基于改进YOLO v4的单环刺螠洞口识别方法

针对养殖池塘内单环刺螠自动采捕和产量预测应用需求，提出一种基于深度学习的单环刺螠洞口识别方法，以适用于自动采捕船的嵌入式设备。该方法通过将YOLO v4的主干网络CSPDarkNet53替换为轻量型网络Mobilenet v2，降低网络参数量，提升检测速度，并在此基础上使用深度可分离卷积块代替原网络中Neck和Detection Head部分的普通卷积块，进一步降低模型参数量；选取带色彩恢复的多尺度视网膜(Multi-scale retinex with color restoration，MSRCR)增强算法进行图像增强；利用K-means++算法对数据集进行重新聚类，对获得的新锚点框尺寸进行线性缩放优化，以提高目标检测效果。在嵌入式设备Jetson AGX Xavier上部署训练好的模型，对水下单环刺螠洞口检测的平均精度均值（Mean average precision，mAP）可达92.26%，检测速度为36f/s，模型内存占用量仅为22.2MB。实验结果表明，该方法实现了检测速度和精度的平衡，可满足实际应用场景下模型部署在单环刺螠采捕船嵌入式设备的需求。

Urechis unicinctus Burrows Recognition Method Based on Improved YOLO v4

In order to realize the real time detection of Urechis unicinctus burrows in the actual aquaculture pond scene, and provide support for the automatic harvesting and yield prediction of Urechis unicinctus, a deep learning based identification method of Urechis unicinctus burrows was proposed. In view of the limited storage space of the embedded equipment of harvesting vessel and high real time requirements for target detection, the YOLO v4 model had a large number of parameters and a slow detection speed. By replacing the backbone network CSPDarkNet53 of YOLO v4 with a lightweight Mobilenet v2 to reduce the amount of network model parameters and improve the detection speed. On this basis, depthwise separable convolution blocks were used instead of the normal convolution blocks in the Neck and Detection Head parts of the original network to further reduce the number of model parameters. For the poor quality of underwater images, the multi-scale retinex with color restoration (MSRCR) algorithm was selected for image enhancement. Finally, for the original anchor box obtained by clustering the COCO dataset, which was not suitable for small target recognition, the K-means++ algorithm was used to recluster the dataset and optimize the linear scaling of the obtained new anchor box size to obtain the most suitable anchor box for the dataset in order to improve the target detection effect. To simulate the automatic capture scene of Urechis unicinctus, a set of image acquisition equipment with an unmanned ship as the main body was built, and an image data set was established through the collected video to conduct experiments. The trained model deployed on the embedded device Jetson AGX Xavier can detect mean average precision (mAP) of underwater Urechis unicinctus burrows up to 92.26% with detection speed of 36f/s and model size of only 22.2MB. Experiments showed that the method achieved a better balance of detection speed and accuracy and can meet the demand of practical application scenarios where the model was deployed in the embedded equipment of the Urechis unicinctus harvesting vessel. It provided a reference for the subsequent automatic harvesting of Urechis unicinctus and yield prediction in breeding ponds.