基于改进YOLOv5s和迁移学习的苹果果实病害识别方法

为实现对苹果果实病害的快速准确识别，提出了一种基于改进YOLOv5s的果实病害识别模型：GHTR2-YOLOv5s （YOLOv5s with Ghost structure and TR2 module），并通过迁移学习策略对其进行优化。在YOLOv5s基础上通过加入幻影结构和调整特征图整体宽度得到小型基线模型，通过卷积块注意力模块（Convolutional Block Attention Module, CBAM）和加权双向特征金字塔网络（Bidirectional Feature Pyramid Network, BIFPN）提高模型精度，使用TR2（Two Transformer）作为检测头增强模型对全局信息的获取能力。改进后模型大小和识别速度为2.06 MB和0.065 s/张，分别为YOLOv5s模型的1/6和2.5倍；IoU阈值为0.5下的平均精度均值（mAP0.5）达到0.909，能快速准确地识别苹果果实病害。研究通过在线图像增强与迁移学习相结合的方式提高模型收敛速度，进一步提高模型精度，其mAP0.5达到0.916，较原始模型提升8.5%。试验结果表明，该研究提出的基于GHTR2-YOLOv5s和迁移学习的苹果病害识别方法有效优化了模型训练过程，实现了占用较少计算资源的情况下对苹果病害进行快速准确地识别。

Identification of apple fruit diseases using improved YOLOv5s and transfer learning

Apple diseases have frequently threatened the yield and quality of fruit production, leading to irreversible damage to the plant growth. Meanwhile, the treatments vary significantly in the types of diseases. Therefore, timely and accurate identification can be highly urgent to prevent the spread of the diseases. Traditional and manual identification is often time-consuming and laborious depending mainly on experienced experts. Fortunately, deep learning can open up a new way to control diseases. However, the general model of disease identification is still lacking, due to the complex real environment in the field and few large datasets of apple fruit diseases. In this study, a lightweight GHTR2-YOLOv5s (YOLOv5s with Ghost structure and TR2 module) and transfer learning-based model was proposed to identify the apple fruit diseases, particularly with the high accuracy and less complex architecture. The specific procedure was given as followed. Firstly, an image dataset of apple fruit diseases was collected in the field, including the bitter pit, anthracnose, ring disease, and fruit rust. A series of data enhancement operations were then adopted on the image dataset to avoid over-fitting for the convergence. Secondly, The Ghost structures (Ghost Conv and Ghost Bottleneck) were added to the backbone and neck network of YOLOv5s. A small baseline model was then obtained to adjust the width of the feature map, in order to reduce the storage occupation of the model for the high detection speed. The Convolutional Block Attention Module (CBAM) was adopted to assign the feature weights to the baseline model, where the effective features were selected after evaluation. The Bidirectional Feature Pyramid Network (BIFPN) was used to enhance the robustness and generalization ability. Two Transformer (TR2) encoder modules were stacked for the detection head to enhance the global information of the model. Thirdly, the transfer learning was obtained to improve the convergence speed and generalization ability, where the knowledge was firstly learned from the image dataset of apple leaves diseases, and then transferred to the GHTR2-YOLOv5s model in the disease identification of apple fruits. Finally, the performance of the model was verified by various experiments under different conditions. The experimental results demonstrated that: 1) The proposed GHTR2-YOLOv5s model reduced the model size for high accuracy. The model size and average identification speed of the GHTR2-YOLOv5s model were 2.06 MB and 0.065 s/sheet, which were 1/6 and 2.5 times that of the original YOLOv5s. The mean Average Precision under Intersection over Union (IoU)=0.5 (mAP0.5) reached 0.909, which was higher than that of the original. 2) Transfer learning improved the model accuracy, while shortening the convergence time of the model. The mAP0.5 of the model reached 0.916 by combining the online data enhancement and the secondary transfer learning, which was 8.5% higher than that of the original version. Consequently, the improved model can be expected to identify the apple fruit diseases rapidly and accurately, indicating higher accuracy and faster convergence with a less model size than before. The finding can provide a feasible and promising reference for the intelligent diagnosis of apple fruit diseases in apple production.