A Detection Systems For Molting Scylla Paramamosain Based On YOLO v4

For a long time, most data acquisition and data processing of crab breeding were collected manually. This paper designs a molting crab detection system from production of data set to model training. We proved an efficient data set generation method with less manual participation. By using the local maximum value of the inter-frame difference with a mean filter and data enhancement, greatly enhances the discrimination of the data. It also reduces the cost of labeling and model training. Through the stratified sampling method based on the number of target and the overlap rate, a data set for the monitoring of molting Scylla is effectively constructed. The data division is effective and the model verification results are more persuasive. It makes easier to monitor the effectiveness of training. The improved NMS maximum suppression method is used to effectively improve the model's ability to capture the crab molting state. The model precision rate of molting crab reaches 98.64%, and the recall rate reaches 100%.