基于改进Faster R_CNN的苹果叶片病害检测模型

在实际条件下，苹果叶片病害图像背景复杂且病斑较小，难以进行实时检测。针对该问题，提出一种改进的Faster R_CNN模型。通过特征金字塔网络将具有细节信息的浅层特征和具有语义信息的深层特征融合，以提取丰富的苹果叶片病害特征。同时采用精确感兴趣区域池化，避免感兴趣区域池化中2次量化操作对病斑较小的苹果叶片病害造成像素偏差。实验结果表明，该模型能对自然条件下5种苹果叶片病害进行有效检测，平均精度均值达82.48%，与Faster R_CNN、YOLOv3和Mask R_CNN模型相比，其平均精度均值分别提高了6.01、14.12和5.06个百分点。

Apple Leaf Diseases Detection Model Based on Improved Faster R_CNN

In practice,it is difficult to detect apple leaf diseases in real time due to the complex background of apple leaf disease images and the small disease spots.To address the problem,an improved Faster R_CNN model is proposed for apple leaf disease detection.Through the Feature Pyramid Networks(FPN),the shallow features with detailed information and the deep features with semantic information are fused to extract the rich features of apple leaf disease. At the same time,the Precise Region Of Interest Pooling(PrROI Pooling) is adopted to avoid the pixel deviation caused by the two quantization operations in the Region Of Interest Pooling(ROI Pooling) to the small disease spots of apple leaf diseases.The experimental results show that the model can effectively detect five kinds of apple leaf diseases under natural conditions at an average precision of 82.48%.Compared with Faster R_CNN,YOLOv3 and Mask R_CNN,the proposed model increases the average accuracy by 6.01,14.12 and 5.06 percentage points respectively.