Automatic classification method of side scan sonar image targets based on transfer learning and depth learning

Abstract

The invention belongs to the field of automatic recognition and classification of underwater targets, in particular to an automatic classification method of side scan sonar image targets based on transfer learning and depth learning. The method includes acquiring a conventional optical image data set with segmentation annotations; carrying out the contour segmentation by using the annotated imagescorresponding to each image in the data set; selecting a convolution neural network structure for training to obtain the source domain target classification network; fozening the parameters of the front part of the fully trained source domain target classification network, and setting the parameters of the back part of the classification network to the trainable state; continuing to train the setclassification network using the training set; when the training is complete, using the verification set to evaluate the performance of the classification network. The method uses the transfer learning method to transfer the convolution neural network trained with non-side scan sonar images, and pre-processes the source domain data set according to the similarity principle, so as to improve the transfer learning efficiency and prevent the negative transfer phenomenon.

Description

technical field [0001] The invention belongs to the field of automatic recognition and classification of underwater targets, and in particular relates to a method for automatic classification of targets in side-scan sonar images based on transfer learning and deep learning. Background technique [0002] The automatic classification of side-scan sonar image targets is of great significance to ocean detection and underwater search, especially in the search for crashed planes and sunken ships. At present, the common search method is to use the sonar carried by AUV to scan a large area of ​​the target seabed, and then copy the data after a complete scan of a sea area, and then manually judge whether there is a target, because it does not have the ability to detect and identify autonomous targets. The operating efficiency of this search process is low, so how to improve the ability of autonomous detection and recognition of targets in sonar images has become more and more importa...

AI Technical Summary

Problems solved by technology

For convolutional neural networks, migration learning is to successfully apply the "knowledge" trained on a specific data set, that is, the source domain, to a new domain, that is, the target domain. This is undoubtedly to solve the problem caused by too little sample data. Unable to apply the best solution for deep learning

Images