Research On Fish Detection Method Based On Deep Learning

Object detection is widely used in intelligent transportation,image retrieval and analysis and intelligent robot.As a practical application of object detection,fish shoal detection is of great value for detecting the regularity of fish-school activity in lakes and oceans and analyzing the size and species of fish-school.Due to the diversity of the environment and the small size of the fish,the traditional feature extraction method has a low accuracy and a high time complexity,so that the research of fish detection algorithm faces a great challenge.In this paper,the fusion detection method based on binocular vision fish images is studied by improving SSD algorithm.The main research work in this paper is as follows:(1)Realizing the real time fish-school detection method based on end-to-end deep learning models.In this paper,the end-to-end YOLO(You Only Look Once)algorithm and SSD(Single Shot Multibox Detector)algorithm are used to extract image features to detect fish-school,which are verified in the underwater fish dataset named Labeled fish in the wild.The result shows that using YOLO model and SSD model to extract the fish features can detect the fish-school more accurately.The detection accuracy reaches 88.7% and 92%,and the real-time detection speed of 40 FPS and 30 FPS can be maintained respectively.The recall rate of SSD algorithm is about 30% higher than that of YOLO,which can reach 42.7%,indicating that SSD algorithm has a significant improvement over YOLO in the detection of small target fish.(2)Proposing an improved SSD algorithm.Aiming at the problem of missed detection caused by the convolution neural network in the extraction of small fish-school,a feature pyramid with the help of multiple convolution layers extracted by SSD model is constructed.High-level features are used to enhance the semantic expression ability of the lower level by fusing multi-layer features,which can improve the detection accuracy of small size fishschool.The underwater fish dataset named Labeled fish in the wild is used for verification.The detection accuracy and recall rate are improved by 2.7% and 30% respectively compared with the original SSD algorithm,and the processing speed is 22 FPS.(3)Proposing a fish detection method based on binocular vision.The traditional binocular image splicing method is prone to artifacts.In this paper,the improved SSD algorithm is used to extract the image features of the binocular fish-school.Based on the receptive field characteristics of the neural network,the feature matching twin tower model and the back propagation mechanism are established to map the high-level feature matching results to the low-level level step by step,which can limit the computational range of feature matching of each feature layer.The matching accuracy of the low-level feature layer can be improved.The underwater fish dataset named Labeled fish in the wild is used for verification.The proposed method and vgg16-based splicing method are used for fish fusion detection and achieve the accuracy of 90.3% and 80.4% respectively.It is shown that the proposed algorithm can significantly improve the mosaic and detection accuracy of low-contrast fish images.So that it can effectively suppress the appearance of splicing artifacts.(4)Realizing the fish detection system based on deep learning.In real application scenarios,this paper trains the underwater fish-school image detection model based on the popular deep learning open source framework named Tensorflow,and builds the fish-school detection system combined with the WCF framework to realize real-time fish-school detection.