Research Of Ear Detection In Complex Static Color Image

As a novel biometric identification technology, more and more attention was paid to the recognition of human ear because of its unique physiological characteristics. A whole ear recognition system is mainly composed of four steps: image inputting, ear detection, feature extraction and recognition. Compared to ear-feature extraction and recognition, little attention was paid to ear detection in the early research. Now, with in-depth study of recognition system, ear detection in different conditions becomes more and more important as the first key step in the whole recognition system. But so far, most of ear detection methods are suitable for simple background or even only ears.Based on the research of human-face detection, existing ear-detection methods were investigated in this paper, and a new optimized detecting method was put forward to solve difficulties in ear detection such as small region, few common features, complex background and so on. This method aims at detecting frontal ears in complex static color images and developing a feasible scheme for human-ear recognition system in complex background.There are four phases in this method. Firstly, images are segmented into different regions with skin-color information and potential candidate regions are selected. After analyzing the feature of skin-color distribution in different color spaces, YCbCr color space is considered as the segmentation space and Gaussian model as the skin model. Then segmentation is accomplished with dynamic threshold. Secondly, for each segmented region, morphological operations and prior knowledge of side-face are applied to eliminate impossible regions. The prior knowledge includes size of skin areas, height-to-width ratio of side-face and so on. Thirdly, according to the rich edge information that internal ear has, image edge is detected with wavelet modulus maxima method in different scales. Then these edge images are superimposed to a binary edge image. Finally, edge characteristics of side face regions in those edge images are statistically analyzed. From the analysis we know that closely spaced edges only appear in ear region. Other parts of the side face only have obvious contour lines. According to this prior knowledge, morphological operations such as dilation, filling, thinning and reconstruction are applied to the binary edge images. After that, the independent contour lines are eliminated while closely spaced edges of the ears are filled to edge regions. Thus ears are detected with these regions. In conclusion, learning from the technology of human face detection, we put forward a method for frontal ear detection in complex background in this paper, which includes selecting candidate regions, extracting edges, determining common ear features and so on. Experimental results show that ear detection is feasible with our method, which is expected to be beneficial to the development of automatic ear recognition system.