Lightweight detection method for microalgae based on improved YOLO v7
WU Zhigao, CHEN Ming*
Key Laboratory of Fisheries Information, Ministry of Agriculture and Rural Affairs, College of Information Technology, Shanghai Ocean University, Shanghai 201306, China
Abstract: Traditional microalgae detection methods rely on a large number of manual operations with complex equipment, which are time consuming and the results easily influenced by the detectors’ knowledge and experience. A lightweight real-time microalgae detection method, YOLO v7-MA, was proposed by clustering anchor frames with K-means++ algorithm based on YOLO V7 model. In this method, GhostNet was introduced into YOLO v7 model as the backbone feature extraction network, which reduced the parameters of the network. At the same time, the ordinary convolution block in the feature fusion network was replaced by the deeply separable convolution block, which further decreased the computational complexity of the model, and added CBAM attention module to the feature fusion network to improve the feature expression ability of the network. The results showed that the YOLOv7-MA model had mean average precision of 98.56%, increased by 0.95%; the recall rate of 96.88%, increased by 1.15%; the F1 score of 97.42%, increased by 0.23%; the parameters quantities of 22.64×106, decreased by 14.63%, and the floating-point operations number of 38.45×109, decreased by 66.55% of the original compared with YOLOv7. Compared with FasterRCNN-VGG16, FasterRCNN-Resnet50, YOLOv4, YOLOv4-Mobilenet v3, YOLOv4-VGG16, YOLOv4-Resnet50, and YOLOv5s models, the mean average precision of YOLOv7-MA model was also improved, and the numbers of parameters were decreased. The findings indicate that it seems that the YOLO v7-MA model could provide a lightweight, real-time and efficient detection method for the identification and classification of microalgae, greatly reducing the workload of the detectors.