Backbone Selection for Packaging Defect Classification: A Controlled Study of Accuracy, Robustness, and Computational Efficiency
Abstract
Automated packaging inspection requires classification models that are accurate, robust, and efficient enough for practical use. Although previous studies have widely explored defect detection, the selection of suitable CNN backbones for defect-type classification remains less investigated. This paper presents a controlled empirical comparison of four CNN backbones—ResNet18, EfficientNet-B0, MobileNetV3-Small, and ConvNeXt-Tiny—for seven-class packaging defect classification using a production-line dataset of 1,217 images, consisting of one non-defect class and six defect classes. To ensure a fair comparison, all models used the same dataset split, preprocessing pipeline, classification head, loss function, and training protocol. The evaluation considered three dimensions: classification performance on test images, robustness under five synthetic image condition changes, and computational efficiency under a shared GPU setting. ResNet18 achieved the best test performance, with 96.72% accuracy and a macro F1-score of 0.9663, while also providing the fastest inference time at 1.21 ms per image. EfficientNet-B0 showed the strongest robustness, with the lowest mean accuracy drop of 0.88 percentage points, and achieved the best overall ranking in the balanced production setting. MobileNetV3-Small had the smallest model size, with 0.93 million parameters and a 10.9 MB checkpoint, but it was the least robust under image condition changes. ConvNeXt-Tiny showed competitive accuracy and relatively strong robustness, but its higher inference latency on the evaluated GPU may limit its use in time-sensitive applications. Therefore, CNN backbone selection for industrial packaging inspection should consider not only test accuracy but also robustness and deployment efficiency.
Authors: Amonpan Chomklin, Siam Yamsaengsung, Pornchai Mongkolnam, Niwan Wattanakitrungroj