A method based on the feature extraction of a wavelet packet energy spectrum and the recognition of a Support Vector Machine(SVM) was presented to solve the problem of recognizing the floating and suspending impurities in ampoules. Firstly, an impurity zone’s image was extracted as an object image through the image sequence difference and point detection division. Then, a 1D signal could be obtained through adding the ROI row by row in the axis direction of an ampoule. The 1D signal was decomposed by a wavelet packet, the independent primary components in the wavelet packet feature vector were extracted by using Primary Component Analysis(PCA), and the wavelet packet energy spectrum of the independent primary components was taken as the feature of impurity types.Furthermore,the extracted feature was taken as the input vector of a SVM,and the sample features could be classified rapidly by a sequential minimal optimization method through training. Different types of core functions and corresponding parameters were selected for training and testing in the experiments,and obtained results show that the recognition period of SVM has decreased by 60% and the recognition precision improved by 35%,respectively, as compared with those of the BP network. This method can meet the requirements of the floating particles for feature extraction and rapid recognition in production.