卷烟包装盒外观瑕疵无监督细粒度视觉识别算法设计

Design of unsupervised fine-grained visual recognition algorithm for cigarette packaging box appearance defects

摘要:

卷烟生产线上包装盒外观瑕疵检测是保障产品质量的关键环节。随着生产线速度的提升,传统基于模板匹配或有监督学习的视觉检测方法面临着对微小瑕疵(如细微划痕、局部漏印)识别率低,且难以应对复杂反光材质和未知缺陷的问题。针对这些问题,为了提高卷烟包装质量检测的准确性和鲁棒性,本研究提出了一种基于无监督学习的细粒度视觉识别算法。首先,设计基于预训练卷积神经网络的特征提取模块,提取包装盒图像的多尺度语义特征。然后,引入坐标注意力机制强化高频局部特征,以适应细粒度瑕疵的识别。最后,利用马氏距离构建正常样本的多元正态分布模型,计算测试样本的异常得分以实现瑕疵定位。根据实验可知,该方法在无缺陷样本标注的条件下,对各类型卷烟包装瑕疵的检出率达到 96.5%,单张图像处理时间仅为 12ms,能够完美适配高速生产线的实时监测需求。

Abstract:

The design of an unsupervised fine-grained visual recognition algorithm for cigarette packaging defect detection. Packaging defect inspection on cigarette production lines is a critical quality assurance process. With increasing production line speeds, traditional visual inspection methods based on template matching or supervised learning face challenges in identifying minute defects (e.g., fine scratches, localized omissions) and handling complex reflective materials and unknown defects. To address these issues and enhance the accuracy and robustness of cigarette packaging quality inspection, this study proposes an unsupervised fine-grained visual recognition algorithm. First, a feature extraction module based on pre-trained convolutional neural networks extracts multi-scale semantic features from packaging images. Then, a coordinate attention mechanism is introduced to strengthen high-frequency local features, enabling fine-grained defect recognition. Finally, a multivariate normal distribution model of normal samples is constructed using the Mahalanobis distance, and anomaly scores are calculated for test samples to achieve defect localization. Experimental results demonstrate that this method achieves a 96.5% detection rate for various cigarette packaging defects under no-defect sample annotation conditions, with a single-image processing time of only 12ms, perfectly meeting real-time monitoring requirements for high-speed production lines.