Feature Extraction And Recognition Technology Research On Target Images Of USV Vision System

As a recently emerging offshore work platform, Unmanned Surface Vehicle(USV) cancomplete the missions of detecting object, avoiding obstacle, searching and rescuingautonomously according to the artificially set procedure, and the autonomous target detectionand recognition is the key technology. Relative to the long-distance radar, infrared targetdetection, the new vision system is the main equipment system to achieve and process closetarget, however, the currently domestic existing target image recognition processingtechnology for USV visual system is not mature. According to the common characteristics ofsurface target images and the real-time demand of target recognition, a set of targetrecognition processing technology for USV visual system is proposed and the correspondingexperiment simulation is completed.Taking the surface targets as research object, according to the research approach ofaccording to target detection, target characteristics analysis, and target recognition, imagesegmentation algorithm, feature extraction method and the target image recognitiontechnology are deeply studied respectively:Firstly, the image library of surface targets is established by searching data from the weband taking pictures of the actual distance-control boat, mainly including: different kinds of3dship types and its rotation appearance in the range0to360degrees; images ofdistance-control boat on water shot under different angles of view; images of sea reefs andislands under different angles of view.Secondly, considering the complexity of sea-sky background and the change of the depthof target images collected by the actual visual system, the traditional threshold segmentationalgorithm cannot get complete detection of target, so the Mean-Shift segmentation algorithmis put forward to detect surface targets, and the adaptive Mean-Shift segmentation technologyis proposed based on the algorithm. In the experiment, the complete segmentation is achievedand the real-time processing requirement is ensured.Thirdly, with surface ships as the main part and reefs and islands as the secondary part,the features of surface targets are analyzed respectively from the aspects of surface textureand shape. The surface texture features and the geometric features, hu moment invariants,affine moment invariant of shape are extracted, and furthermore for the ship target, somefeatures do not have invariance when rotating out-of-plane, thus features under differentangles of view are richly extracted. The experiment results show that the features of different types of targets extracted in the proposed methods have good separability, and are suitablelyused as the reference data for target recognition.Finally, according to the feature extraction technology, a comparatively complete featurelibrary of surface targets is set up and used as the input sample library of BP neural network;Analyze the common problems of BP algorithm and optimize the network learning samplesusing principal component analysis; Combine multiple single-output neural network parallel,extract texture and shape features of marine and non-marine and adjust the network learningstep adaptively to train network to distinguish Marine and non-marine and identify thedifferent types of ships. The experimental results show that the method on one hand raises thelearning speed of neural network, and on the other hand improves the recognition rate.