Analysis Of Changes In Facial Features Of Drug Users’ Faces

Drugs can be said to be a sensitive topic of high concern in today’s society.Due to ignorance or curiosity,many people covet temporary pleasure,inject or take drugs,and have since fallen into the "swamp" of drugs.He lost his original working ability and was severely traumatized at the psychological level.Therefore,all drug-related violations of laws and disciplines(drug trafficking,drug use,etc.)need to be severely cracked down.With the development of face recognition and deep learning research,researchers have begun to try to apply various face recognition technologies to drug control operations.It is hoped that through the research on the face recognition detection technology of drug addicts,drug abusers and normal people can be screened out quickly and efficiently,so as to effectively strengthen the behavioral restraint and management control of drug addicts,thereby reducing the social impact they bring.At the same time,it can remind the general public to cherish life and stay away from drugs.At present,the research on facial recognition detection of drug addicts is still in its infancy in China,and no published literature on facial recognition technology for drug addicts faces can be retrieved.However,some scholars abroad have tried a variety of identification methods,such as dictionary learning identification based on binarized statistical image features(BSIF)and GIST feature descriptors,and network feature extraction based on scattering wavelets.However,the above methods all have a common feature,that is,they are all manual image feature extraction based on prior information.For example,some scholars replace the entire image with several important parts of the face and directly use them as training data.Considering that the effects of drugs on the faces of individual addicts are different,under the idea of replacing the overall features with a small part of the features,all the accuracy and robustness of the designed models are often disrespectful.In order to break through the bottleneck of robustness and accuracy,this paper explores how to use all the features of the image to identify and detect drug abuse faces and judge the correction time of drug addicts without preprocessing according to prior information.The main work of this paper includes:(1)Aiming at the problem of insufficient model training data,the method of data expansion is analyzed and studied,in order to reduce the overfitting phenomenon as much as possible.These methods mainly include adding Gaussian noise and salt and pepper noise to the original data,adjusting the brightness,and performing flip rotation.The processed data volume is 3-4times the original data volume.Finally,the validity of the data augmentation is verified by the comparison experiment of the fitting degree.The error value of the correct rate of the model training and testing in the expanded dataset is significantly smaller than the error value obtained by the experiment on the initial dataset.(2)Aiming at the structural problem that local image feature extraction based on prior information lacks satisfactory accuracy and robustness,a neural network model based on deep learning attention mechanism is proposed.The model refers to the hyperparameter settings of Res Net18,and embeds a mixed-domain attention mechanism in each residual module.Each mixed-domain attention mechanism is composed of spatial attention and channel attention modules,and finally is fully connected.The layer outputs the corresponding face results.This model has two advantages.First,it directly avoids manual image feature extraction based on prior information.With the help of the attention mechanism,the network can learn and utilize the parts and channels that should be focused on in the image.Secondly,this is an end-to-end model structure.The advantage is that it is convenient for deployment and training.Since the data characteristics of the main training set and test set are consistent,the accuracy rate will achieve a good result.In the demonstration part of the experiment,the effectiveness and comparative advantage of the proposed model are demonstrated by ablation experiments.(3)In view of the ambiguity of the parameters of the neural network algorithm,a new detection framework method is proposed that combines the deep learning network and the image feature processing algorithm.The proposed framework uses the deep residual network Res Net50 as the backbone network,and extracts the image features after network training from the pool5 layer of Res Net50,and then uses PCA and Fisher linear discriminator to filter all the extracted feature information.Below are the important distinguishing feature information,and finally use the SVM classifier to classify and judge the information.Compared with those frameworks that rely on deep learning alone or directly extract image features,this framework takes into account the advantages of both,and the neural network can be used to mine deep information as much as possible,but at the same time,the reprocessing of this information is fixed.It is ensured that the processed information has a certain stability,and there will be no large deviation with the change of network parameters.During the training process,the information contained in the facial images that can discriminate whether the subject is an addict will be encoded as abstract features by CNN.Since there is a lot of redundancy between the feature information extracted from the pool5 layer,we need to The extracted information is subjected to PCA dimensionality reduction to reduce the redundancy of information.Since in this framework,there is no judgment on local features of face images,all the information is learned by the neural network autonomously,which greatly avoids the influence of subjective prior information.In the demonstration part of the experiment,the effectiveness and comparative advantage of the proposed model are demonstrated by ablation experiments.