Research And Application Of Deep Weighted Multiple Instance Learning On Pornographic Image Recognition

With the rapid development of network technology,the network information appears explosive growth.The pornographic information has brought negative impact on the daily life of people especially teenager.The recognition of pornographic images is of great significance to the protection of teenager's physical and mental health.Because of the uncertainty of background,light,scale and human gesture,recognition of the pornographic images on the Internet is very difficult.Conventional methods based on regions of interest detect human skin area and extract features near the region of interest.However,skin area detection is difficult itself,and some pornographic images do not contain a large area of skin nudity.The method that based on global feature of image directly captures the local feature in image,and concentrated to obtain the feature representation.This method has limited performance because it cannot cover the pornographic content in the image and neglect the spatial relationship of the bad content.The human body part detectors obtain the semantic feature of the image by detecting the human semantic components.This method is very prone to false detection due to ignoring the context information of body parts and intensive scanning on the original image.We summarize the shortcomings of the existing methods,and propose a region-based deep weighted multiple instance learning method.The main work of this paper is as follows: Firstly,each image is regarded as a bag composed of multiple image regions,each image region has different contribution to the bag.Secondly,based on very few annotations of pornographic part,our method can generate a large number of image regions with different contributions.Meanwhile,we present a simple quantitative measure of each region's contribution,and calculate the contribution of each image region based on the proportion of pornographic part.Finally,we introduce the bag probability function to combine the instances with different contributions,and formulate the problem under the framework of deep convolution neural network and trained to obtain the recognition model.In addition,in order to verify the effectiveness of our algorithm,we collect a large-scale data set from the Internet and tested the performance of each algorithm on a test set containing 100,000 pornographic images and 100,000 normal images.Our deep weighted multiple instance learning algorithm achieve the best performance,achieving an accuracy with 97.52% true positive rate at 1% false positive rate.