DPG-Unet for reflections removal on cigarette box surface

The variety of cigarette box printing inks, diverse patterns, and extensive use of reflective laser materials pose significant challenges for quality inspection. Currently, cigarette box quality inspection relies on slow and subjective manual methods, falling short of accurate and objective standards. However, while digital inspection holds promise, it faces challenges from strong reflections during image capture, leading to data distortion. Aiming to address this concern, we propose a network named Dual-stream Pyramid Gated-SwinUnet (DPG-UNet). Firstly, the dual-stream interaction mechanism enhances information exchange between features at different scales. Secondly, the pyramid model more effectively captures multiscale features, improving reflection removal accuracy. Finally, the proposed Gated-SwinUnet architecture, which significantly enhances feature representation capabilities by integrating the gating mechanism with the SwinUnet model. Our design uses the adjustment effect of the gating mechanism to optimize the interaction between the reflection layer and the projection layer, achieving deep fusion and effective separation of features. The experimental outcomes confirm the effectiveness of our approach significantly enhances the ability to remove reflections from cigarette box surfaces, providing higher-quality image inputs for subsequent analysis and automated inspection.