A global and local context integration DCNN for adult image classification. (December 2019)

• Record Type: Journal Article
• Authors: Cheng, Feng; Wang, Shi-Lin; Wang, Xi-Zi; Liew, Alan Wee-Chung; Liu, Gong-Shen
• Is Part Of: Pattern recognition. Volume 96(2019:Dec.)
• Publication Date: 2019-12
• Subjects: Adult image recognition -- Deep convolutional network -- Global context -- Local context -- Multi-tasks learning; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2019.106983
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We propose a DCNN with multi-tasks learning for adult image classification. Our network seamlessly integrates the global and local contexts. It achieves excellent performance in three-class adult image classification. Abstract: With the wide availability of the Internet and the proliferation of pornographic images online, adult image detection and filtering has become very important to prevent young people from reaching these harmful contents. However, due to the large diversity in adult images, automatic adult image detection is a difficult task. In this paper, a new deep convolutional neural network (DCNN) based approach is proposed to classify images into three classes, i.e. porn, sexy, and benign. Our approach takes both the entire picture (global context) and the meaningful region (local context) information into consideration. The proposed network is composed of three parts, i.e. the image characteristics subnet to extract discriminative low-level image features, the sensitive body part detection subnet to detect adult-image related regions, and the feature extraction and fusion subnet to generate high-level features for image classification. A multi-task learning scheme is designed to optimize the network with both the global and local information. Experiments are carried out on two datasets with over 160, 000 images. From the experiment results, it was observed that the proposed network achieved high classification accuracies (96.6% in the AIC dataset and 92.7% in the NPDI dataset) and outperformed the other approaches investigated.