Counterfeit packaging detection faces a critical dilemma in which rigorous authentication measures often compromise visual aesthetics, whereas unobtrusive micro-features frequently suffer from low detection rates due to their minute scale and background interference. To address this, we propose an enhanced YOLOv8 framework that harmonizes high-precision micro-feature recognition with computational efficiency suitable for edge deployment. Specifically, we introduce Mini-CBAM, a lightweight attention mechanism that substitutes standard multi-layer perceptrons with 1 × 1 convolutions to significantly reduce parameter redundancy while sharpening the focus on anti-counterfeiting regions. To mitigate the loss of texture details in deep convolutional networks, we design PA-FPN++, a multi-scale feature fusion architecture that establishes a high-resolution shortcut connection between shallow and deep layers. Furthermore, addressing the localization bias inherent in small target detection, we propose IoU-E Loss, a novel objective function incorporating an exponential penalty term to amplify gradient feedback during training. Complementing these architectural innovations, a GAN-based adversarial training strategy is employed to synthesize hard-negative samples, thereby enhancing model robustness against diverse forgery techniques. Extensive experiments on a multimodal dataset of 30,000 images demonstrate that our method achieves a mean Average Precision of 0.97, outperforming state-of-the-art baselines in micro-feature recognition while maintaining real-time inference speeds.
Jiang et al. (2026) studied this question.