Object Detection and Recognition Using Small Labeled Datasets

Abstract

Object detection and recognition is a vibrant research area in the computer vision community. Several methods that came into scenario of object detection and recognition are expensive. This paper proposes another methodology for the same. We use selective search algorithm for providing region proposals where there is good chance of finding the object. The method is based on segmenting and eventually merging regions with good similarities. In this paper, we also propose a method for object recognition with a small labeled dataset for training. We use effective methods of unsupervised pre-training to effectively train the network. This paper tries to recognize objects using convolutional neural networks which are pre-trained using a sparse auto-encoder. The region proposals for the objects are forwarded to a convolutional neural network for feature extraction and finally into a fully connected layer for classification.