Abstract: The YOLOv8n model, which employs a Convolutional Neural Network (CNN) as its feature extraction core, typically imposes significant hardware demands on embedded target detection systems, thereby hindering effective detection under low-configuration conditions. In response to such issues, a lightweight improved algorithm based on YOLOv8n is proposed. Firstly, the DualConv module, utilizing both 3×3 and 1×1 convolutional kernels, is introduced to process input feature maps concurrently, with group convolution techniques optimizing the arrangement of convolutional filters to minimize model parameters and computational load. Secondly, the Lightweight Adaptive Downsampling Module (LADM) replaces standard convolution, effectively reducing parameters and computational volume while adaptively weighting features across different channels to preserve critical feature information. Additionally, the LSKA attention mechanism is integrated to refine the feature fusion layer, bolstering its multi-scale feature extraction capabilities. Lastly, the Lightweight Parameter Sharing and Detail-Enhanced Detection Head (LPSDED) leverages parameter sharing to efficiently predict the locations and categories of variously sized and shaped targets within images, substantially decreasing computational and parametric demands. Experimental results on the combined VOC07+12 dataset demonstrate that the enhanced YOLOv8n model achieves a 43.3% reduction in total parameters and a 37% decrease in computational load, with only a marginal 2% decline in detection accuracy. This algorithm successfully realizes model lightweighting, striking an optimal balance between model accuracy and computational efficiency.

Key words: YOLOv8, lightweight improvement, target detection, feature extraction, group convolution technique

摘要： 以卷积神经网络(CNN)为特征提取内核的YOLOv8n通常对嵌入式目标检测系统硬件要求较高，难以实现低配置下有效检测的目的。针对此类问题，提出一种YOLOv8n轻量化改进算法。首先引入双卷积核DualConv来构建轻量级深度神经网络，利用3×3和1×1卷积核同时处理输入特征图，并利用组卷积技术高效排列卷积滤波器，从而减少模型参数量与计算量；其次利用轻量型自适应下采样模块LADM替换普通卷积，保证在降低参数量与计算量的同时，适应性地权衡不同通道的特征，从而更好地保留重要的特征信息；再之引入LSKA注意力机制改进特征融合层，增强其多尺度特征提取能力;最后利用轻量型参数共享与细节增强检测头LPSDED检测参数共享的优势高效地预测图像中不同大小和形状的目标的位置和类别，有效降低了网络的计算量与参数量。实验结果表明，改进型YOLOv8n目标检测模型在VOC07+12联合数据集上实现了参数总量减少43.3%、计算量降低37%，同时检测精度仅下降2%的显著优化效果。该算法成功地实现了模型轻量化，在模型精度和计算效率之间取得了最佳平衡。

关键词: YOLOv8, 轻量化改进, 目标检测, 特征提取, 组卷积技术