Abstract
Cigarette manufacturing requires high-speed, high-precision quality control to ensure product integrity. Traditional anchor-based object detection models often struggle with the irregular shapes and small scales of cigarette defects, such as wrinkles or tiny spots. This paper presents C-CenterNet, an enhanced anchor-free detector specifically optimized for industrial cigarette inspection. By incorporating a Feature Pyramid Network (FPN) for multi-scale feature fusion, Deformable Convolutions (DCN) for geometric adaptability, and a modified Convolutional Block Attention Mechanism (CBAM), the proposed model achieves a mean Average Precision (mAP) of 95.01%. Experimental results demonstrate that C-CenterNet significantly improves detection accuracy for complex defects while maintaining a real-time inference speed of 8.9 ms per image.
Introduction
In the high-speed environment of tobacco manufacturing, automated visual inspection is critical. Cigarettes are long, narrow objects produced at blistering speeds, making manual inspection unstable and expensive. Traditional digital image processing often lacks the generalization required to handle varying factory conditions.
Deep learning has revolutionized object detection, yet standard models like Faster R-CNN or SSD face challenges in this domain. Cigarette defects—including tiny spots ("dotted"), wrinkles ("folded"), or misalignments ("malposed")—are often small and irregular. These geometries do not align well with the rigid rectangular "anchor boxes" used by most detectors. To address these limitations, this research proposes C-CenterNet, an anchor-free approach that treats defects as points rather than boxes, allowing for greater flexibility and precision.
Methodology
The authors selected CenterNet as the baseline architecture due to its efficiency in treating objects as single points (centers) and regressing properties like width and height. To adapt this for the specific rigors of cigarette inspection, the architecture was upgraded to include a ResNet-50 backbone coupled with a Feature Pyramid Network (FPN).
The FPN is essential for industrial defect detection because small defects often vanish during standard down-sampling. By merging high-level semantic features with low-level precise shapes, the FPN ensures that minute "dotted" defects are preserved throughout the network layers.
Architectural Innovations
C-CenterNet introduces several key modifications to the standard residual blocks to enhance feature representation and geometric adaptability:
- Deformable Convolution (DCN): Standard convolutions use fixed grids, which are poorly suited for irregular "folded" or "malposed" defects. DCN allows the network to adaptively shift its receptive field to match the actual shape of the defect.
- CBAM with h-Sigmoid: The Convolutional Block Attention Mechanism helps the network focus on relevant spatial and channel-wise features. The authors replaced the standard Sigmoid with h-Sigmoid to reduce computational overhead without sacrificing performance.
- ACON Activation Function: Instead of the rigid ReLU, the Activate-Or-Not (ACON) function adaptively determines neuron activation, allowing the model to learn more complex representations.
Experimental Results
The model was evaluated on a custom dataset of 10,000 augmented industrial images, covering categories such as Normal, Dotted, Folded, Malposed, and Unfiltered.
C-CenterNet achieved a 95.01% mAP, representing a 6.14% improvement over the baseline CenterNet. Notably, the model achieved a precision of 99.89%, which is vital for industrial applications where false positives lead to unnecessary waste. While YOLOv5 remains faster (4.8 ms vs. 8.9 ms), C-CenterNet provides the superior accuracy required for the irregular geometries found in cigarette defects, while still operating well within the requirements for real-time production line speeds.
Conclusion
C-CenterNet demonstrates that anchor-free detection, when combined with adaptive convolutions and attention mechanisms, is highly effective for specialized industrial tasks. The model's ability to handle irregular shapes and small-scale features makes it a robust solution for cigarette quality control. Future work will focus on model compression and deployment on embedded hardware, such as NVIDIA Jetson, to facilitate direct integration into factory floor systems.
Summary updated