FingerVision is becoming a promising solution to tactile manipulation tasks. This paper introduces a modular custom FingerVision and a new method of specific contact surface recognition implementing KNN (k-Nearest Neighbor) algorithm. This method uses the markers array in the sensor as feature points to describe the deformation in contact, and generates a database recording all simulated conditions to identify surfaces of different shapes. Adjusted parameters of the blob detection algorithm and a frame correction algorithm by linear interpolation in the feature extraction process are proposed as well. Recognition effects of different sample sizes, k values and other factors are discussed with an accuracy test. With optimization, accuracy rate of recognition achieves 100% among 570 test samples with 1140 given samples from 7 surfaces in 19 categories. The conclusions are that the accuracy of this method hardly depends on the k value and the dimension of coordinates analyzed, so the default k value 5 and only abscissas are applied to simplify the calculation. As for sample size, when the given sample number of each category reaches about 40, the accuracy rate is almost 100%, and thus it is unnecessary to continue to increase the sample size.