Research On Underwater Fish Recognition Technology Based On Embedded

With the increase in recent years,people’s attention on aquaculture,fish detection and recognition play a crucial role in environmental monitoring aquaculture and fishery development.At present,Underwater target recognition technology has become an important part of exploring the aquaculture industry,and the accurate identification of fish in the turbid waters of aquaculture is the guarantee of fishermen’s economic benefits.And with the increasing availability of underwater recording equipment such as action cameras and unmanned underwater equipment,it is possible to take pictures efficiently and safely without the logistical difficulties that are usually caused by manual data collection.However,underwater equipment collects a large amount of image data that needs to be processed manually,which has some impact on target recognition.Using deep learning to automate image processing has great benefits,but it is rarely used in the aquaculture field.In recent years,with the success of the depth of learning technology in various fields,more and more depth learning technologies have begun to be deployed on hardware embedded platforms.At the same time,embedded systems have made great breakthroughs in real-time detection and information collection.At present,Low cost and high efficiency of K210 development board has been applied to real-time target detection and direction recognition.Through the built-in KPU convolutional neural network accelerator and camera real-time acquisition of the K210,the real-time acquisition data of the camera is used as the input of the KPU and displayed the target category and confidence level in real time.By studying the depth of conventional neural network,and in conjunction with underwater optimize the image enhancement algorithms to achieve image enhancement in turbid waters.And underwater target detection model using enhanced binding characteristics extraction algorithm instead of the traditional image enhancement algorithm,higher accuracy compared to the conventional image enhancement based on underwater target recognition.And transplant it to the K210 hardware platform to realize target detection and recognition.In this paper,the research work as follows:1)Enhancement technology of underwater images.The adverse effects of underwater object recognition in order to solve the image quality is poor,we proposed an improved MSRCR(multi-scale Retinex with color restoration)from the target to improve the recognition rate,and angle of the image enhancement algorithms incorporated methods FUn IE-GAN.This method not only greatly improved the perception of underwater images,underwater images and the color distribution is more balanced,but also makes the color distribution of underwater images more balanced,which is beneficial to improve the recognition rate of underwater fish.In addition,the method is compared with image enhancement methods such as dark channel prior,MSRCR,restricted contrast histogram equalization,FUn IE-GAN algorithm,etc.,and we feel the traditional evaluation(information entropy,sharpness,contrast)and improve the recognition rate YOLOv4 model of subjective visual aspects from concept to verify the effectiveness and advancement of the proposed image enhancement algorithm.2)Research on target recognition model.First of all,in view of the problem of small target blurring in underwater images,multi-target and multi-category underwater fish target detection was performed on the image-enhanced data set,and YOLOv4 achieved an accuracy rate of 98.20%,a recall rate of 95.34% and a MAP(mean average precision)value of 95.09%.Then,Faster R-CNN method for detecting a target contrast YOLOv3(You only look once),demonstrate the effectiveness of the image enhancement methods described herein.By comparing the underwater fish target recognition results before and after image enhancement,the effectiveness and meliority of the mentioned method can be checked.At the same time,it also shows that in small samples and underwater scenes,a suitable image enhancement algorithm can improve the performance of target detection,and this improvement is proportional to the results of objective image quality evaluation.3)detection and recognition based on K210.First,the K210 boards only meet image resolution size 224×224 problem resolution underwater image scaling,image enhancement and image enhancement according to the proposed method.Second,select a lightweight network model for training and perform model conversion.Finally,the KPU convolutional neural network accelerator built into the K210 development board solves the problem of slow processing speed and realizes real-time detection.This paper performs image enhancement on poor quality underwater images,and proposes an improved image enhancement method combining MSRCR and conditional generation confrontation network.The superiority of this method is evaluated subjectively and objectively,and the target detection model is used in the data set.Perform training on the K210 development board,and Finally,target detection in K210 development board.Results of this study offer significant reference for optical vision system underwater robot.