基于LID-YOLO的小目标昆虫轻量化检测算法

Light weight detection algorithm for small target insects based on LID-YOLO

摘要:

针对复杂背景下新疆棉田昆虫识别误检率高、小目标昆虫难以检测等问题,提出了一种基于YOLOv5s改进的LID-YOLO(Lightweight Insect Detection-YOLO)轻量化昆虫检测模型。首先,主干网络使用GhostNet网络替换原CSPDarknet53网络,并采用Slim-Neck模块对颈部网络进行改进,以实现模型轻量化；其次,引入BottleNet Transformer融合模块,减少模型参数量并增强网络特征提取能力,更好的检测小目标；最后,加入NAM注意力机制,通过应用权重稀疏性惩罚抑制不显著权重来提取细节特征,提高模型准确率。实验结果表明,LID-YOLO模型在参数量、计算量、模型权重大小方面,相比YOLOv5s模型,其分别减少了30.9%、45.6%和29.7%。LID-YOLO模型的准确率达到了97.4%,检测速度为55.25FPS,与原YOLOv5s模型相比,提高了1个百分点和2.62FPS。LID-YOLO模型在保证轻量化的同时进一步提高了检测精度,更好满足农作物昆虫实际检测的需要。

Abstract:

Aiming at the problems of low Insect identification accuracy and difficult Detection of small target insects under the complex background of cotton field in Xinjiang, a Lightweight insect detection-YOLO (LID-YOLO) detection model based on YOLOv5s was proposed. Firstly, GhostNet network is used to replace the original CSPDarknet53 network in the backbone network, and Slim-Neck module is used to improve the neck network to achieve lightweight model. Secondly, the fusion module BottleNet Transformer is introduced to reduce the number of model parameters and enhance the capability of network feature extraction to better detect small targets. Finally, NAM attention mechanism is added to extract detail features by applying sparse weight penalty to suppress non-significant weights to improve model accuracy. The experimental results show that compared with the YOLOv5s model, LID-YOLO model reduces the parameters, calculation amount and model weight by 30.9%, 45.6% and 29.7% respectively. The accuracy rate of LID-YOLO model reached 97.4%, and the detection speed was 55.25FPS, which was 1 percentage point and 2.62FPS higher than that of the original YOLOv5s model. LID-YOLO model not only ensures lightweight, but also further improves the detection accuracy to better meet the needs of actual detection of crop insects.