Refined Trademark Sticker Detection Method Using Minimum Pooling Features *

Trademark sticker detection stands as a crucial visual task within industrial quality inspection scenarios, de-manding precise determination of the accurate positioning of adhered trademarks on products. Despite the remarkable suc-cess achieved by current deep learning networks, their training requirements often involve extensive datasets, posing significant challenges for direct application in trademark sticker detection systems. A major obstacle lies in the vast variety of trademarks encountered in real-world industrial settings, coupled with the rapid speed of trademark updates and iterations. This renders it impractical to gather samples of every trademark type for training purposes. Furthermore, the process of collecting and annotating industrial data samples can be both time-consuming and costly. To overcome these limitations, this paper introduces a novel approach that utilizes image subtraction to analyze the differences between the image with adhered trademarks and the image without adhered trademarks. This approach circumvents the need for vast sample sizes and extensive data annotations. By integrating minimum pooling techniques into the analysis of these difference images, this paper introduces an efficient approach that effectively filters out noise and alleviates the generalization challenges arising from frequent trademark replacements. Experimental findings underscore the robust flexibility of our method in precisely and adaptively detecting and recognizing packaging box labels. Consequently, this approach significantly enhances the quality of product packaging and trademark sticker detection, offering a viable solution for addressing the challenges posed by the diverse and rapidly evolving nature of industrial trademarks.