Deep Learning Based Scylla Measurement And Detection Model

Phenotype data and monitoring of molting Scylla were all measured manually.It caused higher requirements for the operator,personal injury and Scylla damage.Research on breeding and fattening of Scylla for different farming scenarios has raised the need for more accurate and robust measurement methods.Due to the excellent performance of artificial intelligence in aquatic research,this thesis uses a deep learning approach to propose a high-precision keypoint-based ph enotype data measurement model and molting detection model for Scylla.Deep learning is based on a huge dataset.There are often no enough datasets in practical projects.This thesis is devoted to the study of the design of Scylla dataset,the improvement and training of the low-shot model.The main contributions of this thesis are as follows:1.An efficient design of Scylla dataset is proposed.In the keypoint detection dataset,the existing labeling software is improved.It can reduce the low-error-point rate and increase the labeling speed.The self-detection area algorithm is proposed by combining the method of image preprocessing and cluster analysis.It can eliminate invalid information.With these algorithms and data enhancement based on constraint information,the accuracy index PCK＿0.15 of Scylla is increased by nearly 15 percentage points.In the target detection dataset,the data is generated using the filtered inter frame difference method.It can generate highly differentiated data with less manual participation which means it will reduce the cost of labeling and training.2.The improvement of Hourglass model is proposed based on the existing keypoint model.The high-precision keypoint detection model of Scylla is based on channel attention mechanism.It achieved a high-precision keypoint detection of PCK＿0.15 up to 96.04%.And it combines with multi-frame strategy to achieve the phenotype data measurement model of Scylla.3.The improvement of YOLO v4 model are proposed based on the existing target detection model.The prediction head and Non-Maximum Suppression method of YOLO v4 are redesigned respectively.The priority-based multi-head Scylla molting detection model is proposed.With a hierarchical sampling training method based on data distribution,the precision rate of final model reached 98.64% and the recall rate reached 100%.Tested in video streams,the temporal overlap bewteen model detection results and manual discrimination results reached 99%.