Dehazing Network Based on Combined Loss and Fusion Attention Mechanism

The image defogging model based on neural network deep learning mainly improves the performance of the algorithm by adding an attention mechanism and increasing the depth of the network, which often requires a huge data set for support during the training process, but this requires more computational resources. In order to better solve the problem of unsatisfactory image defogging effect and time consumption caused by huge computation, a lightweight and efficient image defogging model is proposed. First, ODConv (omni-dimensional dynamic convolution) is introduced instead of the traditional convolution (CNN) to improve the feature extraction effect of hazy images through multi-dimensional feature attention; then, the use of the pixel attention mechanism (PA) enables the model to pay more attention to the areas of the image that are seriously affected by haze or rich in details; finally, a combined loss function is used to improve the feature extraction effect of hazy images. regions; finally, a combined loss function is used for training to balance the pixel-level similarity (via L1 loss) and the similarity of high-level feature representations (via perceptual loss).