Sonar Image Segmentation Based On BP Neutral Network Optimized By Genetic Algorithm

The recognition technology of underwater target image provides powerful technological support for the exploitation of seabed mineral resources,submarine landform and submarine targets.The segmentation for sonar images is a key step in the recognition process of underwater targets.Only when sonar images are segmented accurately,can underwater targets be recognized precisely.Because the quality of filtered sonar images affects the accuracy of the segmentation for sonar images directly,effective methods of filtering and segmentation for sonar images need to be provided.On the base of BP neural network research,a filtering algorithm based on multi-feature BP neural network and a segmentation algorithm based on multi-feature BP neural network for sonar images are proposed in this article.In the part of sonar image filtering,five statistical characteristics,namely,8-neighborhood mean,variance,minimum,mode and median for each pixel of sonar image,which are filtered by the algorithm in this article,are selected as the training input of BP neural network.The gray values of the target bright area,target dark area and reverb zone are set as the mean of gray values of all pixel points in the corresponding area and used as the training output of BP neutral network and the sonar images are filtered by the trained network.Because of the possibility of slow convergence speed and falling into local minimum of BP neural network,an improved genetic algorithm is used to optimize its training process.Two kinds of evaluation index(mean square error,peak signal noise ratio)are used to evaluate the quantity of filtered image.By comparing with other four classical filtering algorithms(median filtering,Frost filtering,Kuan filtering and Lee filtering),effectiveness of the filtering algorithm in this article is proved.In the part of sonar image segmentation,five statistical characteristics,namely,8-neighborhood mean,variance,minimum,median and mode of each pixel of sonar images,which are filtered by the algorithm in this article,are selected as the training input of BP neural network.The gray value of the target bright area,target dark area and reverb zone is set as 255,0,125 separately and used as the training output of BP neutral network.The sonar images are segmented by the trained network.Similarly,an improved genetic algorithm is used to optimize its training process.Two kinds of evaluation index(false positive rate,similarity)are used to evaluate the quality of segmented images.By comparing with other three classical segmentation algorithms(Kapur threshold segmentation,Otsu threshold segmentation and fuzzy C-means clustering segmentation),effectiveness of the segmentation algorithm in this article is proved.