Attention-Enhanced Feature Extraction for Image Defogging Algorithms

The AOD-Net suffers from issues such as low image quality and color distortion after defogging. An improved defogging algorithm is proposed, which applies pooling operations and dilated convolution to the convolutional layers of AOD-Net to enhance feature extraction. Multi-scale feature fusion is used, and the extracted information features are fed back to improve propagation, accelerate model training, and enhance the model's generalization ability. Additionally, an ECA attention module is introduced, allowing the network to dynamically adjust the focus on different feature channels, enhancing the network's feature modeling ability and suppressing redundant information, thus improving the defogging quality of the model. Experiments are conducted using the public dataset NYU2 and existing lightweight defogging algorithms for comparison. The defogging quality has improved compared to defogging algorithms such as DCP and CAP. Compared with AOD-Net, the peak signal-to-noise ratio of this method has increased by 4.037 dB, and the structural similarity has reached 0.918, effectively enhancing the network's defogging capability.