基于改进YOLOv8的虫草小目标检测算法

Cordyceps Small Object Detection Algorithm Based on Improved YOLOv8

摘要: 针对青藏高原特殊环境下虫草检测面临的复杂高原背景、目标遮挡频繁、虫草形态细长且易与自然背景混淆等挑战，本文基于YOLOv8模型提出改进方法。首先，在可变卷积(Deformable Convolution)的基础上设计双层可变卷积(Double-layer Deformable Convolution)，建立双层动态卷积调整机制，利用特征偏移量自适应调整卷积核的大小和形状，提高特征饱和度，缓解下采样带来的信息失衡。其次，针对虫草因遮挡导致的漏检问题，融合空间增强注意力机制(SEAM, Spatially Enhanced Attention Module)，通过深度可分离卷积和残差模块增强未遮挡部分的语义特征，优化空间通道中的权重信息，有效提升模型对遮挡环境下的信息提取和检测能力。最后，引入新的检测头FASFF-head，以自适应学习多尺度特征图的空间权重，进行空间特征融合，确保多尺度特征的协调性，且在原有检测层之上添加小目标专用检测层，使得在高密草丛环境下，虫草的检测精度得到显著提升。以上实验表明，改进模型在自建藏区虫草数据集上的mAP@0.5和mAP@0.5:0.95对比YOLOv8模型分别提升4.2%和2.9%；在Flavia Dataset公开数据集上的实验结果可以发现，YOLOv8-DSEAM 除了参数量略高于YOLOv10n，mAP@0.5比YOLOv10n提高了1.3%，mAP@0.5:0.95比YOLOv10n提高了0.8%，充分地展现了改进后的模型在高密草丛场景下的检测力和泛化力。

Abstract: To address the challenges of caterpillar fungus detection in the complex plateau environment of the Qinghai-Tibet Plateau, including intricate high-altitude backgrounds, frequent target occlusion, and the elongated morphology of cordyceps that easily blends with natural surroundings, this paper proposes improvements based on the YOLOv8 model. First, we design a Double-layer Deformable Convolution building upon Deformable Convolution, establishing a dual-layer dynamic convolution adjustment mechanism. This utilizes feature offsets to adaptively adjust convolution kernel size and shape, enhancing feature saturation and alleviating information imbalance caused by downsampling. Second, to tackle missed detection due to occlusion, we integrate the Spatially Enhanced Attention Module (SEAM). Through depthwise separable convolution and residual modules, this enhances semantic features of unoccluded regions and optimizes weight information in spatial channels, effectively improving information extraction and detection capabilities in occluded environments. Finally, we introduce a novel FASFF-head detection head to adaptively learn spatial weights of multi-scale feature maps for spatial feature fusion, ensuring multi-scale feature coordination. Additionally, a dedicated small-target detection layer is added above the original detection layers, significantly improving detection accuracy in dense grassland environments. Experimental results demonstrate that the improved model achieves 4.2% and 2.9% increases in mAP@0.5 and mAP@0.5:0.95 respectively compared to YOLOv8 on our self-built Tibetan Cordyceps dataset. On the public Flavia Dataset, YOLOv8-DSEAM shows superior performance: while slightly higher in parameters than YOLOv10n, it improves mAP@0.5 by 1.3% and mAP@0.5:0.95 by 0.8%, fully demonstrating the enhanced detection capability and generalization power of our model in dense vegetation scenarios.