Research On The Classification Of Non Time-sequential Image And Time-sequential Speech

Data can be classified as non time-sequencial data and time-sequencial data. Data classification, especially non time-sequencial image and time-sequencial speech classification, is one of the hot topics in data mining. High content screening of small compounds on stem cells and language independent speech recognition are two important applications in image and speech classification, respectively. The former is essentially to classify images with multiple samples for each image, and the latter is basically to classifiy speeches with single sample for each speech.The current classification algorithms for the above two applications need to be improved. For conventional high content screening algorithms, they need the image segmentation results to acquire the features, which are used for the classification. Human intervention is involved in image feature statistics, and huge space is also required to store the feature information. For current general speech recognition algorithms, they are normally based on the training of statistic model. The difficulty of multi language training data collection and the inherent risk of personal privacy disclosure make the general speech recognition algorithms hard to fit the language independent speech recognition application. To avoid the disadvantages and weakness of the current image and speech classification algorithms, this thesis researches on novel efficient image and speech classification algorithms.The mainly contributions and novelty of this thesis are as follows:1. Base on Kolmogorov complexity and information theory, we design cytoplasm image classification(CIC) algorithm and the information distance classification(IDC) algorithm that is an improvement of the CIC algorithm.Unlike the current image classification algorithms, CIC and IDC do not need image segmentation results to do image classification, thus, they avoid to extract image features, do not take huge feature dasta storage space, avoid staining’s cytotoxic/cytostatic problems and segmentation errors, and allow biologists’ continuous kinetic studies. Moreover, they are easy to be used and do not require human intervention, which is required in the current high content screening image classification algorithms, to obtain image information such as cell count and brightness.The results show that CIC algorithm successfully classifies two different compound classes represented by their images and acquire a similar result with the biologists’ conventional analysis. By considering the high cost of secondary biological confirmation experiments, we do not carry out the secondary experiments for the stem cell differentiation image data processed by CIC algorithm. However, we finish the secondary biological conformation experiments for the large batch of stem cell differentiation image data proceesed by IDC algorithm, conventional anlaysis, and five different machine learning algorithms. Compared with the conventional analysis and the five different machine learning algorithms, the IDC algorithm achieves better efficiency. Generally speaking, considering the “black box” classification process and easy-to-use of our algorithms, they are suitable for biologists to perform efficient and robust high content screening small compounds on stem cells.To the best of our knowledge, CIC and IDC algorithms are the first work to apply information distance to high content screening.2. Based on dynamic time warping(DTW) and fuzzy logic, we design merge-weighted dynamic time warping(MWDTW) algorithm and one-against-all weighted dynamict time warping(OAWDTW) algorithm that is improved from MWDTW algorithm.Compared with current general speech classification algorithms, MWDTW and OAWDTW algorithms offer a kind of light weight speaker dependent speech classification approach for language independent speech recognition. Here, light weight speaker dependent means that each speech class has only one sample. MWDTW and OAWDTW do not need to train large amount of data, thus avoiding the collection difficullty of multi-language training data. Forthermore, they avoid the risk of personal privacy disclosure as they perform off-line speech classification rather than upload data to remote server.For the classification of the speeches recorded in a quiet environment, MWDTW acquires better performance than DTW, merged-DTW and HMM. Compared with the DTW, the OAWDTW achieves better accuracy in the classification of speech recorded in noisy and bad recording environment.As far as we know, MWDTW and OAWDTW are the first weighted DTW algorithms especially deisnged for speech recorded in quiet environment and adverse conditions, respectively.3. In this thesis, we combine support vector machine(SVM) and DTW to design the SVM-merged DTW(SVM-MDTW) algorithm, and expect to perform accurate and fast speech recognition in adverse conditions. Even though the results are not as good as DTW, SVM-MDTW is an early attempt of SVM and DTW combination. Thus, SVM-MWDTW has intrinsic research value.