This article aimed to address the challenges of underwater target recognition algorithms in the face of large differences in lighting conditions, image blurring, and distortion by improving deep convolutional neural networks. By introducing the attention mechanism, the purpose of strengthening the representation of target information in the convolutional feature map is achieved, thereby reducing the interference of irrelevant information. Firstly, the pre-trained and improved deep convolutional neural network model ResNet-50 was used to fine tune the underwater target dataset to capture local and global information of the target and construct an attention mechanism network; secondly, attention mechanism was used to calculate the weight of each pixel in the feature map, making the target area more prominent. The attention weighted features can be fused with the original features, and the channel attention mechanism can be used to weight the importance of features between channels; finally, this article designed a classifier to recognize underwater targets, based on the Softmax classifier for target classification, and outputs the probability distribution of target categories. The research results indicate that the deep convolutional neural network improved by applying attention mechanism performs the best compared to other models in terms of accuracy, recall, and F1 value, reaching 0.85, 0.82, and 0.83, respectively. This indicates that deep convolutional neural networks improved by applying attention mechanisms can play a greater role in underwater target recognition.
