Multi-scale Attention Map Guided Image Rain Removal Network
Gu Kunyuan · Academic Journal of Computing & Information Science · 2023
To solve the problem of low running speed of image rain removal models, a multi scale attention map guided image rain removal network is proposed. Firstly, an original resolution main network with residual structure is constructed for image rain removal. Secondly, in order to enhance the rain removal ability of the main network, especially the recognition ability of multi-scale features, a multi-scale attention map generation network is designed as an auxiliary to the main network. The generated multi-scale attention maps are used to enhance the features of rain stripes in the image, guiding the main network to identify and remove them in turn. In addition, this paper also designed the dual attention guided convolution block, which adds the ability to pay attention to spatial features on the basis of traditional channel attention mechanisms to achieve joint attention to spatial and channel features. Experimental results show that the proposed method maintains the comparability and stability of the restored image quality while significantly reducing the amount of network parameters and effectively improving the speed.