A fast and area-efficient architecture for classifying images based on binarization and a binary dual-feature set




Attarmoghaddam, Narges

Journal Title

Journal ISSN

Volume Title



Image classification is an active research area in computer vision with many applications. Image classification is a computationally complex task, but for embedded applications, only limited resources are allowed. Moreover, additional constraints such as reliable classification accuracy, high-throughput performance, power-efficiency, and real-time speed must be fulfilled. Due to potential for parallelism, low power consumption, scalable resource utilization, and reconfigurability, Field-Programmable Gate Array (FPGA) devices are well-suited to overcome these challenges in image classification system implementation. In this dissertation, we implemented two feature-based image classification systems using binary feature sets. A binary feature of an image pixel is represented using one bit, while a non-binary feature is represented using more bits. Therefore, implementing an image classification system based on a binary feature set needs fewer hardware resources for storage and computation. The first proposed system is founded on a single binary Histograms of Oriented Gradients (HOG) feature set. Using a binary feature set generally leads to an area-efficient and faster architecture; however, accuracy is lost because binary features contain less information than non-binary features. We, therefore, proposed the second system that uses a dual feature set that combines HOG and Local Binary Pattern (LBP) features to improve classification accuracy performance. The Support Vector Machine (SVM) classification algorithm is utilized as the classifier in both proposed systems. To obtain binary features, two steps of binarization are applied to the HOG descriptor. First, HOG features are extracted from binary images to simplify the feature extraction process. Second, the block histogram normalization of the both HOG and LBP descriptors is replaced using binarization to reduce hardware resource utilization in the descriptors and the SVM classifier. Binary-based computation in the proposed systems results in resource utilization reduction, thus allowing an area-efficient architecture even with two feature descriptors implemented. Compared to similar existing works, our system speeds up the classification process while utilizing fewer hardware resources, with comparable accuracy.



Image classification, Binary dual-feature set, SVM classifier, HOG feature descriptor, LBP feature descriptor, Area-efficient