Object recognition using deep learning features

Abstract

Image recognition is a hot research area and still a changeling in computer vision. Many algorithms are available for automatic recognition of many different object categories. However, previous algorithms need to train classifier to rank proposal and learn from object class recognition datasets. That is time consuming since those classifiers need to be trained again when new object category comes and database grow. This project proposed a more efficient algorithm called Data Driven Object Proposal which doesn’t need to train classifier. This project labels the elements of an image into plausible object regions and provide a label for each object without knowing a-prior which objects are present in that image. Our methods extract features from an annotated image dataset and transform the features into hash code. After that we store those features and related object labels into database. Around 131067 images and 3819 object categories are carried out. For a test image, we retrieve object regions from images and obtain those region features. After that, we search our pre-trained database to get nearest region features. Those pre-trained features are coupled with object labels. We found most common labels and assign the labels to our teat image regions. Different parameters are explored and analysis is carried out to determine which approaches achieve the highest accuracy. The results for each class category are evaluated. Specially, a GUI was developed to allow the users retrieve object labels of images.