关键词: 水下图像增强, Transformer, 残差卷积, U-Net网络, 平行注意模块, 通道注意, 像素注意, 卷积神经网络, 深度学习

Abstract:

In view of the problems of color distortion, low contrast and blurred details in underwater images, an U-Net network based on parallel Transformer and residual convolution is designed for underwater image enhancement. In the new U-Net structure, the HCTB(Hybrid Convolution Transformer Block) is placed in the encoding and decoding parts, which integrates the ability of the Transformer to capture global information and the ability of the convolutional block to capture local information, and builds a number of PAM(Parallel Attention Module) in the hopping connection part to extract more important pixel and channel information. The existing UIEB(Underwater Image Enhancement Benchmark dataset) paired dataset is used to train the network. In order to verify the effectiveness of the proposed algorithm, underwater images with different color degree are selected for experiments and tests. The experimental results show that the PSNR(Peak Single-to-Ratio) value of the proposed model is increased by 4.3% compared with other advanced models, and the subjective and objective evaluation results are obtained, which effectively improves the enhancement level of underwater images.

Key words: underwater image enhancement, Transformer, residual convolution, U-Net network, parallel attention module, channel attention, pixel attention, convolutional neural networks, deep learning