Abstract
The key of image recognition and classification based on machine learning is to extract image feature points effectively and classify image features correctly. In this paper, SURF (Speeded-Up Robust Features) is used to extract image feature points. The extracted feature points are clustered by K-means algorithm, and the center of clustering represents this kind of feature. The BOW (bag of word) trainer for each image is created according to the image features completed by clustering, and then the trained KNN (k-Nearest Neighbor) classifier is used to recognize and classify each image. The descriptor of SURF algorithm is 64-dimensional, which is faster than traditional feature extraction algorithm. KNN algorithm is suitable for multi-class label classification problem and has good generalization ability. The experimental results show that the accuracy of image classification reaches 90% when good parameters are selected, but there are some shortcomings for image classification with high similarity.