Adaptive texture and low-light feature learning in enhanced MobileNetV4 for industrial packaging quality inspection

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Automation

Background:

• Food packaging integrity inspection faces challenges due to subtle textural differences in defects and varying industrial lighting.
• Class imbalance and high similarity between normal and defective regions complicate automated detection systems.

Purpose of the Study:

• To develop an enhanced MobileNetV4 architecture for accurate and efficient seal defect detection in food packaging.
• To address limitations in texture discrimination, suboptimal illumination, and class imbalance in industrial inspection scenarios.

Main Methods:

• Implemented an enhanced MobileNetV4 with a novel LocalAttention module for improved texture analysis and Pinwheel-shaped Convolution (PConv) for low-light conditions.
• Integrated an Adaptive Similarity-Modulated Loss function to handle class imbalance and refine feature separation for similar defect categories.
• Utilized dynamically weighted sliding windows and Gaussian-inspired patterns for adaptive feature learning.

Main Results:

• Achieved 99.54% accuracy and 99.75% F1-score on a grain packaging dataset with only 3.7M parameters and 25.27 FPS on a Raspberry Pi 4B.
• Demonstrated robust performance under simulated LED degradation and sensor noise, outperforming baseline architectures.
• Showcased stable 8-hour edge deployment without thermal throttling and achieved 99.72% accuracy on a cross-domain NEU-CLS dataset.

Conclusions:

• The proposed MobileNetV4 framework effectively addresses challenges in food packaging seal defect detection, offering high accuracy and efficiency.
• The adaptive feature learning, specialized convolution, and loss function contribute to superior performance in complex industrial environments.
• The model's parameter efficiency and robustness suggest strong transferability to other industrial defect classification tasks.