Research And Implementation Of Face Detection System Based On Multi-core DSP

With the rapid development of the field of computer vision,a large number of machine learning algorithms have been applied to image analysis and processing,among which the face detection technology has attracted much attention due to its wide range of application fields.In this thesis,a complete front-end video face detection system has been implemented.The functions of the system include real-time video capture,transmission,face detection,filtering non-face image and face image deduplication,encoding and uploading.The NPD(Normalized Pixel Difference)face detection algorithm based on the AdaBoost classifier is used and optimized in this system as the software foundation.Ultimately,the whole face detection system is implemented on a high-performance,low-power multi-core DSP platform.The main tasks of implementing the entire system include algorithm research and system construction.In terms of algorithms,the general idea,basic principle,and some face features of AdaBoost-based face detection algorithms are introduced.Then according to the face detection algorithm evaluation standard,the performance of some common detection algorithms is compared.Finally,the AdaBoost-based NPD face detection algorithm is selected for research and improvement.On the one hand,the false detection rate of the algorithm is reduced by retraining classifiers and superimposing eye detection to filter non-face image.On the other hand,the classifier data of deep quadratic tree structure of the detection algorithm is analyzed,and the efficiency of algorithm execution is improved by reconstructing the original data into a complete binary tree.In addition,due to the continuity of the video,there are a large number of near-duplicate images in the face images captured by the system.Therefore,the image deduplication algorithm is studied.After comparing various features,the LBP feature is used to implement the face image deduplication algorithm.In terms of system construction,the McFW software development framework is used to build video chains on the DM8168 multi-core DSP to achieve video capture,preprocessing,and transmission.After the face detection algorithm is simulated and optimized,it is ported to the DSP core for video detection.In addition,in order to achieve system requirements,the system has been improved.Firstly,the moving object candidate region is extracted by the moving object detection algorithm,and then face detection is performed on the candidate region.The method improves the real-time performance of the system by filtering a part of the background area,and reduces the false detection caused by the background image to some extent.Secondly,the face image data from the DSP core is send to the ARM core through the inter-core communication mechanism.In the ARM core,some function,including face image deduplication,jpeg encoding,and upload to server is achieved.During the development process,the algorithm performance is tested and compared in two aspects.Firstly,the overall detection performance on a large number of images is tested using the FDDB database and the benchmark.Secondly,in order to compare the detection algorithm performance on video,a video face detection test tool is implemented on the PC.The tool can mark faces in video and use FDDB benchmark to determine missed detections and false detections.After performing the face detection algorithm on the video,information such as the number of missed detections,the number of false detections,and the execution time can be obtained.