Amoral Digital Images Detection Using Deep Dual Attention Network and Inception

The ever-increasing growth of amoral content in cyberspace has caused the idea of detecting such contents in images, then creating an alarm or limiters in different age ranges around the world. This content not only detects harmful visual items, but also it detects instances of abusing individuals. This study has represented a new approach to detect images with amoral content. The suggested model includes two deep models based on inception blocks and attention and residual based convolutional layer which acts with two different approaches: the first model is a five-class recognition model and the second one is an estimator model which maps the image into levels of inappropriateness between 0 and 1. Combining these two models by an aggregator based on first-order rules leads to representing a model which improved 2.1% detection precision in not suitable for work (NSFW) dataset and 2.3% on TI-UNRAM dataset in comparison with other state-of-the-art models. Our approach also shows promising results to model culturally based definition of NSFW, especially in Iran.