Study On Face Detection Based On Deformable Part Model

Deformable part-based model (DPM) has been more and more popular for objectrecognition. It takes both the global and local information into account, and couldcapture richer information to achieve a better classification result. However,achievements at present have seldom concerned the importance degree of parts for theobjects with strong structural prior knowledge. For example, in the face area, the partsof eyes and mouth often play a more important role than others, and the equaltreatment strategy can’t achieve the best result in detection.In this paper, a novel multi-view face detection method is proposed based on theweighted deformable part-based model (WDPM). Accounting to the differentcontributions of the parts to the detection process, we assign a weight to each part toimprove DPM, and give a computing method of weights. The greater the weight is, themore important the corresponding part is for detection. The weighted deformable partmodel is also extend to detect faces with three kinds of views, front, half-profile, andfull-profile, and the three models for different views are made fusion to achievemulti-angel face detection according to special fusion strategy. Because the profilefaces are more difficult to be detected, we also make the lower threshold of the modelfor profile view to ensure a better result. Skin color confirmation is used to decreasethe error detection caused by lower threshold. In the detection for the objects withfewer pixels, bilinear interpolation method is made use of to enlarge pictures toimprove the robustness and detection efficiency of algorithm. Experiments show thatthe proposed method has better performance on PASCAL databases, LFW database,and the face database consisted of Internet pictures.