AN TRANSFER LEARNING APPROACH FOR IMAGE CLASSIFICATION USING BINARY IMAGE SEGMENTATION ON LIMITED DATASET

Abstract

Image classification has become a part of our daily routine whether it is classifying between traffic signals or different types of species. However, to differentiate between similar texture and shapes is a difficult task with a naked eye. Latest advancements in the field of computer vision can make this task of image classification easier with deep learning techniques, especially neural networks. However, training neural networks require large datasets, otherwise, it cannot give accurate classification. Inspite all the data availability, there are some subjects which lack enough data. Medical images, rare animals species to name a few examples with relatively less number of information. In our experiment, we have taken those animal species datasets with a minimum number of data and achieved higher classification accuracy. We have examined the various state-of-the-art neural networks like DenseNet and Convolutional Neural Networks that could classify between various animal breeds, and flower species. Furthermore, we compared their results based on accuracy achieved on the test set to determine the most efficient approach. Thus, we could assess which network is most suited for image classification. Moreover, we proposed a two-phase algorithm which differentiates between multiple image dataset through transfer learning via pre-trained Convolutional Neural Network. Initially, images are automatically segmented with the Fully connected network to allow localization of the subject through minimum bounding box around it. Second, we built a robust convolution neural network fine-tuned with a dense network according to our vi

image datasets. We also proposed novel steps during the training stage to ensure a robust, accurate and real-time classification. Finally, we have evaluated our method on the well known dog breed dataset, and bird species dataset. The experimental results outclass the earlier methods and achieve an accuracy of 95% to 97% for classifying these datasets.