Method for identifying tea diseases in natural environment using improved YOLOv5s

Tea is easily affected by diseases in the growing process, leading to the decline of tea yield and quality. The conventional visual judgement on the disease cannot fully meet the large-scale production in recent years, due to the low accuracy, time-consuming and laborious. Therefore, an accurate knowledge of tea diseases is in high demand for timely and effective prevention and control measures, in order to reduce the abuse of pesticides for the stable development of tea industry. The existing research is focused mainly on the classification and recognition of tea diseases. However, it is difficult to achieve an accurate detection of tea diseases in different periods, especially early diseases. The reason is that there is a complex background in tea images taken in the natural environment, together with the small early spots of tea diseases. In this study, the YOLOv5-CBM model was proposed to recognize tea diseases in the natural environment. Firstly, a C3 module of YOLOv5-CBM was integrated with a Transformer and a Coordinate Attention (CA) mechanism into the feature extraction network in the stage of backbone feature extraction, in order to realize the extraction of important features of diseases. Secondly, the weight of each scale feature was adjusted adaptively using the weighted bidirectional feature pyramid (BiFPN) as the Neck of the network. The network was also better integrated with the features of different sizes for better recognition accuracy. Finally, a small target detection head was added to the detection end to reduce the missing detection in the early stage of tea disease. The disease data set was collected from the tea garden of Babu Purple Tea in Tielu Village, Wangmo County, Qianxinan Prefecture, Guizhou Province. Three common diseases were contained, namely Pestalotiopsis theae, tea anthracnose and Tea blister blight. The number of samples was 6,092 after data enhancement. The training set, verification set and test set were randomly divided, according to the ratio of 8:1:1. The experiment was conducted to verify the improved network. The pytorch 1.9.1 deep learning framework was selected to evaluate the Accuracy, Precision, Recall, AP and mAP. Firstly, the weights were initialized using the YOLOv5s pre-trained model, and the data set was then trained, according to hyperparameters. Then an ablation experiment was designed to verify the effectiveness of the added modules in the improved model. Finally, three main target detection models were selected for the comparative test. The performance of the improved model detection was further verified. The results showed that the improved YOLOv5s model significantly improved the detection of early disease spots in a natural environment. The AP value of early tea cake disease and early tea wheel spot disease increased by 1.9 and 0.9 percentage points, respectively, compared with YOLOv5s. The mAP detection of different diseases reached 97.3%, with a detection speed of 8ms/ sheet. Both models were superior to other target detection models. The improved model realized the detection and classification of three kinds of tea diseases. The detection accuracy and speed can provide a strong reference for the intelligent diagnosis of tea diseases.