基于改进YOLOv8的野生菌分类方法

Wild Fungi Classification Method Based on Improved YOLOv8

摘要:

针对野生菌种类繁多，传统的人工识别准确率低、效率低下的难题，提出了一种基于YOLOv8的野生菌分类新方法；引入C2f-MSBlock卷积，更好地获取野生菌的多尺度特征，减少了计算成本；引入多头自注意力机制模块，避免模型陷入局部最优从而提升分类模型的精度；针对传统检测头难以有效捕捉野生菌在菌伞和斑点等细节上差异的问题，提出了一种不对称双检测头的方案；针对边界框回归中出现的损失问题，引入了Inner-CIoU损失函数，使得模型可以灵活调整边界框的尺度大小，减少了边界框损失。实验结果表明，在野生菌分类任务中，相较于野生菌领域最优的3DRe-YOLO算法，改进后方法使mAP@0.5提升了0.3%，精确率提升了1.2%，召回率提升了0.3%，验证了改进的有效性。

Abstract:

A novel method for wild fungi classification based on YOLOv8 is proposed to address the challenges posed by the diverse range of wild fungi species and the low accuracy and efficiency of traditional manual identification methods. The approach introduces the C2f-MSBlock convolution to enhance the extraction of multi-scale features from wild fungi while reducing computational costs. Additionally, a multi-head self-attention mechanism is incorporated to prevent the model from converging to local optima, thereby improving classification accuracy. To better capture fine-grained details such as fimbriae and spots, an asymmetric dual detection head scheme is proposed, overcoming the limitations of traditional detection heads. To address the issue of loss in bounding box regression, the Inner-CIOU loss function is employed, enabling flexible adjustment of bounding box scales and reducing regression loss. Experimental results demonstrate that, compared to the best 3DRe-YOLO algorithm in the field of wild fungi, the proposed method achieves a 0.3% improvement in mAP@0.5, a 1.2% increase in accuracy, and a 0.3% improvement in recall, validating the effectiveness of the proposed enhancements.