Ma Yukun, Wei Yajun, Ma Minsheng, Ning Zhilong, Qiao Minghui, Awada Uchechukwu
School of Software, Henan Institute of Science and Technology, Xinxiang, Henan, China.
School of Information Engineering, Henan Institute of Science and Technology, Xinxiang, Henan, China.
Front Plant Sci. 2024 Dec 19;15:1501043. doi: 10.3389/fpls.2024.1501043. eCollection 2024.
Pests are important factors affecting the growth of cotton, and it is a challenge to accurately detect cotton pests under complex natural conditions, such as low-light environments. This paper proposes a low-light environments cotton pest detection method, DCP-YOLOv7x, based on YOLOv7x, to address the issues of degraded image quality, difficult feature extraction, and low detection precision of cotton pests in low-light environments.
The DCP-YOLOv7x method first enhances low-quality cotton pest images using FFDNet (Fast and Flexible Denoising Convolutional Neural Network) and the EnlightenGAN low-light image enhancement network. This aims to generate high-quality pest images, reduce redundant noise, and improve target features and texture details in low-light environments. Next, the DAttention (Deformable Attention) mechanism is introduced into the SPPCSPC module of the YOLOv7x network to dynamically adjust the computation area of attention and enhance the feature extraction capability. Meanwhile, the loss function is modified, and NWD (Normalized Wasserstein Distance) is introduced to significantly improve the detection precision and convergence speed of small targets. In addition, the model detection head part is replaced with a DyHead (Dynamic Head) structure, which dynamically fuses the features at different scales by introducing dynamic convolution and multi-head attention mechanism to enhance the model's ability to cope with the problem of target morphology and location variability.
The model was fine-tuned and tested on the Exdark and Dk-CottonInsect datasets. Experimental results show that the detection Precision (P) of DCP-YOLOv7x for cotton pests is 95.9%, and the Mean Average Precision (mAP@0.5) is 95.4% under a low-light environments, showing improvements of 14.4% and 15.6%, respectively, compared to YOLOv7x. Experiments on the Exdark datasets also achieved better detection results, verifying the effectiveness of the DCP-YOLOv7x model in different low-light environments.
Fast and accurate detection of cotton pests using DCP-YOLOv7x provides strong theoretical support for improving cotton quality and yield. Additionally, this method can be further integrated into agricultural edge computing devices to enhance its practical application value.
害虫是影响棉花生长的重要因素，在复杂自然条件下，如低光照环境中准确检测棉花害虫是一项挑战。本文提出一种基于YOLOv7x的低光照环境棉花害虫检测方法DCP-YOLOv7x，以解决低光照环境下棉花害虫图像质量退化、特征提取困难和检测精度低的问题。
DCP-YOLOv7x方法首先使用FFDNet（快速灵活去噪卷积神经网络）和EnlightenGAN低光照图像增强网络增强低质量棉花害虫图像。目的是生成高质量害虫图像，减少冗余噪声，改善低光照环境下的目标特征和纹理细节。接下来，将DAttention（可变形注意力）机制引入YOLOv7x网络的SPPCSPC模块，动态调整注意力计算区域，增强特征提取能力。同时，修改损失函数，引入归一化瓦瑟斯坦距离，显著提高小目标的检测精度和收敛速度。此外，将模型检测头部替换为DyHead（动态头部）结构，通过引入动态卷积和多头注意力机制动态融合不同尺度的特征，增强模型应对目标形态和位置变化问题的能力。
该模型在Exdark和Dk-CottonInsect数据集上进行了微调与测试。实验结果表明，在低光照环境下，DCP-YOLOv7x对棉花害虫的检测精度（P）为95.9%，平均精度均值（mAP@0.5）为95.4%，与YOLOv7x相比，分别提高了14.4%和15.6%。在Exdark数据集上的实验也取得了较好的检测结果，验证了DCP-YOLOv7x模型在不同低光照环境下的有效性。
使用DCP-YOLOv7x快速准确地检测棉花害虫为提高棉花质量和产量提供了有力的理论支持。此外，该方法可进一步集成到农业边缘计算设备中，以提高其实际应用价值。
