A Semantic Feature-Driven Multiscale Grid Dehazing Method
Hao Zhang, Hongmei Jin, Zhanli Li, Ning He · 2024
Aiming at the problems of incomplete dehazing and detail distortion in the dehazing process of existing algorithms, an end-to-end dehazing method SemanticGridDehazeNet is proposed, which is semantic feature-driven and combines multi-scale feature fusion. The method consists of three modules: preprocessing, backbone, and post-processing. First, preprocessing is performed using a newly designed residual dense fusion block (RDFB) and convolutional layer. The Backbone module is based on the extended GridN et, which contains three processing scales, and semantic features extracted based on VGG16 Net are added in front of each scale to enhance learning. Each scale contains five RDFB blocks to fuse the feature map to the last column. Then an RDFB post-processing module improves the output quality. The main innovations include: (1) the introduction of the RDFB fusion module to focus on haze features, dynamically adjust the level of attention paid to channel or pixel locations to more effectively learn and utilize key information about the haze distribution in the input data; (2) Semantic features are integrated into the image dehazing process by incorporating semantic features extracted from VGG16 Net before each processing scale. This enhances multi-scale learning and enables more accurate focus on haze features, improving dehazing effectiveness. Experimental results demonstrate superior performance in target enhancement, robustness, and generalization. The method effectively removes residual hzae and addresses detail distortion issues.