A Study On Low Quality Leg Image Recognition Based On Texture Feature

Low-quality skin recognition,a supplement of fingerprint recognition,DNA recognition and facial recognition,is an important and original task in forensic analysis.Therefore,it is meaningful to do some researches on it.Although skin recognition technology has a long development,the object used for experiment is still limited to the high quality skin image,if match directly without any alignment,the result is not good.At the same time,though there is much high definition equipment in the real life,the image obtained is still low quality because of the influence of viewpoints,pose and illumination.These low-quality images mainly refer to low-resolution images and defective images.Moreover,the skin image we obtained is limbs image in most situations.In order to solve the identification problem in low quality of leg skin images,some researches on low quality of skin images based on the texture feature will be done.The detail research works are as follows:This paper proposed a method based on the trend of skin texture for solving the recognition problem and alignment problem.Firstly,the paper uses the mean filtering in preprocessing to smooth the image for removing noise and improving the accuracy of identification.Then,utilizing the principle of contour line in geography,and regarding the skin image as a three-dimensional topographic map.The paper designed a sliding analysis for extracting the texture information in a certain region.According to the texture information in region,the directional histogram can be calculated.Then,a twice-dynamic alignment for alleviating the influence caused by different viewpoints,poses,illumination is proposed.Finally,the paper use residual distance to measure the similarity between the probe and the gallery,and do some experiments in relative datasets.The accuracy of identification is 84.01% and 81.35% in 25 dpi and 6.25 dpi,respectively.The results illustrate that the proposed method has a robustness of resolution,viewpoint,pose and illumination.In the occlusion problem,the feature in low-resolution problem is used as identification feature.The alignment can avoid the situation that we can hardly obtain the geometrical relationship and do affine transformation because of incomplete leg edge.In this paper,we utilize the orientation feature and the new method of alignment and do some experiments in supplementary datasets.The experiment is in the low resolution and the occlusion conditions,the data in images lost about 20%.The accuracy of recognition is 76.79% and 75.00% in 25 dpi and 6.25 dpi,respectively.The results illustrates again that the identification characteristic is also effective in occlusion problem,and the new alignment is feasible too.