Mask-Guided Joint Single Image Specular Highlight Detection and Removal

Detecting and removing specular highlights is a highly challenging task that can effectively enhance the performance of other computer vision tasks in real-world scenarios. Traditional highlight removal algorithms struggle to accurately distinguish the differences between pure white or nearly white materials and highlights, while recent deep learning-based highlight removal algorithms suffer from complex network architectures, lack of flexibility, and limited object adaptability. To address these issues, we propose an end-to-end framework for single-image highlight detection and removal that utilizes mask guidance. Our framework adopts an encoder-decoder structure, with EfficientNet serving as the backbone network for feature extraction in the encoder. The decoder gradually restores the feature map to its original size through upsampling. In the highlight detection module, the network layers are deepened, and residual modules are introduced to extract more feature information and improve detection accuracy. In the highlight removal module, we introduce the Convolutional Block Attention Module, which dynamically learns the importance of each channel and spatial position in the input feature map. This helps the model better distinguish foreground and background and improve adaptability and accuracy in complex scenes.Experimental results demonstrate that our proposed method outperforms existing methods, as evaluated on the public SHIQ dataset through comparative experiments. Our method achieves superior performance in highlight detection and removal.