Citation

Ruan F, Dang L, Ge Q, Zhang Q, Qiao B, Zuo X. Comput. Intell. Neurosci. 2022; 2022: e6962838.

Copyright

(Copyright © 2022, Hindawi Publishing)

DOI

10.1155/2022/6962838

PMID

unavailable

Abstract

The underwater environment is complicated and changeable and contains many noises, making it difficult to detect a particular object in the underwater environment. At present, the main seabed detection technology explores the seabed environment with sonar equipment. However, the characteristics of underwater sonar imaging (e.g., low contrast, blurred edges, poor texture, and unsatisfactory quality) have serious negative influences on such image classification. Therefore, in this study, we propose a dual-path deep residual "shrinkage" network (DP-DRSN) module, which is a simple and effective neural network attention module that can classify side-scan sonar images. Specifically, the module can extract background and feature texture information of the input feature mapping through different scales (e.g., global average pooling and global max pooling), whereas scale information passes through a two-layer 1 × 1 convolution to increase nonlinearity. This helps realize cross-channel information interaction and information integration simultaneously before outputting threshold parameters in a sigmoid layer. The parameters are then multiplied by the average value of the input feature mapping to obtain a threshold, which is used to denoise the image features using the soft threshold function. The proposed DP-DRSN study provided higher classification accuracy and efficiency than other models. In this way, the feasibility and effectiveness of DP-DRSN in image classification of side-scan sonar are proven...

... To further verify the performance, we present a confusion matrix for each model based on the ResNet34 structure (Table 8). Compared with other models (SE, CABM, and RSBU-CW), the proposed model slightly increased the number of correct predictions for three image types: aircraft, seafloor, and shipwreck. Although the number was relatively small, it provided a significant improvement for fewer numbers of SSS image samples. While there was no improvement in the seafloor category compared with the SE module, four more images were predicted correctly in the aircraft category, especially when the total number of aircraft in the test set was 18 images, demonstrating a particularly objective improvement. However, in the category of drowning victims, the proposed model's performance was slightly insufficient, registering fewer correct images less than CABM and RSBU-CW. This may be due to the proposed model being relatively complex in structure and the small number of drowning victims preventing adequate network parameter training. This would cause discretely inferior classification performance. In addition, all models possessed a large error rate in the aircraft category due to the target objects of some aircraft images being too small or incomplete. Accuracy could be improved by redefining the size of target objects...

Language: en