Marine Animal Target Detection And Classification Based On GPU Embedded System

In the world,many countries are striving to develop the construction of marine pastures.Marine pastures can bring huge income through farming fish,shellfish,sea cucumbers,etc.In current marine aquaculture,observation of marine animal population density and fishing operations are carried out by trawl fishing or artificial underwater operations.These methods are inefficient,the data information is inaccurate,and the low-temperature and high-pressure underwater environment will have a certain impact on personnel safety.In recent years,the development of convolutional neural networks in deep learning theory has made researchers from various countries achieve good results in many fields.Therefore,this paper introduces deep learning theory into the target detection of marine animal images.We perform feature extraction and classification through convolutional neural networks,and use GPU embedded hardware suitable for underwater environments to complete underwater real-time target detection.Such a method can provide a basis for future biological population monitoring and intelligent fishing.The main research contents of this article are as follows:(1)The real-time image acquired by underwater equipment is enhanced.Due to the poor lighting conditions in the underwater environment,the water has a scattering effect on the light,the image will have the problem of lack of clarity and contrast.Therefore,the image needs to be enhanced.The CLAHE algorithm and MSRCR algorithm with color restoration are used to enhance the underwater images respectively.The two algorithms are tested from the collected underwater unclear pictures.After comprehensive analysis and comparison,the MSRCR algorithm was finally selected to enhance the images collected underwater.(2)MobilenetV2 convolutional neural network for marine animal image classification is established.In order to obtain a model with excellent classification effect and speed on GPU embedded devices,this paper proposes an image classification method based on convolutional neural network and transfer learning.By migrating the parameters of the MobilenetV2 network and fine-tuning the parameters for retraining using the fine-tune method,we can obtain a classification model suitable for marine animal images.The experimental results show that MobilenetV2 network model can obtain the best accuracy and faster classification speed,and its performance on GPU embedded systems also proves that it can meet both accuracy and classification Speed requirements.(3)A target detection algorithm for marine animal target detection is proposed.Based on the "two-stage" target detection algorithm faster-RCNN and the "single-stage" target detection algorithm SSD,we propose an improved algorithm based on the MoblenetV2 network.We will use MobilenetV2 convolutional neural network to replace the original faster-RCNN and VGG16 network in SSD for feature extraction,use the fusion of RPN layer and multi-scale feature in the two algorithms for target area frame calibration,and use the parameters in Chapter 3 Perform initialization.The experimental results show that the m-ssd algorithm is more suitable for marine animal target detection in terms of accuracy and real-time performance.(4)A light weighted pruning M-YOLOV3+ target detection algorithm is proposed.In order to have a better detection speed in the GPU embedded system and ensure the real-time detection,a target detection algorithm based on improved YOLOV3 is proposed in this paper.This method replaces the original darknet53 feature extraction network with mobilenetv2 network,and improves the target location network to obtain the m-yolov3 + network model and prune it.Compared with the original M-YOLOV3 network and the pruned network,the pruned M-YOLOV3+ network model can obtain the best detection effect under the condition of balancing the detection accuracy and speed.And on the GPU embedded system,the detection of msrcr enhanced image can better meet the requirements of real-time.Compared with m-ssd detection algorithm,pruning M-YOLOV3+ network is more suitable for processing images with smaller target size,and M-SSD network is more suitable for processing images with partial occlusion.