Additional Learning on Object Detection: A Novel Approach in Pornography Classification

Hoang-Loc Tran, Quang-Huy Nguyen, Dinh-Duy Phan, Thanh-Thien Nguyen, Khac-Ngoc-Khoi Nguyen, Duc-Lung Vu

Abstract: In this paper, we proposed a new approach for pornographic classification by recognizing sensitive objects on images. To handle the misdetection and wrong judgment, a novel training strategy named additional learning was developed to help object detection model learns from mistakes, therefore increasing the method performance. Furthermore, a separate SVM classifier was trained to classify pornography and benign images from sexual object detected using Mask R-CNN model. Benchmarked by the NPDI-800 dataset, our proposed method achieved an accuracy of 84.625% and 90.125%, before and after applying additional learning strategy respectively. Besides, our proposed model also improves the false positive rate from 22.16% to 3.56% in our manually collected dataset.