Research On Underwater Fisheye Image Classification Based On Sphere-Net Framework

In the field of intelligent animal husbandry,in order to make full use of underwater resources,underwater objects need to be classified and recognized.In practice,due to the scattering and absorption of light when it propagates in the water,the underwater image taken will show some problems such as color distortion,unclear image,and even affect the expression of semantic information,thus affecting the subsequent processing of the image.In addition,in order to obtain underwater images,people tend to carry fish eye lenses on underwater robots for shooting.Fisheye lens is cheap and has a wide field of view,but because of its special imaging method,the image has a certain degree of distortion.Traditional classification methods based on object detection lack feature processing methods for fish eye images,which makes the accuracy and generalization of the model poor.Finally,in order to provide undistorted images to the observer as much as possible,distortion correction is needed,and the traditional methods are insufficient in correction ability,so the visual restoration effect of distorted images is poor.In this paper,depth learning,digital image processing and other technologies are comprehensively used to solve the above problems,and an underwater fish eye image classification system based on depth learning is designed and implemented.(1)Aiming at the problem of poor image quality taken underwater,a hybrid underwater image enhancement algorithm based on MSRCR and Dehaze Net is proposed through comprehensive comparison and analysis.The hybrid enhancement algorithm uses MSRCR algorithm and Dehaze Net to perform enhancement processing in turn,and comprehensively improves the image contrast and texture detail information,and the algorithm operation time is within the acceptable range,and the enhanced image strengthens the image semantic features,laying a foundation for subsequent classification.(2)Aiming at the problem of insufficient correction effect of traditional correction algorithms,we propose a multi angle fish eye image correction algorithm.The multi angle correction algorithm realizes the correction of the optional angle of each region of the fisheye image through the transformation of the coordinate system.On the basis of basically retaining the complete information of the image,it can also take into account the image quality and meet the visual habits of the human eye.The experimental results show that the multi angle correction algorithm can correct images with different shooting angles and different distortion angles,and has wide applicability.In addition,compared with chessboard correction method and longitude and latitude correction method,the multi-scale correction method proposed in this paper has performance advantages in both subjective and objective evaluation criteria.Finally,for the distortion of the fish image taken by the actual underwater fish eye,the algorithm can also better correct the texture details of the fish target object.(3)In the fish target classification task of underwater fisheye image,the traditional convolutional neural network has weak feature extraction ability and low recognition accuracy,and a SConv-OD model based on Sphere-Net is proposed.SConv-OD uses spherical convolutional neural network instead of ordinary neural network for modeling,so that the image of fisheye lens can be captured by targeted abstract features,which effectively improves the accuracy of fish identification and model generalization.Experimental results show that compared with other target classification models,SConvOD improves the detection accuracy and intersection ratio of targets of different scales,especially for low-resolution targets.In addition,the spherical convolutional neural network introduced by SConv-OD has little impact on the inference speed,which ensures that the model still has high inference efficiency.(4)We analyze the functional and nonfunctional requirements of the underwater fish eye lens target classification system,and design the system according to the actual business process.We design and implement an underwater fish eye lens target classification system.The system integrates various functions such as data set making mode,input and output interaction mode,data enhancement and so on.The system can realize the mixed enhancement of MSRCR and Dehaze Net for the pictures uploaded by users,and realize the operation of target detection and classification and recognition for the interested targets in the fisheye image,and finally display the results after completing the multi angle correction of the fisheye image.In addition,the actual running time of the system is completely within the acceptable range of users,which has a wide range of practicality.