Research On Face Attribute Recognition And State Detection Technology Based On Machine Vision

For the needs of security,privacy and public order,identity recognition is gradually being valued,and face recognition is widely used in life and work because of its unique and difficult-to-replicate characteristics.At present,face recognition mainly focuses on a small number of extended tasks such as identity recognition,age recognition,and gender recognition.The extraction of depth information of human faces is still a less-touched area.In addition,part of the attributes of a human face are usually difficult to be limited to a certain area of the face when describing,and a target is often given multiple labels.Therefore,the description of facial attributes is usually regarded as a multi-label classification task.This paper studies the task of multi-label recognition starting from face detection.In view of the different requirements for feature extraction capabilities due to different facial attribute description areas,after comparing the feature extraction framework,select YOLOv4 as the feature extraction network,and the modified Resne Xt feature extraction structure is integrated according to the needs,design a global feature extraction module that focuses on contours and a local feature extraction module that focuses on details to improve network classification capabilities;for the problem of complementary spatial geometric information between different areas of the face,partbased detection ideas are used to extract feature images of different areas,and local parts of the PS-MCNN network are used.The shared structure builds the network architecture,combines high-level semantic features,and improves information interaction and feature sharing between different branch networks;for the semantic related issues between facial tags,a tag weighting module is proposed to map tags in the feature space,and through improvements The latter LSTM structure continuously transmits the previous prediction information backwards,and then assists the network in classification and improves the accuracy of network classification.Based on the above improvements,the MFARNet network is designed and tested and compared under the same conditions.The network has achieved better results in the application of facial attribute recognition,and the accuracy has been effectively improved.In terms of state recognition,this paper implements one of the application directions of face attribute recognition-fatigue driving detection.In view of the differences in lighting conditions that may occur in practical applications,the image is self-adapted to equalize the local histogram in the image preprocessing step to strengthen the adaptability to light changes,improve the face resolution in low-light environments,and retain the face Detailed information to assist the next step of detection and identification tasks.After the processed video frame is obtained,the face area is recognized,the landmark points are further located,and the coordinate positions of the facial landmark points are obtained.Extract the position information of key landmarks in the eye and mouth area,determine the range of the two,and calculate the aspect ratio respectively,and analyze the open and closed state of both eyes and mouth.Extract the position information of the basic landmarks of the face,construct the threedimensional coordinate axis of the face,calculate the elevation angle of the face,and recognize the nodding state that appears under fatigue.In addition,the attributes of the video frame are analyzed,and the two attributes of "Blurry" and "Mouth Slightly Open" are detected,which are used as an auxiliary proof in the case of inaccurate positioning to participate in the classification of the fatigue state.After real video testing,under different light conditions,the system can accurately locate key points on the face,and further achieve more accurate calculation and statistics of different parameters such as eye closure,eye closure frequency,yawn detection,and nodding detection.Facial area status discrimination and fatigue status recognition can realize real-time screen prompts and warning sounds when it is identified as fatigue,which confirms its applicability in actual scenes.