Image Enhancement Effects On Adult Content Classification

Abstract

Adult content filtering is an essential part of digital media platforms. With the extensive usage of social media, it becomes harder to overcome this problem. Traditional methods consist of human supervision and standalone image processing techniques. These approaches are not accurate enough according to the massive size of the social media generated content. Also, the methods are not discriminative enough on a variety property of the images. Colour, shadow, frequency features of the images can vary, even the context is the same according to lumination features. The problem can be solved more accurately with deep learning techniques. Notably, the specific type of deep learning architecture called convolutional neural network is suitable for the problem space. In this study, the state of the art model has been used with transfer learning to test image enhancement effect on the success of the architecture. Colour vivid, sharpness value incrementation and histogram equalization approaches have been tested for adult content classification problems.