Single Image Dehazing Based on Contrastive Learning and Transformer
Yang Zhao, Yigang Wang · Journal of Physics Conference Series · 2023
Abstract For the single image dehazing problem, an end-to-end multi-stage dehazing algorithm is designed. The algorithm contains two distinct parts to extract features. For shallow features, texture-level information is mined by stacking pixel and channel attention mechanisms. The proposed method uses multi-head self-attention (MHSA) to capture high-level features. MHSA improves dehazing performance by mining the dependencies of a wide range of abstract information. The superiority of the transformer architecture is extended with cascaded attention mechanisms and convolutions to improve feature extraction capabilities. Multilayer perceptron (MLP) is used in the decoding stage to equalize the context information. Furthermore, a contrastive loss function that introduces multiple negative samples and correction terms is proposed. The correction term is generated according to the difference between the precise and blurred images, which can enhance the training effect when dealing with different concentrations of dehaze. The training result of this loss function assists the model in approximating clear images and staying away from blurry images. According to the experimental results compared with other methods under the same conditions, the proposed method achieves good results in both subjective visual effects and objective evaluation indicators. The proposed contrast loss function also improves the dehazing performance of the algorithm.