Research And System Implementation Of Feature Fusion Face Recognition Based On Panoramic Stereo Video

As the most accurate and convenient biometric recognition technology,face recognition is now widely used in many fields,but in most application scenarios it requires active cooperation from the user to obtain good recognition results.Panoramic stereo video is a special form of video,consisting of a combination of panoramic video and stereo video.In contrast to normal video,the image of a panoramic stereo video is a sphere,with a large number of different degrees of deformation in the projected image of a two-dimensional plane.When faces appear in these areas,there will be changes such as rotation,stretching,etc.,which will affect the recognition effect.In wild scenes,the accuracy of face recognition will be greatly reduced due to the influence of image quality such as lighting and noise,as well as uncontrollable factors such as face posture and occlusion.Based on these characteristics,this thesis investigates face recognition methods in panoramic stereo video.The specific research contents are as follows:Firstly,the imaging principle of panoramic video is analyzed.Due to the structure of the fisheye lens,there will be barrel distortion in the equidistant cylindrical projection image captured by the fisheye lens.The degree of distortion is related to the position of the picture.The farther away from the center of the fisheye image,the greater the degree of distortion.When the face appears in this areas,it will be deformed.Aiming at this problem,a new lightweight convolutional neural network is proposed,which extracts deformable face features by deformable convolution.At the same time,a panoramic face dataset is constructed for training and testing.Experiments show that the model not only has good accuracy in some large scale public regular face datasets,but also achieves better results in panoramic face datasets.Binocular stereo video simulates the effect observed by the human eye by placing two cameras in parallel at a small distance apart and shooting videos simultaneously at different angles.In a panoramic stereo video,the face of the same person will be captured by cameras with different angles to different images,such as different angles and different occlusion ranges.Therefore,a face recognition model that fuses features from binocular face images is proposed.The model integrates the feature between channels during feature fusion,and redistributes the information between channels in the deeper network,making full use of the information between different channels for feature extraction.The experimental results show that the method has higher accuracy in panoramic stereo video.The thesis builds a panoramic stereo video streaming and face recognition system.System divides 8 cameras into four groups of binocular cameras for video collection.After the images are processed and stitched,they are encoded and streamed to the cloud server.The receiving side can watch it through the different devices.At the same time,a face recognition system is constructed using the method proposed in this thesis.First,the panoramic stereo video stream is received from the cloud server.After the face images are detected,cropped and paired,and the left and right face images of the same person are input into the model for feature extraction.Finally,the feature vector is matched and retrieved in the face database and the recognition result is displayed in the video.