Exploring Shadow Removal in Image Processing Pipelines

Figure 1: Proposed pipeline and result comparison

This project explores shadow removal as a crucial image preprocessing technique to enhance classification performance, with a primary focus on the challenges of shadows present in computer vision applications. Shadows can significantly impact object detection and render systems vulnerable to adversarial attacks, particularly in safety-critical domains like autonomous driving and medical imaging. To address this issue, our work utilizes a comprehensive traffic sign dataset, the German Traffic Sign Benchmark (GTSRB). We first trained a ResNet classifier on this dataset and then applied shadow adversarial attacks on the images. We then used an unsupervised domain-classifier guided shadow removal network known as DC-ShadowNet, which introduces novel loss functions grounded in physics-based shadow-free chromaticity, shadow-resistant perceptual characteristics, and boundary smoothness. The study delivers interesting numerical results: the ResNet model achieved a remarkable 98.12% classification accuracy on the original GTSRB test set. However, the introduction of shadow attacks to these images led to a notable drop to 93.19%. Strikingly, when shadow removal was applied to these adversarial images, the overall accuracy further decreased to 87.14%, unveiling the gaps in shadow removal algorithms' effectiveness…