ResHOGNet: A Novel Hybrid Multi-Scale Residual CNN with DiffHOG Features for Lightweight and Efficient Traffic Sign Classification

Read full text Summarise with VIRAT
Authors: M Somanna, H.S. Nagendraswamy, B L Thejashwini, Manjuchethan N, B.M. Chethana Kumara
Cited by: 0
References: 23

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Objectives: The objective of this study is to develop an efficient and accurate model for traffic sign classification under diverse benchmark conditions. It aims to design a small dual-stream network that combines residual CNN features with a trainable HOG module (DiffHOG), learns gradient-based descriptors end-to-end, fuses multi-scale features without pyramid layers, and maintains a lightweight architecture with 1.18 million parameters, suitable for real-time deployment. Methods: ResHOGNet has been trained end-to-end on seven benchmark datasets: the German Traffic Sign Recognition Benchmark (GTSRB, v1/v2), Belgian, Arabic ArTS (v1/v2), and Indian ICTS and ITSDD, with class counts ranging from 17 to 62. Accuracy, precision, recall, F1-score, misclassification rate, and parameter count were all tracked, with macro and weighted averaging used to address class imbalance. To put the results in context, comparisons were made with classical methods such as HOG-SVM, well-known deep learning models including ResNet18 and ResNet50, and recent state-of-the-art architectures from the literature. Findings: ResHOGNet achieved test accuracy between 98.26% and 99.92%, averaging 99.10% across all seven datasets, with misclassification rates below 2% in every case, as low as 0.08% on GTSRB v2. Novelty: The hybrid CNN-Hog architecture with fully differentiable DiffHog enables end-to-end gradient optimisation. The SACMF mechanism dynamically gates multiple-scale features without the need of pyramidal structures. The information-preserving multi-stream design and efficient attention mechanisms in the residual blocks enable the multi-stream model to achieve 99% or higher accuracy across seven datasets while containing only 1.18 million parameters, making it suitable for real-time deployment. Keywords: Traffic Sign Classification, ResHOGNet, DiffHOG, Multi-Scale CNN, Residual Learning, Lightweight Networks, Hybrid CNN-HOG, Intelligent Transportation Systems

Fields of Study & Subjects

Advanced Neural Network Applications Computer Science and Engineering Autonomous Vehicle Technology and Safety

Publication details

Journal / source
Indian Journal of Science and Technology
Volume
19
Issue
10
Pages
736–750
Publication date
26 March 2026
Language
EN