Pornographic Image Recognition via Weighted Multiple Instance Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Image Recognition

Background:

• Accurate recognition of pornographic images is vital for safeguarding children online.
• Identifying explicit content is challenging due to the small size and localized nature of key pornographic features within images.

Purpose of the Study:

• To develop a robust region-based model for pornographic image recognition using a multiple instance learning (MIL) approach.
• To enhance the accuracy of detecting explicit content by considering the degree of pornography within image regions.

Main Methods:

• Modeling images as a 'bag of regions' and employing a multiple instance learning (MIL) framework.
• Developing a quantitative measure for a region's degree of pornography to weight region importance.
• Formulating the recognition task as a weighted MIL problem within a convolutional neural network (CNN) framework, incorporating a bag probability function.

Main Results:

• The proposed method effectively generates relevant image regions for recognition, even with minimal annotations.
• A quantitative measure for region pornography degree aids in weighting important regions within positive images.
• The weighted MIL approach using CNNs achieved 97.52% true positive rate at 1% false positive rate on a large dataset (100K pornographic, 100K normal images).

Conclusions:

• The developed region-based MIL method significantly improves pornographic image recognition accuracy.
• The approach is effective in identifying explicit content, even when key features are small and localized.
• This technique offers a promising solution for online child protection by enhancing the detection of harmful visual content.