Detecting Faces in Low-resolution Images

Face detectio n is o ne o f th e h o t research to p ics in C o m p uter Visio n, and greatly p rogressed in past decade. However, as far as we know, face detectio n in low-reso lutio n im ages h as no t been studied (m o st sy stem detects faces bigger th an 2 0×2 0 o r 2 4×2 4 p ixels). A co nventio nal AdaB oo st based face detectio n m eth od (Vio la & Jo nes) can detect m erely 3 2 % o f faces in 1 / 4 reso lutio n M IT +C M U fro ntal face test set. In th is paper, we p ro po se a new face detectio n m eth od fo r low-reso lutio n im ages by co m b ined use o f two classifiers: o ne classifier detects faces and th e o th er detects up per-bodies. T h ese classifiers are ap p lied to m agnified low-reso lutio n im ages. T h e co m b inatio n o f classifiers is realized by using a neural netwo rk. As th e result, o ur m eth od ach ieved 8 3 % o f th e face detectio n rate fo r th e 1 / 4 reso lutio n M IT +C M U test set.