Multi-Patch and Feature Fusion Network for Single Image Dehazing

Abstract

The dehazing method based on deep learning has made significant progress in the field of image dehazing, but most methods still have the problem of incomplete dehazing and color distortion. To solve this problem, an image dehazing network based on multi-patch and feature fusion is proposed. The network consists of preprocessing, feature extraction, feature fusion and post-processing modules. The preprocessing module can adaptively extract image feature information from the patch. The feature extraction module uses cascaded dense residual blocks to extract deep feature information. The feature fusion module performs channel weighting and pixel weighting on the feature map to achieve the fusion of main features. The post-processing module performs nonlinear mapping on the fused feature map to obtain a dehazing image. Experiments show that this network has achieved ideal dehazing effects on both synthetic and real-world images, and can avoid color distortion after dehazing.