The research level of ultrasound image recognition is cumbered by the complexity of the contents and the low signal-noise-ratio of ultrasound images, therefore an effective learning system is required. Besides the research of the specific classifiers, the data preprocessing mechanisms to improve the quality of the training feature data are also important, however, data preprocessing is almost equal to feature selection to most of the related researchers in the field of ultrasound image recognition. In the paper, the quality of feature data is considered at two different views and the corresponding algorithms are discussed. To solve these two problems together, the combination algorithms are proposed. The experiments are arranged on four UCI datasets and two feature data derived from the real ultrasound images, and the results show that the proposed algorithms can slightly improve the performance of the classifier applied on the datasets, comparing with other related algorithms.