Cross-Layer Attention Feature Interaction and Multi-Scale Channel Attention Network for Single Image Dehazing
In recent years,U-shaped convolutional neural networks(CNNs)have achieved remarkable progress in im-age dehazing.However,most U-shaped dehazing networks directly pass encoder features to the decoder at the corresponding scale,ignoring effective utilization of multi-scale features.In addition,channel attention widely used in dehazing networks is restricted by receptive fields,failing to sufficiently leverage contextual information,which adversely affects learning of chan-nel weights.To address the above issues,this paper proposes a novel dehazing algorithm with cross-layer attentive feature in-teraction and multi-scale channel attention.Specifically,the cross-layer attentive feature interaction module learns hierarchi-cal weights for multi-scale encoder features,and aggregates these cross-layer features for transfer to the decoder,thereby re-ducing feature dilution during the dehazing network's reconstruction of clear images.Moreover,to uncover channel informa-tion that is critical for dehazing networks,we devise a multi-scale channel attention mechanism that extracts multi-scale fea-tures by dilated convolutions with different dilation rates,forming a parallel learning scheme of channel attention with multi-scale contexts for more effective weight allocation for dehazing network features.Experimental results demonstrate that the proposed dehazing algorithm achieves better objective metrics and visual performance compared to 12 existing methods on 4 public datasets.The code for this paper has been uploaded tohttp://github.com/bohuisir/AAFMAF.
万方数据
维普
