Abstract
Conventional methods of detecting packaging defects face challenges with multiobject simultaneous detection for automatic filling and packaging of food. Targeting this issue, we propose a packaging defect detection method based on the ECA-EfficientDet transfer learning algorithm. First, we increased the complexity in the sampled data using the mosaic data augmentation technique. Then, we introduced a channel-importance prediction mechanism and the Mish activation function and designed ECA-Convblock to improve the specificity in the feature extractions of the backbone network. Heterogeneous data transfer learning was then carried out on the optimized network to improve the generalization capability of the model on a small population of training data. We conducted performance testing and a comparative analysis of the trained model with defect data. The results indicate that, compared with other algorithms, our method achieves higher accuracy of 99.16% with good stability.
Highlights
To guarantee high-quality products from automatic packaging production lines, defect inspections are indispensable
Conventional packaging defect inspections are mostly made using equipment based on image processing techniques
A total of 1200 samples were collected with cap and label defects
Summary
To guarantee high-quality products from automatic packaging production lines, defect inspections are indispensable. To improve the model’s generalization capability, the model must be designed and the data must be processed Regarding the latter, data augmentation and transfer learning are effective techniques to improve model accuracy with small datasets. Hu et al [17] proposed the squeeze-and-excitation network (SE-Net), which makes predictions on channel importance during convolution to improve the overall accuracy of the model. We propose a fast packaging defect detection method based on the ECA-EfficientDet transfer learning algorithm. We incorporated the ECA mechanism in a backbone feature extraction network and designed an ECA-Convblock convolution block that is capable of predicting the channel importance during convolution This suppresses channels that carry no information, to make specific representations of object features and improve the defect detection accuracy.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
