Image Dehazing Algorithm Based on Multi-Scale Detail Enhancement and Parallel Attention Mechanism
XiaoLan Yang, Juan Wu · 2024
Previous single image defogging methods based on deep convolutional neural networks have worked on improving the performance of the network by increasing its depth and width. Current approaches focus on increasing the size of the convolutional kernel to obtain a larger sensory field to improve performance. However, directly increasing the size of the convolutional kernel introduces a large number of parameters and increases the computational overhead. Therefore, an image defogging algorithm based on multi-scale detail enhancement and parallel attention mechanism is designed to solve the above problems. A parallel large convolutional kernel module with large receiver field and multi-scale characteristics is designed to recover texture details while capturing large blurred regions. The enhanced parallel attention module is able to handle inhomogeneous haze distribution efficiently. Low-level and high-level features of the encoder are adaptively fused through the learned spatial weighting features. The evaluation results show that the designed multi-scale detail enhancement and parallel attention mechanisms are effective. On the SOTS outdoor dataset, the algorithm makes a significant improvement in PSNR and SSIM with only 6.92M parameters.