Underwater sonar target detection system and method

Abstract

The invention provides an underwater sonar target detection system and method, and belongs to the field of underwater sonar target detection. The system comprises a dividing module, a preprocessing module, a PP-YOLOv2 module, a training module and a calling module. In the invention, a training module is utilized to call image data after filtering and segmentation processing to train a PP-YOLOv2 module until error loss of a prediction result and a real result meets a threshold value, and training of the PP-YOLOv2 module is completed; and inputting the filtered and segmented image to be detected into the trained PP-YOLOv2 module by using a calling module to obtain an underwater sonar target detection result. According to the invention, end-to-end one-stop definition and use are completed. The global feature information sensing capability of the underwater sonar image target detection network can be effectively enhanced, and the underwater sonar image detection capability of the target detection network is improved.

Description

technical field [0001] The invention belongs to the field of underwater sonar target detection, in particular to an underwater sonar target detection system and method. Background technique [0002] Underwater sonar has important application value in the detection and identification of mine-like objects, seabed appearance, shipwreck salvage, and the search for wrecked aircraft. At present, the traditional search method is to use side scan sonar to conduct large-scale After all areas are scanned, the area image recognition is performed manually. This method causes a lot of time wastage. At the same time, due to manual interpretation, it will inevitably lead to misjudgment and misjudgment. [0003] In recent years, with the emergence of deep learning, the deep target detection network composed of multiple layers of neural networks has been greatly improved in the accuracy of sonar image target detection and recognition. With the deepening of parameters, the network extracts fe...

AI Technical Summary

Problems solved by technology

[0002] Underwater sonar has important application value in the detection and identification of mine-like objects, the appearance of the seabed, the salvage of sunken ships, and the search for crashed aircraft. The current traditional search method is to use side-scan sonar to conduct large-scale After the scanning of all the areas is completed, the area image recognition is performed manually. This method causes a lot of time wasted, and at the same time, due to the situation of manual interpretation, it will inevitably lead to misjudgment and misjudgment.

[0003] In recent years, with the emergence of deep learning, the deep target detection network composed of multi-layer neural networks has greatly improved the accuracy of sonar image target detection and recognition. With the deepening of parameters, the network extracts features and The ability to combine features is increasing, but the target detection network requires a large amount of training data, and it is difficult to obtain underwater sonar images. Therefore, someone proposed a target detection network based on transfer learning, which is to migrate the trained optical target detection network to Acoustic direction has greatly improved the performance, but the convolutional neural network is limited by the size of the convolution kernel, so it can only obtain local feature information, but cannot obtain global feature information. How to improve the global information extraction of the network Ability to become an urgent problem to be solved

[0004] At present, the latest PPYOLOv2 deep target detection neural network has high accuracy while taking into account the detection speed. Using V100 on the COCO dataset has 50% Map accuracy and the speed can still reach 68.9FPS, but due to the sonar image The shape uncertainty and other interferences, directly applying PPYOLOv2 to the image recognition effect of underwater sonar cannot achieve the application effect

Images