A Novel Lightweight Algorithm for Sonar Image Recognition.

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Sonar images suffer from low resolution, high noise, and blurred edges, challenging traditional Convolutional Neural Networks (CNNs).
• Existing CNNs exhibit large sizes and high computational demands, hindering deployment on resource-constrained embedded systems.
• Inadequate target recognition accuracy is a significant issue in sonar image analysis.

Purpose of the Study:

• To develop a lightweight and accurate sonar image recognition algorithm.
• To enhance the feature extraction capabilities for sonar image characteristics.
• To enable the deployment of advanced recognition models on embedded devices.

Main Methods:

• Modified the MobileViT block by redesigning the jump connection layer to better capture critical sonar image features.
• Replaced the 1x1 convolution in the MV2 module with a multi-scale convolution Res2Net for improved global and local feature learning.
• Applied an imbalance (IB) loss function to manage sample category imbalance in the sonar dataset, assigning differential sample weights.

Main Results:

• Proposed modifications demonstrated varying degrees of improvement in sonar image recognition accuracy.
• The enhanced algorithm maintains a lightweight profile, suitable for embedded system deployment.
• The integration of Res2Net and modified jump connections effectively improved feature learning capabilities.

Conclusions:

• The optimized MobileViT algorithm offers a viable solution for accurate and efficient sonar image recognition.
• The study successfully addressed limitations of traditional CNNs in sonar applications.
• The proposed method facilitates the practical application of deep learning in embedded sonar systems.