Ear Recognition Method Based On Locally Linear Embedding And Its Improved Algorithm

As a new biometric identification technology, ear recognition can be widelyapplied in information safety and public safety field, and has drawn more and moreattention from the scientists, because of the unique physiological position and structurecharacteristic of ear. After more than10years of the exploration and research, thescientists have made some achievements in ear recognition, but these methods do notapplied to the practice, and many problems need to be solved. According to the specificcharacteristics of the human ear, the paper deeply studies the key technologies andkernel algorithms of an ear automatic identification system.Firstly, Gabor wavelet transform is used to extract features of the human ear.2DGabor filter function is robust to resist the variation of lighting conditions and headpostures, so the result of the transformation can better reveal the inherent characteristicsof ear images. Secondly, the locally linear embedding (LLE) algorithm is used to reducethe dimensionality of high-dimensional data. But traditional LLE algorithm choosessample’s neighborhood simply, so it may not work well when high dimension dataalready has class information. To solve this problem, the double LLE algorithm basedon Gabor wavelet transform and correlation coefficient is proposed, which can select allthe same kind of sample points as the nearest neighbor points and reduce the effect onthe dimension reduction due to the improper selection of the local neighborhood. Lastly,the improved algorithm is a supervised learning algorithm, and cannot be applied to theear recognition with a few identity labels. On this basis, combining with the principleand characteristics of semi-supervised clustering, the double LLE algorithm based onGabor wavelet transform and semi-supervised clustering is proposed.The experimental results in USTB ear database are as follows:①2D Gaborwavelet transform can decrease the influence of illumination and head postures on earrecognition.②The double LLE algorithm based on Gabor wavelet transform andcorrelation coefficient can acquire better recognition accuracy and running speed thanthe traditional LLE algorithm.③Though the ear recognition accuracy of the doubleLLE algorithm based on Gabor wavelet transform and semi-supervised clustering is alittle lower than the previously improved algorithm, the recognition curve of the newimproved algorithm gradually approach to the previously improved algorithm’s with theincrease of the number of class labels. In conclusion, this paper makes a thorough study of ear recognition based on LLEalgorithm, and proposes two improved LLE algorithms. Experimental results show thatour methods are feasible with high accuracy in ear recognition, which is expected toprovide a solid foundation for practical application of ear recognition.