Multi-category sensitive image recognition based on RefCA-EfficientNetV2
Miao Yu, Dingju Zhu, Kai Leung Yung, W.H. Ip · International Journal of General Systems · 2025
In an era where the Internet has become a vital component of our daily existence, an unforeseen explosion of information has surfaced, including sensitive images containing pornographic, political, and horror themes. This content, threatening our well-being and polluting the Internet environment, compels the necessity for effective filtering and identification measures. Existing multi-category sensitive image detection methods, however, grapple with issues like scale, significant inference time, and unsatisfactory accuracy. In response to these challenges, we present RefCA-EfficientNetV2, a novel method built on the EfficientNetV2 model. This innovative solution enhances channel correlations and incorporates Coordinate Attention, thereby refining spatial coordinate information for eased region localization. Demonstrating marked accuracy improvement, our method attains a remarkable 97.64% accuracy level on sensitive images. With minimal parameter increase and time, RefCA-EfficientNetV2 not only enhances multi-category image accuracy but significantly reduces computational amount, offering a robust framework for cyberspace governance.