Research On Sonar Image Classification Algorithm Under Small Samples Based On Deep Learning

Due to the presence of a large amount of clutter and noise in sonar images,as well as the unclear boundaries of the targets,the extraction of features from sonar images is greatly affected,which in turn affects the performance of classification algorithms.With the development of artificial intelligence technology,image classification algorithms based on deep learning are gradually being applied in the field of sonar image recognition.However,due to the difficulty in obtaining sonar datasets,the algorithms often face the problem of overfitting with limited samples,and simply augmenting the dataset can also result in decreased convergence speed and failure to meet the speed requirements of practical applications.To address the tasks of improving the quality of sonar images,increasing the accuracy of underwater sonar image classification,and enhancing model training efficiency,this paper proposes a classification method for a small amount of sidescan sonar image samples.This is achieved by considering three aspects: improving the quality of the images and increasing the number of samples,improving the classification network model,and changing the training mode.The performance of the proposed algorithm is verified on different datasets.The main work and research contents of this paper are summarized as follows:(1)To improve the quality of sonar images and enhance the classification accuracy,the study proposes a block-matching-based denoising method that matches and groups similar pixel blocks to restore clear images and remove speckle noise that significantly affects sonar image quality.Compared to other traditional denoising methods,the proposed method demonstrates improved peak signal-to-noise ratio and structural similarity indexes.Moreover,the classification performance of the network trained on the denoised dataset outperforms that of the network trained on the original noisy dataset.(2)To address the overfitting problem caused by the limited sample size,the study proposes a feature mining-based sonar image classification algorithm that introduces residual modules into CNN to deepen the feature mining process and alleviate overfitting.The study proposes a residual module that combines feature mining and multiscale target adaptive feature extraction modules,which exploits the correlations between feature channels to select convolutional kernel sizes that correspond to different target scales,maximizing the utilization of limited samples.The study demonstrates the feasibility of solving the limited sample problem through experiments on a small digital dataset and shows that the proposed algorithm outperforms other classic image classification networks on sonar image datasets.(3)To improve the training efficiency of CNN using data augmentation and meet the strict time requirements of sonar detection technology,the study applies the idea of transfer learning to sonar image classification tasks.The study trains the network model through a transfer learning method based on model parameters,which preserves the network’s ability to extract low-level features and integrates feature mining modules in the high-level feature extraction stage to improve feature utilization.The study’s experiments on sonar image datasets verify that the proposed algorithm improves model convergence speed while maintaining sufficient classification accuracy,achieving the goal of trading off a small amount of classification accuracy for higher training efficiency.