As LLE algorithm is apt to be influenced by the deflection angle and the selection of the proximal points, a method of Double-layer LLE (DLLE) human ear recognition is put forward. And a further algorithm combined with Gabor is proposed with the name of GDLLE (Gabor and DLLE). The DLLE algorithm calculate the Euclidean distance between sample point and the center of the sample points first, then take all of the K-class sample points whose Euclidean distance are the minimum as neighbor points of LLE, extract the discriminate features, and classify them by the nearest neighbor classifier finally. The experiments on USTB3 ear image database reveal that the DLLE method proposed in this paper can reduce the impacts of the deflection angle and the selection of the proximal points on LLE algorithm, and recognition rate is further improved with the combination of Gabor wavelet.