Research Of Applications Based On Ensemble Learning And Deep Learning

Pattern recognition systems based on machine learning usually consist of two parts:feature extraction and classification. For feature extraction, in this thesis, we not only try to manual design new features, but also study the advantages of learned features. Generally ensemble classifier is superior to the base one, therefor we adopt ensemble classifier as the pattern recognizer. In short, this thesis studies the ensemble learning、feature learning and their applications, and proposed corresponding solutions for membrane segmentation and retina blood vessel segmentation.The main contents of this thesis have two parts:hierarchical level features based trainable membrane segmentation, feature and ensemble learning based retina blood vessel segmentation.In this thesis, we first propose a method for membrane segmentation based on hierarchical level features (HLFs) and random forest (RF). Considering the local clustering characteristic of membrane Electron Microscopy (EM) images, we first define and extract the HLFs, which can be seen as a kind of more reasonable and natural representation. RF supplied with HLFs is then trained, thus we propose the HLFs-RF method for membrane segmentation. Experiments on ISBI2012EM Segmentation Challenge show us that:1) HLFs combine the merits of pixel level features and superpixel level features, for pixel level features are quite rich and superpixel level features are perceptual meaningful.2) Comparing to conventional neighborhood based on fixed shaped and sized rectangular window, HLFs can adaptively capture context information around single pixel, better describe the local intricate microstructures, and furthermore improve the recognition rate of membrane pixel.3) Comparisons with other published methods show us that in case of low dimensional features and small sample space, HLFs-RF can still achieve promising segmentation performance. In addition, we propose a representative sample selection method based on superpixel, reducing the sample space and the redundancy between samples.This thesis also proposes a hierarchical retina blood vessel segmentation method based on Convolutional Neural Network (CNN) and Random Forest (RF). Since CNN can directly learn features from raw images and RF has superior classification performance, it is natural to combine the advantages of CNN and RF, making the whole pipeline of pattern recognition system automatic and trainable. Specifically, CNN performs as a trainable hierarchical feature extractor and ensemble RFs work as a trainable classifier. It is noteworthy that learned features are extracted using not only the last layer output but also the intermediate output of CNN. Learned features from same layer of CNN are fed into one RF, finally ensemble method of winner-takes-all is utilized to fuse the outputs of multiple RFs.For the proposed membrane segmentation method based on HLFs and RF, HLFs is quite powerful for discrimination and the proposed method acquire a promising segmentation results. Future work will focus on two directions:on the one hand, more discriminative feature definition and extraction for superpixel; on the other hand, design new classification approach so that HLFs can be fully used.Research on the combination of feature learning and traditional classifier not only is theoretical significant but also has various application scenarios. The proposed retinal blood vessel segmentation method based on feature learning and traditional classifier both achieved good experimental results on two public databases. We will continue study the improvement of proposed method’s computational time, try to develop other deep learning technologies and plan to apply the proposed method to other area.