Research And Application Of Water Surface Target Recognition Based On Machine Vision

With the further development and utilization of rivers,lakes and oceans,the number of water surface equipment and the use of ships are also increasing.In addition,the current environmental protection awareness is not strong,and then with the increase of water waste,pollution,too many water surface equipment and ships lead to frequent navigation accidents and other situations.Ship identification,common garbage and marine equipment identification can provide research basis for garbage cleaning and ship traffic monitoring applications.Therefore,how to identify garbage,surface equipment and running ships is of great significance to the treatment of environment and surface traffic,so it has great research value.Machine vision technology based on Deep learning has made great progress in the field of automatic driving or UAV.However,there are few researches on water surface target recognition.Now,the traditional machine vision method is still used to extract the target.The scene is relatively simple,and the algorithm is difficult to engineer.At present,there is little research on the deep convolution neural network in the field of water surface target recognition,which deep learning algorithm should be applied to water surface target recognition needs further research.Therefore,this paper studies the related knowledge of deep learning,and selects three deep convolution neural network algorithms,faster R-CNN,SSD and yolov3,which are highly used in engineering,as comparative experiments.That is,for the same water target training set and water target test set,different algorithms are used for comparison,and the most suitable depth learning algorithm for water target recognition is selected.Through the comparison of recognition rate and recognition speed of water target,it is concluded that yolov3 algorithm is superior to the former two.Then through further research on the network structure and loss function of yolov3 algorithm,the algorithm is further optimized to achieve better recognition effect in water surface target recognition.In the same way,this paper also makes a comparative experiment of yolov3 algorithm before and after the improvement,and draws the conclusion that the recognition rate of yolov3 algorithm after the improvement is higher,and the convergence speed during the training is better.After yolov3 algorithm is selected,the data of water surface data set needs to be enhanced,so that the effect of later algorithm research is better.By collecting a large number of trainable data to build the training set and test set of water surface target,using data enhancement techniques such as random rotation,random brightness and contrast,and adding noise to simulate different weather conditions,such as strong light or evening,which not only increases the diversity of training data but also increases the number of data,but also enhances the generalization of algorithm and improves the recognition rate.In addition,when building a richer surface target data set,data from different angles and scenes are collected to enhance the effectiveness of training results.In order to increase the speed of water target recognition,the algorithm of water area segmentation in various water areas is studied.Because of the particularity of water target recognition,generally,water area does not occupy the whole picture of recognition image.But in the conventional algorithm reasoning,it is necessary to filter the candidate frame in the whole picture.The selection of candidate frames takes up a large part of target detection time when the algorithm is running,so the water area segmentation algorithm proposed in this paper can roughly segment the water area,so that all candidate frames of the target whose bottom frame is not in the water area are eliminated,which can greatly reduce the number of candidate frames,improve the reasoning speed of the algorithm,and contribute to the engineering application of the algorithm.Based on this phenomenon,this paper proposes three water area segmentation algorithms.Firstly,the water area segmentation algorithm based on sea antenna recognition can be adapted to the sea area with obvious sea antenna.At this time,the picture is relatively wide.This algorithm can be used to obtain the water area through the algorithm of extracting sea antenna.For clear lakes,a water area segmentation algorithm based on phase correlation algorithm is proposed.Because of the obvious reflection in this water area,the water shoreline can be calculated by the phase correlation between the reflection and the real object,and then the water area can be obtained.The first two algorithms can react quickly in special water area and get water area.However,the recognition of general waters is relatively low,so this paper proposes a third water area segmentation algorithm,water area extraction algorithm based on region segmentation.This algorithm uses the watershed algorithm based on the mark to extract the water area,and the recognition effect is better than the first two in the case of complex waters.In addition to the research on the algorithm itself,this paper has done some research on the validation and engineering application of the algorithm.Firstly,the algorithm is used to train the self-made water surface target data set enhanced by data,record the middle results of training such as loss value and IOU change rate during the training process,and then analyze them,and then analyze the effectiveness of the algorithm training and observe whether the degree of convergence is the same as the theoretical prediction.Finally,the weight file after training is used to verify the test set,and the recognition results in different situations are compared.Finally,the weight file after training is applied in the engineering of unmanned ship.First of all,the USV(Unmanned Surface Vehicle)system is reformed to control the USV.The USV is equipped with HIKVISION camera,lidar,esp8266 and other equipment,and then a local area network is built to carry out the communication link between USV and the upper computer.Then,the algorithm is applied to the USV for surface target recognition.The experiment shows that the target recognition algorithm designed in this paper has better recognition effect in the aspect of water target recognition.