【摘要】 针对水下探测机器人在视觉感知过程中由于图像退化容易造成辨识困难的问题，提出基于通道特征融合的水下图像轻量化增强网络。通过构建基于通道注意力模块的随机残差结构，并设计通道混洗模块，实现原图像与图像各层特征的细节融合；采用基于倒残差卷积的通道评分模块，通过通道之间的回归分离，提升网络对图像的增强效果；最后，网络以水下图像在颜色、局部特性等方向的增强为目标，设计包括高斯均方损失、结构性相似损失与感知损失等在内的网络损失函数，完成图像增强训练。通过对真实水下环境数据进行实验验证，算法在水下颜色色偏有更好的适应性，增强效果在细节保留上处理更优，模型参数更少，推理速度更快，更适合小型水下探测机器人的应用部署。更多还原

【Abstract】 In order to solve the problem that the image degradation would cause identification difficulties in visual perception process of underwater exploration robot, an underwater image lightweight enhancement network based on channel feature fusion was proposed. By constructing the random residual structure based on channel attention block and using channel mixing module, the detail fusion of original image and features of each layer of the image was realized. In addition, the network architecture introduces a channel scoring module based on inverted residual convolution to enhance the network’s image enhancement effect through regression separation between channels. Finally, aiming at the enhancement of underwater images in color and local characteristics, loss functions including Gaussian mean square loss, structural similarity loss and perception loss were designed to complete image enhancement training. Through experimental verification of real underwater environment data, the algorithm was better adapt to the underwater color bias and the model parameters were reduced to 1/36 of Funir-GAN network, 1/21 of Deep SESR network and 1/2 of Shallow UWNet network. The inference speed was also faster, which was more suitable for the deployment of small underwater exploration robots.更多还原