Underwater Multi-target Recognition Based On Generative Adversarial Network

The accurate identification of underwater target is the premise of underwater robot to grasp,catch and other safe operations,with the development of computer technology and the deep learning technology matures.In recent years,the ability of target detection and recognition for land-based images has been greatly improved,but for the underwater target detection and recognition,people have long committed to the research of underwater sonar technology,the research of underwater target detection based on visual tasks are relatively few.In the process of training underwater object recognition model with deep learning method,it is easy to encounter two problems.On the one hand,due to the influence of harsh underwater imaging environment,underwater images often show the characteristics of low contrast,fuzzy multinoise and color distortion.On the other hand,due to the particularity of the underwater environment,the requirements for underwater image acquisition equipment are high and the technical difficulties are great.Therefore,it is impossible to collect a large number of training data that meet the requirements of different underwater scenes and image quality,and the samples collected are prone to the problem of category imbalance.Therefore,it brings great difficulty for underwater multi-target detection.In view of the above problems,in this paper,a method to deal with small samples and class imbalance problems based on the Generative Adversarial Network was proposed,construct a depth neural network model for underwater multi-target detection,and conduct experiments on the original sample set and the expanded sample set of Generative Adversarial Network.The research work of this paper includes the following aspects.1)The method of underwater sample expansion based on Generative Adversarial Network is studied.For the problem of small sample and class imbalance of underwater image,by intercepting category target images that the number of images are small in the training samples,The target image similar to the real image is generated by training the Generative Adversarial Network,and the purpose of sample expansion is achieved by fusing the generated target and background image.Finally,the similarity between the generated image and the real image is verified by comparing the color histogram and histogram Bhattacharyya coefficient of the generated image and the real image.2)The underwater target detection method based on YOLOv3 algorithm is studied.Firstly,the characteristics of underwater image target detection are analyzed.On this basis,the underwater target detection model based on YOLOv3 algorithm is constructed,and the parameter design and optimization of the model are completed.In order to verify the effectiveness of the proposed method,experiments were carried out on the original image set and the expanded sample set respectively.The experimental results show the feasibility and effectiveness of the sample expansion method based on Generative Adversarial Network.At the same time,the effectiveness of the proposed underwater object recognition model is verified by comparing the improved object recognition model with the original model.3)A complete application system of underwater target detection algorithm is designed,which has the functions of multi-target recognition model training,image detection and real-time detection.In this paper,the method of underwater sample expansion based on Generative Adversarial Network is studied,and the underwater target recognition model is constructed and experiments are carried out on the original sample set and the expanded sample set.The experimental results show that the method proposed in this paper can improve the problem of small samples and class unbalance to a certain extent,improve the accuracy of underwater target identification,and it is of great significance to realize the accurate capture of underwater targets and promote the development of underwater operations and Marine resources.