基于YOLO 11n – ACE的轻量化荔枝病虫害检测方法

Litchi Disease and Pest Detection Method Based on Lightweight YOLO 11n – ACE

摘要:

针对自然环境下荔枝病虫害目标具有多尺度以及背景复杂等问题,本研究提出一种基于YOLO 11n – ACE的轻量化荔枝病虫害检测方法。该方法基于YOLO 11n算法,采用C3k2 – AC模块替换原来的C3k2模块提取荔枝病虫害图像特征,降低了模型参数量和计算存储负载;在此基础上进一步将检测头替换为轻量化检测头Detect Efficient,在特征提取和参数量上优化,使得模型在面对复杂环境时,能够准确定位和识别荔枝病虫害,在减小参数量的同时提高检测精度;最后,针对荔枝病虫害目标密集和尺度不一的问题,采用MPDIoU作为损失函数,最小点距离的引入使得损失函数能够更准确地评估预测框的质量,提高模型的检测性能。试验结果表明,改进后的YOLO 11n – ACE在试验数据集上的精确率和召回率达到88.27%和76.44% ,对比原始模型分别提高了2.54、0.97个百分点,mAP达到83.03% ,在提高检测精度的同时参数量减小为2.186 × 10?,浮点运算量降低1.3 × 10?。与主流的目标检测模型相比,改进后的模型减少了漏检、误检,精确率、检测精度和参数量均处于优势,为现代果园环境下的移动设备的荔枝病虫害检测提供了技术参考。

Abstract:

With the popularization of smart agriculture, in response to the problems of multi-scale and complex background of pest and disease targets of litchi in natural environments, a lightweight detection method for litchi pests and diseases was proposed based on YOLO 11n – ACE. This method was based on the YOLO 11n algorithm and used the C3k2 – AC module to replace the original C3k2 module to extract the features of litchi pest and disease images, reducing the number of model parameters and computational storage load. On this basis, the detection head was further replaced with the lightweight Detect-Efficient head to optimize feature extraction and parameter quantity, enabling the model to accurately locate and identify litchi pests and diseases in complex environments, improving detection accuracy while reducing the number of parameters. Finally, to address the issues of dense and diverse scales of litchi pest and disease targets, the MPDIoU was adopted as the loss function. The introduction of the minimum point distance enabled the loss function to more accurately evaluate the quality of prediction boxes and improve the detection performance of the model. Experimental results showed that the improved YOLO 11n – ACE achieved a precision and recall rate of 88.27% and 76.44% on the experimental dataset, respectively, which were 2. 54 and 0. 97 percentage points higher than that of the original model. The mAP reached 83.03% , and the number of parameters was reduced to 2.186 × 10? while the floating-point operations were reduced by 1.3 × 10?. Compared with mainstream object detection models, the improved model reduced missed detections and false detections, and had advantages in precision, detection accuracy, and parameter quantity, providing a technical reference for litchi pest and disease detection on mobile devices in modern orchard environments.