目的 针对当前卷烟烟组端面缺陷检测技术无法准确识别具体缺陷类型，以及未兼顾识别滤嘴变形或夹沫的局限性，加之在检测精度上存在不足，本文引入一种基于轻量化YOLOv7模型的检测策略。方法 在GDX2包装机上设计并安装图像采集系统，通过集成轻量化网络结构ShuffleNetV2、DWConv、DSConv对YOLOv7架构进行定制化改造，有效减少模型参数量，加速训练与推理过程;结合SimAM注意力机制进一步增强模型对轻微缺陷特征的关注能力，显著提升轻微外观缺陷识别的准确性。结果 实验结果显示，所提算法参数量下降85%，计算量下降81%，mAP@0.5值达到0.973，缺陷类型中滤嘴变形和夹沫的平均精度分别达到0.978和0.945。在低算力(11.15 TFLOPS)环境下，该系统的推理时间低至1.98 ms。结论 该技术已经应用于GDX2型卷烟包装机烟组端面缺陷检测，能够有效提升缺陷检测的综合性能，为烟草行业提供兼具高效性与经济性的外观质量控制解决方案。

In view of the fact that the current detection technology for the end face defects of cigarette sets cannot accurately identify specific defect types, does not take into account the limitations of identifying filter nozzle deformation or smoke foam inclusion, and has insufficient detection accuracy, the work aims to introduce a detection strategy based on lightweight YOLOv7 model. An image acquisition system was designed and installed on the GDX2 packaging machine, and the YOLOv7 architecture was customized by integrating lightweight network structures ShuffleNetV2, DWConv and DSConv, which effectively reduced the number of model parameters and accelerated the training and reasoning process. Combined with the SimAM attention mechanism, the model's ability to pay attention to minor defect features was further enhanced, and the accuracy of minor appearance defect recognition was significantly improved. The experimental results showed that the number of parameters and calculation amount of the proposed algorithm decreased by 85% and 81%, and the value of mAP@0.5 reached 0.973. The average accuracy of filter deformation and smoke inclusion among defect types reached 0.978 and 0.945, respectively. At low computing power (11.15 TFLOPS), the system's reasoning time was as low as 1.98 ms. This technology has been applied to the detection of end face defects of GDX2 cigarette packaging machine, which can effectively improve the comprehensive performance of defect detection and provide an efficient and economical appearance quality control solution for the tobacco industry.