CNN-based Single Image Dehazing via Attention Module
Enguang Wang, Shanfu Shu, Cheng Fan · 2022 IEEE 5th International Conference on Automation, Electronics and Electrical Engineering (AUTEEE) · 2022
Methods based on deep learning have achieved remarkable performance in image deblur processing. Most of the deep learning methods used for image restoration and enhancement have no end-to-end modeling, and too many network layers lead to slow calculation and can not effectively integrate the overall features. Therefore, we propose a neural network with convolution module attention mechanism to jointly estimate the projected picture and atmospheric light, and embed the altered atmospheric scattering model. The CD-Net with Convolutional Block Attention Module proposed by us integrates global features and avoids CNN losing the low-level semantics learned in the early layers. Meanwhile, the reconstructed the fog degradation model model can greatly reduce the amount of computation, making our model simple to construct and efficient to compute. We test our network on NYU2 dataset and SOTS dataset and verify that our method is superior to other advanced image deblurring techniques by quantitative and qualitative methods, and the images after processing are easier for computer recognition or better suited to human eyes' visual qualities. Automatic driving, medical imaging, remote sensing imaging, character photography, and other industries can all benefit from image enhancing technology.