基于改进YOLO-V4网络的浅海生物检测模型

海洋生物智能检测是海洋牧场战略的一部分，而利用水下机器人在复杂的海洋环境中快速、准确地检测海洋生物是关键问题。由于海底环境复杂、亮度分布不均匀、海洋生物与其生存环境的区分性差、生物被遮蔽或半隐蔽等原因，准确识别海洋生物是一个巨大的挑战。随着卷积神经网络的发展，基于深度学习的目标检测算法成为主流，出现了如EfficientDet、RetinaNet和YOLO-V4等典型算法。这些基于深度学习的算法都不是完全尽善尽美的，不能完全满足海洋生物识别的需求。在探测精度、运算速度、密集目标探测效果等方面都有提高的空间。该研究建立了一个海洋生物数据集，采集了原始图片1 810张，数据增强后得到7 240张图片，它们被分成训练集（80%）和测试集（20%）。其次，通过引入跨阶段局部网络的概念，构建了嵌连接EC（Embedded Connection）部件，并将其嵌入到YOLO-V4网络的末端，得到改进的YOLO-V4网络。最后，该研究提出了基于改进YOLO-V4网络的海洋生物检测模型MOD（Marine Organism Detection）。试验结果表明，MOD模型的mAP50、mAP75分别为0.969和0.734，计算量为35.328BFLOPs（十亿浮点运算数），检测帧速为139 ms（具有图形加速器GeForce GTX1650上）。与原始YOLO-V4模型相比，MOD模型的mAP50和mAP75提高了0.9和4.8个百分点，而计算量仅提高0.2%。此外，对比两种模型的PR曲线，MOD模型的精确度与召回率的平衡点更接近（1，1），因此MOD模型能学习精度和效率的平衡性更好。该研究直接面向浅海生物的目标检测问题，所提供的方法可以为水下机器人精准执行智能捕捞等任务提供有益参考。

Model for marine organism detection in shallow sea using the improved YOLO-V4 network

Abstract: Intelligent detection of marine organisms is a significant step of marine ranching strategy. An underwater robot is highly demanding to rapidly and accurately monitor marine organisms in the complex ocean environment. However, there is a relatively low distinction between marine organisms and their living environment, some of which are covered or semi-hidden, due mainly to the low contrast of seabed environment, and uneven distribution of brightness. Therefore, it is a big challenge to accurately identify the specific marine life in the undersea environment. Many target (object) detections using deep learning have emerged, such as EfficientDet, RetinaNet, and YOLO-V4, with the development of convolutional neural networks (CNN) in recent years. Nevertheless, the current network cannot fully meet the specific requirements of marine biological recognition. It is also necessary to improve the detection accuracy, operation speed, and detection efficiency of dense targets. In this study, an improved target (object) detection network using YOLO-V4 was designed to realize the rapid and accurate identification of marine organisms in an aquaculture environment of a shallow sea. A marine biological dataset was firstly established with 7 240 images, which were generated from 1 810 original images after data enhancement. Training (80%) and test datasets (20%) were divided. Data enhancement (suitable for the small data sample learning) effectively enriched the background and elements of the original images, thereby producing much more learning samples than before. As such, an effective expansion of the sample was achieved in the same learning effect as the large sample. Secondly, the cross-stage partial network (CSP) was successfully introduced, while the Embedded Connection (EC) component was designed to detect marine organisms. An improved YOLO-V4 network model was constructed, when the EC was embedded into the end of the YOLO-V4 network. The improved YOLO-V4 network with an EC can be expected to make the gradient flow propagate on different learning paths, while effectively delay the occurrence of gradient disappearance, aiming to improve the detection accuracy and cost-saving calculation. Finally, Marine Organism Detection (called MOD) was presented using the improved YOLO-V4 network to achieve a better performance in the complex seabed environments. The experimental results showed that the mAP50 and mAP75 of the MOD model were 0.969 and 0.734, respectively, while the computational complexity was 35.328 billion floating-point operations (BFLOPs), and the detection frame rate was 139 ms on the computer system with a graphics accelerator GeForce GTX 1650. The mAP50 and mAP75 from the MOD increased by 0.9% and 4.8%, respectively, while the amount of computation only increased by 0.2%, compared with the original YOLO-V4 model. Especially, the evaluating indicators in the MOD model improved in all studied categories, where mAP75 presented the most obvious. In addition, the precision and recall values of balance points in the MOD model were closer to (1, 1) in most cases. It can also be reasonable that the learning performance was better in the MOD than the original YOLO-V4 model, compare with the PR curves. Consequently, the finding can provide promising insightful ideas and useful references for the rapid and accurate detection of the marine organisms in an underwater robot of intelligent fishing.