Intelligent Identifications and Filtering of Unconventional Images Based on Deep Neural Networks

Currently vast improvement of internet access and significant growth of web based broadcasters have resulted in distribution and sharing of informative resources such as images worldwide. Although this kind of sharing may bring many advantages, there are certain risks such as access of kids to porn images which should not be neglected. In fact, access to these images can be a threat to the culture of any society where kids and adults are included. However, many of internet users are members of social websites including Facebook or Instagram and without an appropriate intelligent filtering system, presence of few unconventional images may result in total filtering of these websites causing unpleasant feeling of members. In this paper, an attempt was made to propose an approach for classification and intelligent filtering of unconventional images. One of the major issues on these occasions is the analysis of a large scale of data available in the websites which might be a very time consuming task. A deep neural network might be a good option to resolve this issue and provide a good accuracy in dealing with huge databases. In this research, a new architecture for identifying unconventional images is proposed. In the proposed approach, the new architecture is presented with a combination of AlexNet and LeNet architecture that uses convolutional, polling and fully-connected layers.
The activation function used in this architecture, is the Rectified Linear Unit (ReLU) function. The reason of using this activation function is the high speed of convergence in deep convolution networks and simplicity in implementation. The proposed architecture consists of several parts. The first two parts consist of convolutional layers, ReLUs and pooling. In this section, convolution is applied to the input image with different dimensions and filters. In the next section, the convolutional layer with ReLU is used without pooling. The next section, like the first two parts, includes convolutional layers, ReLU and pooling. Finally, the last three parts include the fully-connected layers with ReLU. The output of the last layer is the two classes, which specifies the degree of belonging of each input to the class of unconventional and conventional images. The results are tested on a large-scale dataset. These tests show that the proposed method is more accurate than the other methods recently developed for identifying unconventional images.