Statistically optimised method for detecting adult image groups

Abstract

Adult image recognition is an important technique for preventing children from accessing offensive material on the Internet. Most of the related works focus on single image recognition. However, adult images usually exist as a group and rarely stand alone. Therefore, considering the entire image group as a whole for classification should be more effective. This paper presents a new method of detecting adult image groups, which aims at achieving optimal recognition accuracy. Adult image group recognition generally includes two components: an adult image recogniser and a final decision rule for classifying the image group. We provide a theoretical analysis to clarify the correlation of the two components and use probability formulae to estimate the recognition rates for different settings of the adult image recogniser and the decision rule. Then, a set of optimal receiver-operating characteristic (ROC) curves for different image numbers is solved. To recognise an unknown image group, a desired recall rate for adult (or benign) image groups is specified and the system is set according to the parameters acquired from the optimal ROC curves. The proposed method can be dynamically adapted to the recall rates that the user expects. This advantage makes the proposed system more suitable for real applications. Our work can be viewed as an extension of single image recognition and the experimental results demonstrate that it can attain higher recognition accuracy than the earlier methods.

Keywords:

• Adult image recognition
• Image group classification
• Roc curve
• Neural network