Abstract
Counterfeit Protection System (CPS) examination is as a crucial method for both brand discrimination and individual identification of colour laser printers. Current CPS recognition methodologies for colour-laser-printed documents primarily rely on manual inspection, leading to low efficiency and suboptimal accuracy. To address this limitation, this study developed an improved ShuffleNet_OD_CA model based on the lightweight convolutional neural network ShuffleNet V2 to enhance CPS recognition across different printer brands. Utilising printed documents from eight dominant colour laser printer brands (collectively representing over 95% of the market share), a dedicated dataset was constructed. This dataset was used to train the proposed CPS recognition model for colour-laser-printed documents, followed by comprehensive testing. The enhanced ShuffleNet_OD_CA model effectively identified CPS patterns in colour-laser-printed documents, achieving a recognition accuracy of 91.18% on the test set. The model exhibited 1.82 million parameters and 80.3 million FLOPs, which were fewer than those of the baseline (pre-improvement) model. Furthermore, compared to classical network models such as ResNet, the improved ShuffleNet_OD_CA model not only achieved higher recognition accuracy but also requires fewer parameters. This analysis demonstrated the capability of the model to extract and analyse key image features of the CPS.