Abstract: Principles of Gabor wavelet transform and kernel locality preserving projections (KLPP) are studied and characteristics of surface defects on hot-rolled steel plates are analyzed. A feature extraction method based on Gabor wavelet and KLPP is presented and applied to automatic recognition of hot-rolled steel plate surface defects. Surface images is decomposed into 40 complex-value components at 5 scales and 8 orientations by Gabor wavelet transform, then means and standard deviations of real parts and imaginary parts of the components and the original image are calculated as features respectively to produce a feature vector with 162 dimensions, which is then reduced to 21 dimensions by KLPP. A multi-layer perceptron classifier is constructed to classify the samples with the 21-dimensional feature vector. The feature extraction method presented in this paper has low computational complexity, high computational parallelism, and can discriminate edges and textures along different directions. The method is examined with samples of surface defects captured from a hot-rolled steel plate production line, and the classification rate is 93.87%.