基于YOLO-V5l与ResNet50的农田害虫检测

Farmland Pest Detection Based on YOLO-V5l and ResNet50

目前,我国农田受虫害影响日渐严重,虫情分析可以针对不同区域的农田虫情状况,制定不同的治理农田害虫方案。传统的虫情分析靠人工收集与统计,耗时耗力,随着深度学习技术在计算机视觉领域的发展,本文提出结合YOLO-V5l目标检测与ResNet50神经网络搭建农田害虫检测模型。昆虫在图像数据中呈现时具有体态多样、鳞片缺失、肢体脱落等特点,对目标检测与分类的影响较大,因此本文将28种害虫按照体态,颜色等进行粗分类为A~G七种后,利用YOLO-V5l模型对其进行检测与计数,再将检测结果代入ResNet50识别模型中确定其种类。这种方法极大降低了农田害虫检测的误检率。并且,本文提出一种预测增强算法,对待检测害虫图像进行增强后,分别带入识别模型中,对识别的结果取其加权平均,得到最终结果。单一的YOLO-V5l模型的mAP.5:.95为71.4%,平均精确率80.91%,漏检率5.39%。本文提出的虫情检测模型其平均精确率为89.56%,提升了对农田害虫的识别准确率。该模型将改善原始人工统计的缺点,推进我国智慧农业的发展。

At present, China’s farmland is increasingly affected by insect pests. Insect situation analysis can formulate different plans to control farmland pests according to the insect situation in different regions. Traditional pest situation analysis relies on manual collection and statistics, which is time-consuming and labor-consuming. With the development of deep learning technology in the field of computer vision, this paper proposes to build a farmland pest detection model by combining YOLO-V5l target detection and ResNet50 neural network. Insects have the characteristics of diverse body shapes, missing scales and falling limbs in the image data, which have a great impact on the target detection and classification. Therefore, this paper roughly classifies 28 pests into seven A~G species according to their body shapes and colors, uses YOLO-V5l model to detect and count them, and then substitutes the detection results into ResNet50 recognition model to determine their species. This method greatly reduces the false detection rate of farmland pest detection. Moreover, this paper proposes a predictive enhancement algorithm. After the pest images to be detected are enhanced, they are brought into the recognition model respectively, and the recognition results are weighted average to get the final results. mAP.5:.95 of single YOLO-V5l model was 71.4%, the average accuracy rate was 80.91%, and the missed detection rate was 5.39%. The average accuracy of the pest detection model proposed in this paper is 89.56%, which improves the recognition accuracy of farmland pests. The model will improve the shortcomings of original artificial statistics and promote the development of Intelligent Agriculture in China.