SVM Based Dynamic Target Detection in Underwater Sonar Image

Abstract

For underwater target detection and recognition tasks, sonar is widely used because it has a larger range than optical detection equipment. However, target detection and recognition could be difficult in underwater sonar images due to the larger amount of noise than optical images. Moreover, this task usually requires real-time execution on hardware-constrained devices in practical applications. Therefore, in this work, we propose a sonar image target detection and recognition method that can run in real-time on low-energy devices. Firstly, this method applies the inter-frame information to continuously track the connected domains in the image to detect possible targets in the image. Secondly, it uses the support vector machine model to classify the manually extracted target features. Finally, the category voting mechanism is adopted to determine the target category. Experiments were carried out on 2122 sonar images containing 5 types of targets, and the results showed that the $$mACC$$ of the 5 types of targets by this method achieved to 96%, and the recognition rate of any arbitrary type of target was not less than 87.75%. The processing time is 0.1 s/frame on Intel I3 seven-generation processors.