基于改进YOLOv5网络的复杂背景图像中茶尺蠖检测

茶叶的产量和品质深受病虫害的影响。茶尺蠖是一种常见的茶叶害虫，精确检测茶尺蠖对茶叶病虫害防治有重要意义。由于茶尺蠖和茶树枝、枯死茶叶的颜色、纹理相近，茶尺蠖的体积小、形态多变、被遮挡等问题，现有方法检测茶尺蠖的精度不高。该研究提出一种基于深度学习的复杂背景图像中茶尺蠖检测方法，该方法使用YOLOv5为基线网络，利用卷积核组增强对茶尺蠖的特征提取，在不增加计算量的条件下减少复杂背景对茶尺蠖检测结果的干扰；使用注意力模块关注茶尺蠖整体，根据茶尺蠖的大小和形状自适应调节感受野，降低因目标大小形状不一导致的漏检；使用Focal loss损失函数减少前景和背景的类不平衡对检测结果的影响。试验结果表明，所提方法用于复杂背景图像中茶尺蠖的检测，可以达到0.94的召回率，0.96的精确度和92.89%的平均精度均值。与基线网络相比，该方法的平均精度均值提高了6.44个百分点。使用相同的数据集和预处理的对比分析表明，该方法优于SSD、Faster RCNN和YOLOv4等其他经典深度学习方法，平均精度均值比SSD、Faster RCNN、YOLOv4分别高17.18个百分点、6.52个百分点和4.78个百分点。该方法可实现对茶尺蠖的智能检测，减少人力成本，有助于实现精准施药，提高茶叶的产量和品质。

Detection of Ectropis oblique in complex background images using improved YOLOv5

Abstract: Diseases and pests have posed a great threat to the yield and quality of tea in recent years. Among them, the Ectropis oblique is one of the most common pests in tea growth. A traditional detection has normally used the appearance of the pests, such as the color, morphology, and texture. But, these are more sensitive to the environments, particularly to the complex background, where the pests appear. A rapid and accurate detection cannot be realized, because: 1) The training samples are taken in different scales, while the pest is normally small in size; 2) The pest with the changeable shape and color may be shielded to obscure during imaging; 3) The color and texture of the pest can be similar to the tree branches and dead leaves of tea. Therefore, it is very necessary to identify and recognize the pest in a complex background in tea production. In this study, a rapid and accurate detection was proposed for the Ectropis oblique in complex background images using the improved YOLOv5 deep learning. Definitely, the YOLOv5 was taken as the baseline network. A labeling operation was first used to manually label the pest samples in the training and validation images. The data was then enhanced using the flipping, and contrast enhancement, particularly that the Gaussian noise was added to prevent data from overfitting. Meanwhile, the contrast of the test image was adjusted to reduce the influence of complex backgrounds, such as the tea pole on the detection of the scorpion. A convolution kernel group was also used to enhance the feature extraction without increasing the computation load. Furthermore, an attention module was utilized to adaptively adjust the receptive field, thereby enhancing the feature representation, according to the size and shape of the Ectropis oblique. More importantly, a Focal Loss function was used to reduce the impact of class imbalances between foreground and background during detection. The experimental results show that the convolution kernel group was effectively reduced the interference of complex background to the detection of tea geometrid. The attention module also presented an excellent performance to reduce the missed detection, due to the varying sizes and shapes of targets. Specifically, the best detection was achieved for the images with a complex background, where 0.94 recall, 0.96 precision, and 92.89% mean average precision. The improved accuracy increased by 6.44 percentage points, compared with the original YOLOv5. Moreover, there were 17.18 percentage points higher than the SSD, 6.52 percentage points higher than the Faster-RCNN, and 4.78 percentage points higher than the YOLOv4, compared with the SSD, Faster-RCNN, and YOLOv4. Consequently, the improved YOLOv5 can be widely expected to realize the intelligent monitoring of ectropis oblique pests in the precise pesticide application for the higher yield and quality of tea.