A real-time defect detection system for cigarette appearance based on machine vision

Abstract

This paper introduces a real-time defect detection system for cigarette appearance based on advanced machine vision and deep learning techniques. The system then uses a specially designed convolutional neural network for multi-scale feature extraction, with adaptive image pre-processing to compensate for lighting changes and reduce sensor noise. A comprehensive dataset with expert annotations and targeted data augmentation was built in the production environment to address various types of defects and class imbalance issues. The model integrates structured pruning and quantization for efficient industrial deployment, and uses cross-entropy and focal loss functions to enhance the sensitivity of certain defects. The experimental results show that the system can maintain a real-time inference speed of more than 50 frames per second and maintain high detection accuracy even under challenging lighting conditions. Comparative evaluation results show that the proposed method outperforms traditional methods based on rules and baseline deep learning. The proposed method provides a reliable and scalable solution for automated quality control in cigarette manufacturing and establishes a technical framework to support future intelligent industrial inspection technologies. The modular architecture can also be directly used for other industrial vision inspection tasks, which highlights the wide applicability and practicality of the method in high-speed production processes.