apparent shape, position, motion, contour, contrast and more. Birds of many species, such as ospreys, kingfishers, and herons, capture fish after observing and aiming from the air and can compensate for these effects and successfully strike their prey. In this project we researched the geometrical distortions caused by refraction of light rays through the air-water interface and expanded the process to a sea surface with waves. Afterwards, we defined the type of distortion caused by the water surface and proposed a way to measure said distortion.

Security and safety problems rank among the most pressing issues of modern times. In the day-to-day life, the army encounters terrorist attacks. To thwart them, it is necessary to monitor the movements of people and their belongings in specific places. Sample videos of intentionally placed objects were provided by the ICT Corps. The projects goal is to provide a tool for locating suspicious objects that were left in the scene and to classify them. The identification and tracking part can be solved by using DeepSORT. After that, an analysis algorithm associates the objects to their owners, and monitors changes in various parameters between frames.

Augmented Reality App for Physics Classrooms, is a work that aims to create an android app that will allow to take a photo of a drawing of a physics system on a blackboard, to create a running simulation of the drawn system. This project uses deep learning tools, classic image processing algorithms and animation programming tools. It is based on previous projects in which a method was developed for detecting, classifying and localizing objects in an image, mapping them, and creating animation. In this project we focused on improving the object detection and recognition accuracy.

In the first part of the project, our focus was on the PyNET network. This network was designed to replace the full ISP pipeline, which is responsible for the conversion of raw information detected by a digital camera sensor (known as a Bayer image or a RAW image) into the color image seen on the screen (of the DSLR camera, of the smartphone, etc.). Specifically, we tested different loss functions in order to improve PyNETs performance. In the second part of the project, we explored additional ways to improve this performance.

This project's main goal is to exam and study the Intel Realsense Lidar Camera, and do to object detection and classification of in-door images taken by L515. The secondary objective of the project is to create a small database with those images and to label them. The first step in the project was to find a deep learning network and the large 3D in-door dataset it was trained on. The second step was to create a small database with labeled images taken by L515 camera, and the third and final step was to input those images to the chosen network, do adjustments, and to achieve the project's goal: do objects detection and classification.

In this work, two architectures of Optical Character Recognition (OCR) systems were demonstrated for the Solitreo font of the Hebrew Language. The first architecture demonstrated was based upon text detection and classification. The second architecture demonstrated was based upon cropping a document into separate text rows, and whole row translation using an LSTM network. In addition, handwritten text document processing algorithms were also demonstrated, such as: Binarization, Connected Components Analysis, Text Row Detections, and more.

Cancer is a major challenge to modern medicine, the disease has many victims everywhere in the world, therefore multiple efforts and resources are invested in the attempt for cancer annihilation. As part of the characterization of diseases in general and especially cancer, early detection has the potential to increase the patient's chances of recovery. The primary goal of the project - early identification of external cancer (tongue/ cheek /Lip) using cooling and heating patterns of these biological tissues.

The project, Physics Classroom Augmented Reality with Your Smartphone, is the second project having the same goal as the previous one creating an android app that will allow, using a photo of a drawing of a physical system, to create a running simulation of the said physical system. This project uses classic image processing algorithms and animation programming tools. The project is based on a previous one that detects, classifies and localizes objects in an image. The first stage of the project was to create an application for presenting a simple animation of a physical interaction.

The project Deep Learning for Classroom Augmented Reality Android App is a second project having the same goal as the previous one creating an android app that will allow, using an image of a drawing of a physical system, to create a running simulation of the said physical system. The goal of this project, similar to that of the previous project didnt succeed, and as part of the overall solution, is to classify and localize different objects in the drawing of the physical system. Our project tries (and usually succeeds) to do so using deep learning algorithms, as opposed to the previous project that tried and hasnt managed to do so using classic image processing algorithms.

Crossing a road is a dangerous activity for pedestrians and therefore pedestrian crossings and intersections often include pedestrian-directed traffic lights. These traffic lights may be accompanied by audio signals to aid the visually impaired. In many cases, when such an audio signal is not available, a visually impaired pedestrian cannot cross the road without help. In this project, we propose a technique that may help visually impaired people by detecting pedestrian traffic lights and their state (walk/dont walk) from video taken with a mobile phone camera.

In recent years, the use of deep features as an image perceptual descriptor is very popular, mainly for measuring the perceptual similarity between two images. In the field of image restoration, this has proved to be very useful for tasks such as super-resolution and style transfer. In this project, we suggest a different direction: rather than using deep features as a similarity measure, we suggest using them to construct a natural image prior. This can be done by learning the statistics of natural image's deep features. Using this special prior, we can gain from both world: the deep one, and the "classic" one.

During communication between two people who are located at a distance from each other, often arises the need to indicate a point of interest in the individuals shared field of view. Such a situation causes a problem since it is not always possible to describe where the point of interest is in the other persons field of view, since they see the scene from a different perspective. For example, when describing a window on a building or a rock in a forest, one must use the surrounding environment in order to convey the location of the point of interest, which creates a communication barrier. In this project, we surveyed the possibility of creating the basis of a Visual Social Network.

Much recent advancement in Computer Vision is attributed to large datasets and the ability to use them to train deep neural networks. In 2016 Google announced the publishing of a public dataset containing about 8-million tagged videos called YouTube-8M. In this project, we used this database to train several deep neural networks for tagging videos in a variety of categories. In the first stage, we downloaded 5000 videos for 5 different categories. Next, we trained two deep networks, with slightly different architectures, to tag a video into one of the five categories. One network uses the LSTM architecture and the other uses the BiLSTM architecture.

The composition of a promotional brochure is the characteristics and relations between the different objects within it, among those their location, color and shape. Composition has a crucial importance to the degree of success in which the brochure manages to convey the desired message. Despite of the great importance of composition to the success of the brochure, the designer is tasked with designing it and matching it to the brochure's topic, without a convenient tool that allows using generic compositions that were found suitable for certain topics.

In this work, we address the problem of Cumulonimbus (Cb) cloud detection from Infrared (IR) satellite images. The Cb Clouds are associated with thunderstorms and atmospheric instability and their detection is of high importance since they pose extreme danger to aviation. We present a joint spatio-temporal detection method that exploits the distinct spatial characteristics of Cb clouds as well as their prototypical evolution over time. The presented method is unsupervised and does not require labeled data or predefined spatial handcrafted features, such as particular shapes, temperatures, textures, and gradients.

This project is a part of a series of projects carried out in SIPL dedicated to creating an Android application that will assist the visually impaired people with pedestrian traffic lights. The current project consists of two parts: 1. Recognition of pedestrian traffic lights in a single image taken with a mobile phone from a pedestrian perspective. We use the Faster RCNN object detector with transfer learning on more than 900 pedestrian traffic light images, and achieve 98% accuracy. 2. Using the recognition module from part 1 along with object tracking to detect light switches from red to green or vice versa, for improved recognition robustness. For this aim, we use the KCF object tracker.

This project deals with the detection of anomalies in seabed surveys collected using multibeam sonar. Detecting anomalies is a common task in the world of signal processing in general and in image processing in particular. Therefore, there are established methods and methodologies for dealing with such problems in these worlds. As oppose to typical data in these fields, information collected by the multibeam sonar presents various problems that are unique - sampling on irregular grids, variability of the sampling nature depending on the position of the boat and the depth of the soil and areas lacking samples.

Dyslexia is a learning disorder characterized by difficulties with accurate or fluent word recognition and by poor spelling and decoding abilities. Current diagnosis of dyslexia lacks objective criteria, which can decrease treatment efficacy. Diagnosis relies on a discrepancy between reading ability and intelligence, a measure which can be unreliable, and has been criticized for its poor validity. Functional magnetic resonance imaging (fMRI) is a fairly new and unique tool that enables widespread, noninvasive investigation of brain functions.

When using the smartphone's camera to take a photo, usually unwanted objects enter the frame, such as unwanted cars on the road, people who walk in the background of the frame, etc. Another example is when using surveillance cameras, sometimes an image free of unwanted objects is desired. The project's goal is to deal with those situations in a way that allows the user to select the unwanted object that he wants to remove from a list of unwanted objects, and remove him in a way that the background of the objects is completed from another frame.

We describes here a JPEG based compression scheme, adjusted specific for the underwater acoustical channel. The project goal was to deal with common bit error rates for the underwater acoustic channel. Tow measures were used to quantify different solutions: compression ratio, which we tried to minimize, and SSIM, a measure that describes the quality of the image, which we tried to maximize. The scheme described in this report was inspired by initial research conducted by RAFAEL, common schemes to deal with errors in images, and intensive acquaintance with the international JPEG compression standard.

High Efficiency Video Coding (HEVC) is a new video coding standard that has recently been finalized. Due to its substantially improved performance, it is expected to replace the H.264 video coding standard and to become the most common video coding technique in few years. A major innovation in HEVC is the use of a quad-tree based coding tree block for images. In this representation, an image is first divided into non-overlapping coding units, which can be recursively divided into smaller coding units. This recursive quad-tree decomposition of the image is an efficient representation of variable block sizes, so that regions of different sizes can be better coded.

This projects goal is estimation of the distance of floating objects, such as boats and personal water craft (water scooters) from a video of maritime environment for the Protector USV, which is a product of Rafael. We propose a novel and efficient algorithm to achieve this goal. The algorithm receives as input a video of a marine environment. In addition, the algorithm receives as input for every video frame the location of a pixel that is on or near the object of interest, which we want to estimate the distance to.

Capsule endoscopy is a method for recording images of the digestive tract. A patient swallows a capsule containing a tiny camera, which captures images that are then transmitted wirelessly to an external receiver for examination by a physician. Due to limited computational capabilities in the capsule and bandwidth constraints derives from dimensions of capsule, low-complexity and efficient compression of the images is required before transmission. In addition, the images are captured using a Bayer filter mosaic, such that each pixel in raw captured images represents only one color: red, green or blue.

Parkinson's Disease (PD) is a degenerative disease of the central nervous system with a profound effect on the motor system. Symptoms include slowness of movement, rigidity of motion and in some patients, tremor. The severity of the disease is quantified using the Unified Parkinson Disease Rating Scale (UPDRS) which is a subjective scale performed and scored by physicians. In this work, we present an automated, objective quantitative analysis of four UPDRS motor examinations of Hand Movement and Finger Taps. For this purpose, a non-invasive system for recording and analysis of fine motor skills of hands was developed.

Video quality assessment becomes increasingly important nowadays. Therefore, it is desirable to develop a visual quality metric that correlates well with human visual perception. In previous projects in SIPL, a new technique for quality assessment has been developed, based on DCT sub-bands similarity (DSS). The new technique shows excellent results in comparison to subjective results. Its low complexity makes it highly suitable for a variety of live-streaming applications. Yet, some changes have to be made. The main goal of this project is to build up a prototype of a video quality assessment system, based on network similar to the IDF network, while using the DSS method.

Created in an undergraduate project in the Signal and Image Processing Laboratory (SIPL), Department of Electrical Engineering, Technion Israel Institute of Technology. We have built a standalone, robust and portable platform that enables an augmented reality pinball game based on virtual and real objects and using common hardware.

...

Sensing in the Shortwave Infrared (SWIR) range has only recently been made practical. The SWIR band has an important advantage it is not visible to the human eye, but since it is reflective, shows shadows and contrast in its imagery. Moreover, SWIR sensors are highly tolerant to challenging atmospheric conditions such as fog and smoke. They can be made extremely sensitive, thus can work in very dark conditions. However, fundamental differences exist in the appearance between images sensed in visible and SWIR bands. In particular, human faces in SWIR images do not match human intuition and make it difficult to recognize familiar faces by looking at such images.

...

...

...

...

...

...

...

...

...

...

...

...

...

...

...

...

...

...

...

...

...

...

...

...

...

...

...

...

...

...

...

...

...

...

...

...

...

...

...