Project DetailsThe goal of this project is to develop an automatic method for detecting and filtering images with various degradations, as part of a data quality improvement process. This is a challenging task, as visual degradations can appear in different forms and affect image characteristics in diverse ways. To address this, we constructed a dedicated dataset of aircraft images that were subjected to a variety of artificial transformations simulating common quality degradations. These included noise addition, grayscale conversion, style changes (such as cartoon-like or toy-like appearance), quality reduction, and watermark insertion. To label the dataset, a dedicated annotation tool was developed, enabling an efficient and consistent labeling process. Using this tool, the images were labeled according to the type of degradation applied, while maintaining consistency and minimizing labeling errors. Next, several convolutional neural network models were trained, while exploring multiple backbone architectures. The models were trained to classify images based on the type of degradation and were used to evaluate the ability to automatically filter corrupted data as part of a data cleaning process. Model performance was evaluated on both validation data from the dataset and on external images, in order to assess generalization ability and robustness to variations. Finally, heatmaps were used to analyze the decision-making process of the models, highlighting the regions in the images that influenced the classification.
