YOLO-TFD: An efficient object detecting model for automatic monitoring of the degree of fermentation of tea

Tea is a popular beverage with significant economic value, but its production process has a low level of intelligence and is mainly monitored manually, which affects the economic benefits. In order to solve this problem, this paper proposes a system to classify different fermentation levels of tea based on computer vision. The system consists of two stages. In the first phase, the YOLO v7 detection model is improved by adding densely connected transformer encoders at the tail of the backbone, which improves the extraction of key features in the target area. Meanwhile, the MobileOne module is used instead of the E-ELAN module of the YOLO v7 backbone to reduce the inference speed of the model. BiFPN is selected instead of PA-FPN to improve the feature fusion performance further and reduce the computational complexity. For the feature of mutual occlusion in tea leaf greening images, in order to capture a broader range of background information, the structure of SPPCSPC in YOLO v7 is changed, the maximum pooling layer running in parallel is changed to serial operation, and the global average pooling layer and the global maximum pooling layer are introduced, in order to reduce the loss of crucial semantic information that may be caused by relying on the edge information only and ignoring the background information. In the second stage, the detected tea leaves are graded on the basis of their color characteristics by the RGB model for the degree of fermentation of the tea leaves. The experimental results show that The mAP, precision, and recall of YOLO-TFD are higher by 6.2%, 8.6%, and 6% than YOLO v7. In addition, the model was deployed in PyQt5 for real-time detection and grading of tea during fermentation. This method provides a new solution for enhancing the intelligence of the tea production process and has potential industrial applications.