Enhanced ageholonet algorithm using age estimation and objectionable image for pornographic image detection

Abstract

With the rapid growth of the internet and emerging technologies, developing media content, and sharing them globally have become simple and fast. Despite the abundance of advantages this phenomenon has brought, it has led to some concerns in exposing people to unwanted and offensive media content. Among unwanted images, objectionable images are the most offensive ones which people are trying to avoid viewing. Although a number of research have been conducted in this area, this field is still scarce and there are challenges that should be addressed. One major challenge in this field is the lack of a well-defined definition for objectionable images. Therefore, different scholars with varied perceptions of the objectionable image came up with algorithms to tackle the problem of detecting objectionable images. In this research, the objectionable image detection model which is called Holistic Local Aware Deep Network or in short HoLoNet has the following novel characteristic: the local and global features are seamlessly integrated into the network and mutually affect each other during training. Moreover, in order to include the age of humans in the image of final decision, Gender Aware Age Estimation Net or in short GeAeNet was proposed. GeAeNet estimates age under condition of identified facial attribute of gender which makes the estimation more accurate. Moreover, the loss function is proposed to supervise the GeAeNet. Using this loss function, the network tends to generate a more reasonable probability distribution of age classes, where the predicted probability of each age class should be inversely proportional to the deviation from the ground truth age class in general. The combination of HoLoNet and GeAeNet formed the proposed AgeHoLoNet excluding the False Positive (FP) cases wherein detected objectionable images would only be humans who are under adulthood borderline age. GeAeNet outperformed state-of-the-art techniques in both controlled and wild environments by achieving Mean Absolute Error (MAE) 2.43 in facial age estimation dataset (MORPHII) and 2.64 in facial aging dataset (FG-NET) and 5.12 in Age Database (AgeDB) datasets. Finally, comparing the objectionable model with state-of-the-art techniques proves that HoLoNet alone outperforms related works with accuracy of 0.956 and AgeHoLoNet with accuracy of 0.964 over Pornography Dataset (NPDI).