《EE368_Final_Project_Report_RGB_Level_Face_Detection》.pdfVIP

EE368 Final Project Report RGB Level Face Detection Jian Zhang 4881272 Miao He 4881074 Jing Chen 4882379 Abstract In this work, we describe a 2-step algorithm based on skin toner mask and Support Vector Machine (SVM) for face detection in RGB images. Accurate face detection is demonstrated on 7 training pictures offered by EE368 project WebPages. The over all accuracy rate is 97.5%, and runtime ranges from 62 sec to 220 sec, depending on the size and number of faces. In addition, we compare this 2-step approach with some other popular methods. Finally, some further possible improvement is suggested. 1. Introduction Our aim is to develop a proper algorithm and use Matlab to implement it, which can automatically detect all faces on an RGB-scale digital photo in a limited time with a high accuracy. While this task looks simplified since all the pictures (training and final test images) are shot at the same time and place, we find it hard to achieve from the following aspects: ♦ There are quite a lot of faces in the images. Thus the face size falls into several scales, making it time-consuming to search the face candidates. And in the limited frame, some faces are so near to each other that they got overlapped. ♦ The photos are shot under sidelight instead of diffused light, so this is severe influence from shade. What’s more, the diversity of complexion, the rotation of head, and the inhomogeneity of background, all these add difficulties to our work. ♦ The images are 1280*960 with high resolution. While it is great to give us more information, the big size greatly slows down the detection speed. And on the other hand, more information also