Detection method for tea leaf blight in natural scene images based on lightweight and efficient LC3Net model

The discovery and localization of tea leaf blight (TLB) are crucial for tea-producing areas to resist the disease in good time and enhance tea yield and quality. TLB images captured in natural scenes present complex issues such as lighting variations, partial occlusion, and small spots. Traditional manual and machine learning-based methods cannot meet the requirements of detection accuracy and speed. To address these issues, a detection method for TLB in natural scene images based on a lightweight and efficient LC3Net model is proposed. In this method, the Retinex algorithm is chosen to enhance the contrast and minimize the influence of uneven lighting. The LC3Net model can detect leaves with different morphologies while reducing the computational cost. It is achieved by implementing image normalization (IN) and reducing the down-sampling frequency of the backbone network. The method can effectively detect mildly diseased and partially occluded leaves due to the integration of a channel attention module (CAM) and a weighted bi-directional feature pyramid structure in the LC3Net model. The experimental results indicate that the LC3Net model achieves an AP value of 92.29%, which is 5.46% and 4.91% higher than the results of Faster-RCNN and EfficientDet-D0 models, respectively. When compared to the baseline model YOLOv5s, our proposed LC3Net improves by 2.56% over the test data. Furthermore, the model has a parameter size of only 2.16M and achieves a detection speed of 63.5 frames per second (FPS), which provides a good solution for TLB detection in natural scene images.