Considering the problems of dimensional disaster and low recognition efficiency that appear when extracting texture feature using Gabor wavelets method only, a method based on both Gabor wavelets filter and LPP dimensionality reduction algorithm was proposed. First, the description of highdimensional feature vectors of five scales and eight orientations of the image were obtained by using Gabor filters. Next, lowerdimensional feature vectors were obtained by using LPP algorithm. Finally, the lowerdimensional feature vectors were used to recognize different types of froth under different conditions using BP (back propagation) neural network to direct actual mineral manufacture. It is demonstrated by experimental results that this method has a less texture feature vector dimension and a higher recognition efficiency relative to the traditional methods based on GLCM and Gabor wavelets only when extracting texture feature.