Research And Implementation Of Bad Image Detection Based On Deep Learning

With the rapid development of the information age,the Internet has become one of the most convenient information dissemination tools.While the Internet has brought great convenience to people,various violent images and adult images have also spread.In the real world,violent images and adult images have great harm to the physical and mental health of children and adolescents.Although people have been aware of this hazard,and also adopted manual screening and computer vision technology and other methods to solve.However,due to the challenges of huge network content,tedious work,and strong subjectivity,traditional methods have been unable to solve the increasing data.In recent years,with the rapid development of deep learning,the use of deep learning methods to detect bad images has become the mainstream and has achieved good results.In the research of this work,this paper found that there is currently no method and corresponding model for combining violence images and adult images for detection.The reason is that there is currently no experimental data set that combines violence images and adult images.Second,because the definition of violent images is relatively vague,there is no clear definition of violent images,and finally there are many categories of adult images.It is difficult to detect violent images in combination.In order to use deep learning methods to simultaneously detect violent images and adult images on the network,this paper has done a lot of related research and comparative experiments.The main research contents and contributions are as follows:1)Constructed a data set containing violent images,adult images and benign images at the same time,which contains nearly 330,000 images in total.Moreover,the data composition has a good diversity,which provides a good benchmark for the training and comparison of neural network model.2)Improving the residual network,accelerating the convergence of the network by improving the activation function,adjusting the position of the BN and the activation function,and constructing a branch structure,embed the SEnet subnetwork,and perform dynamic channel direction feature correction on the improved network.3)A deep learning network model MIDnet for bad image detection is proposed.It isdesigned based on the improved residual network structure and has a very good generalization ability for bad image detection including violence and adult types.4)This paper uses the MIDnet model to compare with the current mainstream violence and adult image detection methods.The experimental results show that the accuracy of the MIDnet model has obvious advantages,and achieves 98.2% accuracy on the proposed comprehensive data set.