A deep learning-based object detection method for ship images

Abstract

The present invention provides a ship target detection and recognition method based on deep learning, comprising the following steps: S1: building a pixel attention model, and preprocessing the ship image; S2: using K-Means clustering to generate a ship anchoring frame, And convert the label bounding box; S3: Build the YOLOV3 network structure based on the pixel attention model; S4: Use the training optimization method to train the network; S5: Use the non-maximum suppression algorithm to post-process the network output to avoid repeated detection. question. The deep learning-based ship target detection and recognition method provided by the present invention can realize ship target detection and recognition under various complex backgrounds and resolutions, and has a good application prospect in the fields of shipbuilding industry and maritime management.

Description

technical field [0001] The invention relates to a deep learning and target detection technology, in particular to a deep learning-based ship image target detection method, which belongs to the field of deep learning and computer vision. Background technique [0002] Ship target detection and recognition methods can be divided into three strategies, including end-to-end network structure, two-stage network structure and improved network structure based on the former two. For the end-to-end ship target detection and recognition network structure, Ling Ziqin, Chang Yang-Lang and Wang Bingde directly use YOLOV1, YOLOV2 and YOLOV3 networks to achieve ship target detection and recognition, but the network effect cannot meet the engineering standards. Xia Ye et al. used the SSD network to build a ship target detection and recognition system, which achieved a certain improvement in the detection accuracy of the network, but sacrificed the real-time performance of the network. For t...

AI Technical Summary

Problems solved by technology

In general, S-CNN can be regarded as an R-CNN optimized by a general method, which has a great improvement in the generation of candidate bounding boxes, but has no improvement in the target recognition of images inside the candidate bounding boxes.

Images