Design and Analysis of Pattern Classification Algorithms

Abstract

Pattern classification is a domain under machine learning that involves classifying the given data into distinct groups called as classes. The samples under the same class share lar properties and characteristics. The main objective of this research work is to rst study the existing pattern classification techniques and then design our own algorithms and architectures for the same purpose. We intend to use these proposed approaches for solving this problem, specically for the recognition of handwritten Kannada numerals with the aid of two dirent datasets.In this thesis, we propose three different algorithms that provide a good extension to the results achieved by previously used algorithms. Each of these techniques makes use of feature extraction and subsequent grouping such that satisfactory results are achieved in pattern classification.We propose a customized version of the Convolutional Neural Network as the algorithm for pattern classification. The stacking of convolution and pooling layers together in a specic order helps for accurate retrieval of feature maps, which in turn help for assigning the appropriate labels to a new test sample and achieve accurate results. The second proposed algorithm is an extension of the previously proposed Customized Convolutional Neural Network in terms of its initialization strategy. Convolutional Autoencoder is used to train the architecture in an unsupervised manner to get better initial architecture weights for faster convergence of the model. Such an initialization helps us outperform previously suggested method as seen from the obtained results. The third proposed algorithm involves coupling CNNs with the Particle Swarm Optimization(PSO) technique. PSO can automatically search for the best possible conguration of the convolutional neural network and it helps in reducing down the human iterative approach of designing the network architecture. Further, as it is a natural algorithm, it derives inspiration in global best solution search and provides a different

Description

Keywords

Citation

item.page.endorsement

item.page.review

item.page.supplemented

item.page.referenced