Enhancing Image Content Analysis in B2C Online Marketplaces

Abstract

The automation of image analysis in Business-to-Consumer (B2C) online marketplaces is critical, especially when managing vast quantities of supplier-uploaded product images that may contain various forms of objectionable content. This study addresses the automated detection of diverse content types, including sexual, political, and disturbing content, as well as prohibited items like alcohol, tobacco, drugs, and weapons. Furthermore, the identification of competing brand logos and related imagery is examined for competition and ethical reasons. The research integrates custom transfer learning models with the established Microsoft and Google Vision APIs to enhance the precision of content analysis in e-commerce settings. The introduced transfer learning model, trained on a comprehensive dataset, exhibited a significant improvement in identifying and categorizing the specified content types, achieving a notable true positive rate that surpasses traditional API performances. The findings reveal that the “Pazarama Model”, with its transfer learning framework, not only delivers a more accurate and cost-effective content moderation solution but also demonstrates enhanced efficiency by reducing the image processing time and associated costs. These results support a shift toward specialized transfer learning models for content moderation, advocating for their adoption to maintain content integrity and enhance user trust within e-commerce platforms. The study advocates for continued refinement of these models, suggesting the integration of multimodal data to further advance the content analysis capabilities in B2C environments