Research On Detection Method Of Underwater Disappeaser Based On Deep Learning

Effectively prevent and reducing the loss of floods is a problem that human society is always facing in development.When the water happens,it is also the most important step in the first step of underwater rescue when the water is happening.With the continuous improvement of science and technology,people gradually use equipment such as underwater robots to replace the artificial rescue method.However,the traditional target detection method has poor detection effect,slow detection speed,and does not realize real-time detection,and there is a problem with not enough missing sonar image data for detection.Therefore,based on the above problems,image data collection is performed by sub-underwater missing people’s morphology and characteristics of underwater disappears,with sonar images as the main research object,performing depth learning underwater missing person image generation method and target detection method research,implementation The purpose of detecting accuracy in sonar data sets and increasing the detection accuracy in sonard images:First,a study of a sonar image generation method based on convolutional neural network is performed for the number of sonar data sets.By analyzing the number of sonar images,there is almost no pair underwater optical image and sonar image characteristics,select a suitable generating counterfeiting network as a basic network,and then analyze the convolutional neural network,increase training with the number of layers Generating gradient disappears,post-transformation partial target spatial information loss and style conversion effect problem,introducing residual connections in the underlying network,style loss function,proposing an improved generating counterfeiting network.Through experiments,the optical image generating sonar images of the diver is realized,and the effectiveness of the methods mentioned herein is verified.Second,a sonar image detection method based on convolutional neural network is performed against sonar target detection.By analyzing the real-time and accuracy requirements of the target detection in the context of the rescue,select a suitable target detection algorithm as a basis,and then conduct training by analytively target detection.The anchor frame is not suitable enough for the target scale and the common small objective identification inaccurate problem,introducing a data enhancement method and adaptive anchor frame method,and add a variety of different attention mechanism modules to compare,put forward an improvement The target detection algorithm is verified by experiments,and the experimental results of the detection algorithm are compared and analyzed.Finally,the pool experiments performed on the number of lack of sonar data sets are introduced,and the diver is collected underwater optical images and sonar images.By using improved generating sonar images,a few methods such as projection transformation and data set increase,the sonar image data set is expanded,and the improved target detection algorithm is used in the expanded data set,compared to analysis.The detection results of the two data sets,verify the effectiveness of the expansion method and further enhance the accuracy of the target detection.