Detection of powdery mildew on strawberry leaves based on DAC-YOLOv4 model

Abstract

Strawberry powdery mildew (PM) is the main disease affecting the yield and quality of strawberries in recent years, which always appears on the back side of leaves in the early stage. Traditional methods of disease detection are labor-intensive and time-consuming. In this paper, we proposed a computer vision algorithm for strawberry leaf PM and infected leaves (IL) detection in complex background. Then we additionally proposed the estimation index of strawberry leaf PM for disease assessment. The original YOLOv4 backbone and neck are replaced by the proposed backbone and neck with depthwise convolution and hybrid attention mechanism, and the improvement can be made to decrease the size of the model and retain the performance. By combining the proposed backbone and neck, four new network structures are designed and evaluated, and the best one was named DAC-YOLOv4. Compared with YOLOv4, the mean average precision (mAP) of DAC-YOLOv4 reaches 72.7%, while the size is greatly compressed. To confirm the effectiveness of the proposed model, we compare DAC-YOLOv4 with five algorithms, and experimentally show that DAC-YOLOv4 performs well. We also deploy the algorithm on the Jetson Xavier NX and Jetson Nano, and the speed of DAC-YOLOv4 is 43 and 20 FPS, respectively, which can meet the real-time detection requirements. In summary, the experimental results indicate that the DAC-YOLOv4 proposed in this paper has good performance in strawberry leaf PM detection on the embedded platform, and the method to attain the disease index provides a solution for the early detection and prevention of strawberry PM.