Research On Lightweight Underwater Object Detection Method Based On Deep Learning

Underwater object detection is one of the most challenging research topics in current computer vision technology.Traditional exploration for marine environment mainly relies on human diving operations,and all tasks related to underwater object detection and recognition must be completed manually by oceanographers.Obtaining more information is not only inefficient and difficult,but also requires more divers.In the deep-sea environment,this long-term exploration is not practical.The vigorous development of computer vision technology enables humans to develop marine ecological resources in a non-intrusive way by using underwater detection equipment,such as remotely operated vehicle(ROV)and autonomous underwater vehicle(AUV).In the underwater object detection task,the traditional object detection methods usually perform poorly in terms of accuracy and generalization ability.Underwater object detection needs to study accurate,stable,real-time and lightweight detection models,so many deep learning-based underwater object detection methods have been proposed and widely used in various scenarios of the marine environment,such as aquaculture,pelagic fisheries,marine species monitoring,underwater archaeology,etc.However,underwater object detection faces many challenges.The complex underwater environment has high requirements for underwater imaging equipment,which makes it more difficult to obtain underwater images compared with atmospheric optical images.Underwater images are usually affected by unbalanced lighting conditions,blurring,low contrast and color deviation,resulting in poorer quality of color and texture information than atmospheric optical images.Some underwater living objects are often clustered and small,so it is difficult to extract rich detail information,resulting in poor model detection effect.The class imbalance of underwater targets makes it difficult for underwater object detection model to learn the characteristics,which has a negative impact on its performance.In addition,in the past,underwater object detection technology mainly used large detection networks to improve detection accuracy.However,lightweight and real-time requirements are crucial in practical applications.The lightweight of underwater object detection model has greater engineering practical value to a certain extent.Therefore,from the perspective of designing a lightweight underwater object detection network,this paper improves the detection effect of underwater small targets and occluded targets,and achieves a great balance between accuracy and speed.At the same time,the proposed model has a lower number of parameters and storage space occupation,which is more suitable for further deployment.Specifically,this paper realizes the lightweight of underwater object detection model from different aspects.First,the network structure of YOLO v4 algorithm is reconstructed by using lightweight backbone network and efficient convolution operators,namely Mobile Net v2 and depth-wise separable convolution,to reduce the amount of network parameters and computation.At the same time,aiming at the problem that underwater small targets are difficult to detect,the AFFM module is used to enhance the feature pyramid structure to fuse the semantic information of high-level feature and the spatial information of low-level feature to improve the detection effect of small targets.On the basis of YOLO v4,the model parameters and model size are compressed to 16.76% and19.53% of YOLO v4,and the parameters are 10.73 M.The model reaches 81.67% m AP on PASCAL VOC dataset and obtains great performance on underwater image datasets,achieving a good tradeoff between the speed and accuracy of underwater object detection.Second,the above detection model is an anchor-based detection model,which requires a lot of adjustment of the anchor parameters,and the calculation is complex,and the anchor with fixed size and shape is difficult to apply to underwater targets with different shapes.This paper proposes the CA-Ghost PAN structure to improve the anchor-free detection network,namely Nano Det-Plus,to improve the detection accuracy of small targets.The ablation experiment shows that this structure can effectively improve the detection accuracy of small targets in DUO dataset.At the same time,the Hide-and-Seek data enhancement strategy is used to simulate the occlusion of underwater targets to improve the detection effect of occluded targets.This strategy increases the AP by 0.1%~0.2% without introducing additional overhead.In order to further reduce the number of parameters,this paper proposes a GAGM module,which has reduced the number of parameters by about 60%.The experimental results show that the method can realize the lightweight of underwater object detection model,achieve 77.8% m AP in the PASCAL VOC dataset,and achieve competitive results in the DUO underwater image dataset,that is,the AP reaches 53.9%,achieving the balance of speed and accuracy.The proposed methods provide a reference value for the future work of lightweight underwater object detection.