Research On The Sensitive Content Recognition Methods Of Two Specific Images

The main research contents of this paper are the recognition of red head document images and the identification of military uniforms in military images.With the increasing popularity of the Internet and the promotion of paperless office,many units and departments use computers to process official documents.And in recent years,the number of users of mobile Internet has been increasing,and smart phones have almost become portable items for people.Some paper documents containing important information are easily stored in pictures by means of photographing and scanning.Paperless office brings convenience but also security risks,and these pictures can easily leak into the network environment.At the same time,due to the convenient use of smart phones,military personnel are also very easy to access the Internet.In the "Ten prohibitions on the PLA's strict prevention of online leaks",there is a clear requirement that images that may contain classified information or video cannot be published in We Chat and other online chatting tools,and images wearing military uniforms cannot be taken as We Chat avatars.The spread of military uniform pictures on the Internet is a major challenge to the military's secrecy work.In this paper,image processing technology and deep learning-based image recognition technology are used to detect red-headed file images and military pictures transmitted in the network.The main work of this paper is summarized as follows:(1)Extracting the color features based on the HSI color space,and extracting the title feature area and the official seal feature area of the red head file in combination with the Hough transform,and using the OCR technique to identify the text of the title feature area.(2)A simplified triplet loss is proposed and the softmax loss is combined to construct a CNN model to classify the feature areas of the official seal.The CNN model adopts a two-level cascade.The purpose of the first-level CNN network is to exclude the non-official seal images that were mis-extracted in the original image.The second-level CNN network finally classifies the official seal: party emblem,national emblem and five-pointed star.(3)Based on the MTCNN algorithm combined with the HSI and YCb Cr dual color space skin color model to achieve rapid face positioning,and further extract the military uniform area in the military image.A military-based identification method based on cascaded multi-task convolutional neural network is proposed.The first-level CNN network model identifies whether the clothing subject is military uniform,excludes non-armed images.The second-level CNN network model is a multi-tasking CNN,which classifies military uniforms.(4)A simplified quadruplet loss is proposed and the softmax loss is combined to construct a CNN model to classify the military feature areas.Simplified quadruplet loss is improved based on simplified triplet loss,with the goal of further reducing intra-class distances and expanding inter-class distances.(5)A threshold needs to be defined in simplified triplet loss and simplified quadruplet loss,which is difficult to determine during training.A method for adaptively selecting a threshold based on a training model is proposed,and this threshold is used to select a triple and a quad that are suitable for the training model.The experimental results show that the official seal identification method combined with the simplified triplet loss can effectively identify the official seal existing in the red-headed document,and can also obtain a better recognition effect for the official seal with partial missing.The filtering of the skin color model combined with the dual color space can effectively filter the mis-detected face region of the MTCNN.The CNN model combined with the simplified quadruplet loss performs well on the military classification task,and is less affected by the standing posture and ambient illumination.Contrast experiments were established on the official seal classification task and the military uniform classification task.The experimental results show that the CNN model combined with the simplified triple loss and the CNN model combined with the simplified quadruplet loss have good effects in classification.