Authors

Abstract

The peach fruit moth was a fruit-eating pest and one of the major pests of fruit trees in China, Korea, Japan, and Australia. Due to long-term problems such as improper control methods, low technical quality, and untimely treatment, the yield and efficiency of fruit products were greatly affected, which constrained the development of the fruit industry. This paper developed a method for detecting adult peach fruit moths based on an improved YOLOv8m to address the challenging problem of manually detecting peach fruit moths. To increase the Receptive Field of the model, v7Down Sampling was introduced in its backbone network. Then, the channel-prioritized Convolutional Attention Mechanism Module (CPCA), which dynamically allocated the spatial attention weights on each channel, reducing the noise and the algorithm’s complexity, was incorporated. Finally, the inner-WIoU loss function was introduced to enhance the convergence and generalization of the bounding box. The precision (P) of the improved model increased by 3.4 percentage points compared to YOLOv8m. The recall (R) improved by 2.1 percentage points, and the mAP improved by 1.2 percentage points. The single-category precision (AP) for peach fruit moth detection improved by 2.4 percentage points. Moreover, the weight size, number of model parameters, and computational volume were reduced by 3.6MB, 1.8M, and 1.7G, respectively. This achieved an improvement in the model's effectiveness in detecting adult peach fruit moths without increasing the model's complexity. The results provided strong technical support for the subsequent real-time monitoring of the peach fruit moth.

Abstract in Chinese

桃小食心虫是一种食果害虫，也是中国、韩国、日本、澳大利亚等果树的主要害虫之一。因为长期的防治方法不当、技术素质低、处理不及时等问题，使果品的产量和效益都受到很大影响，制约果业的发展。本文针对人工检测桃小食心虫困难问题，开发了一种基于改进YOLOv8m的桃小食心虫成虫检测方法。我们在其主干网络引入v7Down Sampling，增加模型的Receptive Field。然后引入通道优先卷积注意力机制模块（CPCA），动态分配各个通道上的空间注意力权重，减少了噪声及算法的复杂度。最后引入Inner-WIoU损失函数，增强了边界框的收敛和泛化能力。改进后模型的精确度P相较YOLOv8m提高了3.4个百分点。召回率R提高了2.1个百分点。mAP提高了1.2个百分点。桃小食心虫单类别精度AP上提高了2.4个百分点。并且权重大小、模型参数量和计算量分别减少了3.6MB、1.8M、1.7G。实现了在不增加模型复杂度的同时提高模型对桃小食心虫成虫的检测效果，其结果可为后续桃小食心虫的实时监测提供有力的技术支撑。