Feasibility of steerable pyramid filters for image classification

Abstract

Image classification, for object recognition or scene classification, has been an extremely active research area since perhaps the advent of computer vison techniques in the 1960s. In this report the feasibility of using an image transformation, generally reserved for texture recognition, to obtain discernable image features has been discussed, which can aid this age old problem. “Steerable Pyramid Decomposition” is investigated both in isolation and in conjugation with “Spatial Pyramid Matching”, to observe whether it can by itself or by augmenting an existing proven technique aid in scene or object recognition. An image classification system is presented, which makes use of this technique, and its performance on 4 datasets, two involving scene recognition and the other two involving object detection has been investigated. It is found that features obtained from “Steerable Pyramid Decomposition”, while not very powerful in isolation, show good potential in augmenting the performance of “Spatial Pyramid Matching”. This is especially visible in scene classification tasks, which it proves more effective at than object detection.