Research on Improved VGG-16 Model Based on Transfer Learning for Acoustic Image Recognition of Underwater Search and Rescue Targets

Abstract

The advancement of underwater search and rescue technology has made target image recognition crucial for improving efficiency and accuracy in side-scan sonar operations. However, the complexity of the underwater environment presents challenges, such as high computational resource requirements, scarcity of samples, and uneven data distribution. To address these challenges, this article proposes an improved visual geometry group 16 (VGG-16) model based on transfer learning for target image recognition in underwater search and rescue. First, the side-scan sonar target image data are processed through manual supervised segmentation, noise addition, and normalization to enhance data diversity and quantity. Second, the VGG-16 model structure is lightweighted and batch normalization is incorporated to enhance training efficiency. Finally, the improved VGG-16 model is trained and tested using a frozen transfer learning strategy on a small-sample side-scan sonar target image dataset. Results show that compared with the traditional machine learning and VGG-16 models, our proposed transfer learning improved VGG-16 model exhibits higher performance in terms of efficiency and accuracy in underwater search and rescue target image recognition. Its recognition accuracy reaches 97.70%, with a faster convergence speed. Additionally, its average precision value is 97.14%, representing the improvements of 9.50% and 6.40% over VGG-16 and improved VGG-16 models, respectively. This indicates the effectiveness and feasibility of our approach in enhancing model recognition capability and training efficiency, validating its practical application potential.