改进YOLOv7的快递包裹检测算法  

Improved YOLOv7 for enhanced express package detection

作　　者：赖素晖 贾振堂 LAI Su-hui;JIA Zhen-tang(College of Electronics and Information Engineering,Shanghai University of Electric Power,Shanghai 200120,China)

机构地区：[1]上海电力大学电子与信息工程学院,上海200120

出　　处：《计算机工程与设计》2025年第2期537-545,共9页Computer Engineering and Design

基　　金：国家自然科学基金项目(62105196)。

摘　　要：针对物流分拣中堆放混乱、相互遮挡的问题,研究改进基于YOLOv7的目标检测算法,提高分拣效率和识别准确率。通过优化主干网络,引入ConvNeXt网络结构、轻量级上采样算子CARAFE,以及轻量级特征融合金字塔结构Slim-Light,减少模型参数和计算复杂度;引入轻量化无参注意力机制模块(SimAm)和新的边界框相似性度量标准归一化Wasserstein距离(NWD),解决密集堆叠包裹的误检漏检问题。实验结果表明,改进后的YOLOv7模型在自制数据集上取得了显著成果,在检测精度方面取得了显著提升,实现了模型参数和计算负担的双重降低,为解决物流分拣难题提供了一种有效方法。To address the challenges posed by disorderly and obstructed stacking in logistics sorting,improvements were made to the YOLOv7-based object detection algorithm,aiming to enhance sorting efficiency and recognition accuracy.The backbone network was optimized,and the ConvNeXt network structure,along with the lightweight upsampling operator CARAFE and the Slim-Light lightweight feature fusion pyramid structure,were introduced to reduce model parameters and computational complexity.A lightweight and parameter-free attention mechanism module(SimAm)and a new bounding box similarity metric,normalized Wasserstein Distance(NWD),were introduced to address the issue of false positives and negatives in densely stacked parcels.Experimental results show that the improved YOLOv7 model achieves remarkable results on the homemade dataset,which not only achieves a significant improvement in detection accuracy,but realizes a double reduction of the model parameters and computational burden,providing an effective method for solving the logistic sorting problems.

关 键 词：物流快递 智能分拣 密集包裹 目标检测 无参注意力机制 轻量化 边框相似性度量 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]