Underwater Target Classification Based on Feature Fusion and Gene Encoding of CNN-BIGRU-Attention

Towards addressing the issues of poor image quality and limited sample quantity in underwater sonar images. A genetic encoding method for the fusion of image spot and local binary pattern texture features is employed in this study. By encoding the fusion features using genetic algorithms, the dimensionality of the fused features is reduced, thereby enhancing the efficiency of neural network operations while ensuring data transmission confidentiality. Additionally, a deep multi-classifier is introduced, which combines residual units with CNN-BiGRU-Attention. This classifier takes the genetically encoded fused features as input and outputs four predicted probabilities. Through an ensemble learning strategy that evenly assigns weights to the predicted probabilities, the final prediction result is obtained, achieving underwater target recognition. Results demonstrate that the proposed fusion features effectively enlarge inter-class differences among various types of sonar images and improve classifier accuracy. The deep multi-classifier model exhibits stable improvement in prediction accuracy compared to the original model, converging after only 15 training iterations with an average recognition accuracy of 99.1%. The research findings offer an effective approach for target sonar image recognition and classification, with potential applicability to a wider range of underwater target categorization.