Research On Specific Image Recognition Technology In Network Video

With the rapid development of the Internet,adopting internet as a carrier for information dissemination is faster and wider.Bad criminals use internet to spread the rapid characteristics,and a large number of inflammatory speech,pictures or videos are released through online social media.The illegal transmission of horror videos or images has become one of the current sources of pollution affecting social stability.How to block the spread of bad information through the Internet has become a hot topic of current research.This paper studies and summarizes the existing methods of blocking bad information,and proposes to detect the specific image content in the network video on the user side.For the detection of specific image content,the main research image contains "people who hold the gun and keep the shooting attitude".The research work of this paper is as follows:Analyze the existing object detection method and choose Faster R-CNN as the gun detection model.Based on the conventional object classification based on convolutional neural network,the model can also give the position information corresponding to the object,which is the follower and gun.The overlapping area detection provides a basis.In this paper,four types of guns are identified:AK47,M16,92,and 95.After collecting the raw data,each type of gun data set was expanded to approximately 3,600 sheets by the rotation increment technique.After training the Faster R-CNN model using different feature extraction networks,record each model to identify the recognition accuracy of each category of guns and record the mAP indicators of each model for comparison,and confirm the use of ResNet-50 network as the feature extraction network.The test results are better than the other two networks.Aiming at the application limitation of the existing depth camera-based human body motion angle feature extraction method,a complete human body bone angle based human body motion angle feature extraction method is proposed.According to the movement angle of the human body,the space is divided into six angles that are bilaterally symmetrical.After the six angles are preprocessed by data normalization and coordinate transformation,the corresponding data set is established.This paper chooses SVR as the baseline model to compare with the output of the regression neural network.Through the model comparison,it can be confirmed that the indexes of the output of the regression neural network are superior to the SVR model.This method has broad application prospects in somatosensory interaction and human motion recognition.Aiming at the problems existing in the application of the existing human motion recognition method to video images,this paper proposes to use the angular features of the regression neural network output to combine the bone data of the human arm to generate new feature vectors.Finally,two classification neural networks are constructed to train the new feature vectors to achieve the purpose of specific human motion classification.The experimental results show that the accuracy index and recall rate of the specific motion recognition of the right arm are 86.3%and 89.7%,respectively;the accuracy index and recall rate of the specific motion recognition of the left arm are 85.1%and 88.5%,respectively.However,limited by the richness of the data set and the accuracy of the bone data,when the untrained image is tested,the recognition rate is reduced.