Performance Benchmarking: Pre-trained Models and Custom Convolutional Neural Networks in Deep Learning

Sheemona Joseph C.; Ganesh, S.; Kannadhasan, S.; Selvipriya, K. · 2025 · DOAJ

DOI: 10.33093/jiwe.2025.4.2.14

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the critical need for accurate and robust traffic sign recognition to support Advanced Driver Assistance Systems (ADAS) and autonomous vehicles. The authors identify that while Convolutional Neural Networks (CNNs) are effective for image classification, standard architectures like VGG16 suffer from limitations such as shallow feature extraction, high overfitting risks, and poor performance under varying real-world conditions like low light, adverse weather, and occlusions. The study aims to develop an enhanced VGG16 model capable of classifying all 43 classes of traffic signs in the German Traffic Sign Recognition Benchmark (GTSRB) dataset, rather than just broad super-categories, to improve road safety and decision-making reliability. The methodology involves modifying the standard VGG16 architecture by adding additional convolutional layers, integrating batch normalization to stabilize training, and applying dropout layers to prevent overfitting. The system utilizes ReLU activation functions for non-linearity and pooling layers to reduce computational costs. Data preparation includes scaling images to 224x224 pixels, normalizing pixel values, and employing extensive data augmentation techniques—such as random rotations, translations, scaling, and brightness adjustments—to simulate real-world variability. The model is trained using Stochastic Gradient Descent (SGD) with backpropagation and L2 regularization. The implementation was conducted using Python within an Anaconda environment, utilizing libraries installed via pip, and the system interface was developed using Django, HTML, and CSS. The results demonstrate that the enhanced VGG16 architecture significantly outperforms the base model and other contemporary methods in handling complex traffic sign features. The authors report high performance metrics, citing an Area Under the Curve (AUC) of 99.73% for the "Danger" category and 97.62% for the "Mandatory" category in detection benchmarks. The model achieved classification accuracy of approximately 99.67% and near 100% recall on clean static image datasets. Evaluation metrics, including training and validation accuracy, loss convergence, and True/False Positive Rates, confirmed the model's ability to generalize effectively without significant overfitting. The system successfully distinguishes between individual sign classes despite challenges in lighting and perspective. The significance of this work lies in providing a more reliable and accurate framework for traffic sign recognition, which is essential for the safe operation of autonomous vehicles and intelligent transportation systems. By addressing the shortcomings of previous CNN models, the proposed architecture offers improved robustness against environmental variations. The authors conclude that this enhanced model is a strong candidate for real-world deployment and suggest future research directions, including the integration of transfer learning, multi-sensor fusion with LIDAR or RADAR, and expanding datasets to include signs from various countries to further enhance global applicability and cybersecurity resilience.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-25; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• adas effectiveness