Open AccessResearch on Lightweight Improvement of Sonar Image Classification Network
Author(s) -
Yan Zhou,
Shaochang Chen
Publication year - 2021
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/1883/1/012140
Subject(s) - convolutional neural network , computer science , bottleneck , sonar , convolution (computer science) , artificial neural network , artificial intelligence , pattern recognition (psychology) , network performance , field (mathematics) , data mining , contextual image classification , image (mathematics) , machine learning , telecommunications , mathematics , pure mathematics , embedded system
According to the constructed sonar common target detection dataset derived classification dataset, the convolutional neural network is used to classify the target. And in order to be able to better apply actual engineering, the lightweight of the network has been studied in depth. First, the performance of the ordinary convolutional neural network VGG-16 and the lightweight convolutional neural network MobileNet on the derived classification dataset is compared. The results show that the lightweight network achieves better results at a smaller cost. Then use the dilated convolution and bottleneck structure to improve the MobileNet network to obtain the IMDNet network and the IMBNet network. Through experimental comparison and analysis, the two improved networks have a significant reduction in the amount of parameters and calculations compared to the MobileNet network. The performance of the IMBNet network is basically the same as that of the MobileNet network, and the accuracy of the IMDNet network is 2.02% higher than that of the MobileNet network. This shows that in the field of sonar image classification research, lightweight convolutional neural networks have good performance, and have certain practical application prospects and values.