Source-linked AI summary
RIECNN: Real-Time Image Enhanced CNN for Traffic Sign Recognition
Reem Abdel-Salam, Rana Mostafa, Ahmed H. Abdel-Gawad
TL;DR
Traffic-sign recognition must cope with variable image quality and the slow processing of complex deep-learning methods. RIECNN enhances captured images before CNN classification, and reports the highest recognition accuracy across three benchmarks. The authors also identify lower accuracy on two GTSRB subsets as a limitation.
Problem
Traffic-sign images vary in resolution, brightness, weather, blur, and occlusion, while deep-learning recognition methods can be complex and slow.
Source: §1 · pp. 1–2Method
RIECNN preprocesses captured images with enhancement techniques before feeding them to a CNN.
Source: §1 · p. 2, §5 · p. 12Results
RIECNN achieved the highest recognition accuracy among prior techniques on GTSRB, BTSC, and rMASTIF, with 99.75% on GTSRB, 99.25% on BTSC, and 99.55% on rMASTIF.
Source: §5 · p. 13Takeaways & Limitations
Across three benchmarks, RIECNN combines high recognition accuracy with fewer parameters than previous techniques.
Source: §5 · p. 13Accuracy on the GTSRB EndOf and Speed subsets is relatively lower than that of other state-of-the-art architectures.
Source: §5 · p. 13
Abstract
from arXiv · showhide
Traffic sign recognition plays a crucial role in the development of autonomous cars to reduce the accident rate and promote road safety. It has been a necessity to address traffic signs that are affected significantly by the environment as well as poor real-time performance for deep-learning state-of-the-art algorithms. In this paper, we introduce Real-Time Image Enhanced CNN ({\our}) for Traffic Sign Recognition. {\our} is a real-time, novel approach that tackles multiple, diverse traffic sign datasets, and out-performs the state-of-the-art architectures in terms of recognition rate and execution time. Experiments are conducted using the German Traffic Sign Benchmark (GTSRB), the Belgium Traffic Sign Classification (BTSC), and the Croatian Traffic Sign (rMASTIF) benchmark. Experimental results show that our approach has achieved the highest recognition rate for all Benchmarks, achieving a recognition accuracy of 99.75\% for GTSRB, 99.25\% for BTSC and 99.55\% for rMASTIF. In terms of latency and meeting the real-time constraint, the pre-processing time and inference time together do not exceed 1.3 ms per image. Not only have our proposed approach achieved remarkably high accuracy with real-time performance, but it also demonstrated robustness against traffic sign recognition challenges such as brightness and contrast variations in the environment.
1 Introduction
Traffic sign recognition must handle uncontrolled image conditions while meeting real-time constraints, but deep-learning methods can be complex and slow. RIECNN combines image enhancement with a CNN and is reported to outperform published work across three benchmarks.
1 Introduction
- Captured traffic images vary in resolution, brightness, weather, blur, and occlusion, making real-time sign detection and classification challenging.The paper attributes these challenges to uncontrolled environments.Source: §1 · pp. 1–2
- Deep-learning recognition methods can be complex and slow, while image enhancement may address image problems with a simpler network.The proposed enhancement runs at 60 × 60 resolution, with low overhead relative to CNN inference.Source: §1 · p. 2
- RIECNN preprocesses captured images with enhancement techniques before feeding them to a deep-learning CNN.The approach is evaluated on GTSRB, BTSC, and rMASTIF.Source: §1 · p. 2
- The paper assesses four image-enhancement stages separately and in combination for their effects on CNN accuracy.The authors frame this assessment as potentially informative for CNNs in other uncontrolled-environment domains.Source: §1 · p. 2
- RIECNN achieved the highest recognition accuracy among published work on GTSRB, BTSC, and rMASTIF.The introduction reports the cross-benchmark comparison without listing accuracy values.Source: §1 · p. 2
2 Related Work
Prior traffic-sign systems use ensembles, spatial transformers, multiscale or hierarchical CNNs, and lightweight or multi-dataset designs. The section identifies a remaining need for a lightweight, stable method with competitive accuracy across diverse datasets.
2 Related Work
- Ciresan et al.’s 25-CNN committee reached 99.46% accuracy, while its roughly 90 M parameters limited real-time performance.García et al.’s three-spatial-transformer network reported 99.71% on GTSRB and 98.86% on BTSC.Source: §2 · p. 2
- Sermanet et al.’s multiscale CNN reached 98.31% and later 99.17% GTSRB accuracy; Saha et al.’s hierarchical residual CNN used dilated skip connections.The 99.17% figure follows increasing network depth and ignoring color information.Source: §2 · p. 2
- Zhang et al.’s distilled student network reached 99.61% on GTSRB and 99.13% on BTSC with 0.8 million parameters, but faced convergence instability.A teacher model must be trained for each dataset, and student performance depends on teacher quality.Source: §2 · p. 3
- Mao et al.’s hierarchical CNN reached 99.67% on GTSRB, but selecting the number of subsets for other datasets may be time-consuming.Source: §2 · p. 4
- OneCNN was tested on GTSRB, BTSC, and rMASTIF, while CNN-ELM reported 99.40% accuracy on GTSRB without data augmentation.OneCNN used augmented duplicate images and was reported to outperform human accuracy and other architectures.Source: §2 · p. 4
- A lightweight, stable approach with competitive accuracy across multiple diverse datasets remains a challenge for automotive deployment.This is the related-work section’s stated research gap.Source: §2 · p. 4
3 Methodology
RIECNN classifies traffic signs in two stages: image preprocessing followed by a deep-learning architecture. Its preprocessing enhances input images before classification, while its CNN uses a shallower VGG16-like structure.
3.1 Pre-processing
- RGB images are resized to 60 × 60, then undergo contrast enhancement, Multi-scale Retinex, histogram equalization, edge enhancement, and grayscale conversion.Contrast enhancement precedes Retinex to address low- and high-contrast images, since Retinex can produce halo-like artifacts in high-contrast regions and perform poorly on high-brightness images.Source: §3.1 · pp. 4, 6
- Multi-scale Retinex transforms images into aperture mode, reducing pixel dependence on illumination, then applies a nonlinear transformation to produce the output image.The authors identify halo-like artifacts in high-contrast regions and poor performance on high-brightness images as limitations.Source: §3.1 · pp. 4–6
- Preprocessing is intended to improve images affected by poor contrast, reduced brightness, low resolution, darkness, and blur; data augmentation addresses blurriness, perspective, and occlusion.The paper reports that preprocessing also makes good-resolution images sharper and visually enhanced.Source: §3.1 · p. 7
3.2 Deep Learning Architecture
- The 8-layer CNN uses three convolutional blocks followed by two fully connected layers, with 32, 128, and 256 filters across the blocks.The architecture is described as shallower than VGG16, with fewer parameters and better training convergence.Source: Figure 3 · p. 5, §3.2 · p. 7
4 Experimental Results
The experiments compare RIECNN across three traffic-sign benchmarks with varied image conditions and class distributions. RIECNN reports high recognition accuracy, short processing time, and gains from combining image-enhancement stages.
4.1 Dataset Preparation
- The evaluation uses GTSRB, BTSC, and rMASTIF, whose images vary in quality and whose sign classes are unbalanced, especially in BTSC.GTSRB has 43 classes; BTSC has 62 and rMASTIF has 31. The reported challenges include lighting, blur, perspective, aging, and class imbalance.Source: §4.1 · pp. 7–8
4.2 Performance Evaluation
- Training images were resized to 60×60 and augmented with shifts, zoom, shear, and rotation; the training data were split into 90% training and 10% validation.The experiments used mini-batches of 32, stochastic gradient descent, orthogonal initialization, and kernel L2 regularization.Source: §4.2 · pp. 8–9
- 0.8–1.3 ms per image is the reported average processing time for RIECNN.The paper compares total processing time with other techniques and notes a 16.67 ms frame interval at 60 fps, leaving about 15 ms for other applications.Source: §4.2 · p. 10
- 99.75% on GTSRB, 99.25% on BTSC, and 99.55% on rMASTIF; RIECNN is reported to outperform state-of-the-art architectures across all three benchmarks.The performance comparisons are shown in Tables 3, 4, and 5, respectively.Source: §4.2 · p. 10, Table 3 · p. 10, Table 4 · p. 11
- RIECNN achieves the highest-ranked accuracy with 4 times fewer parameters than STDNN on GTSRB and 2 times fewer than DHCNN on BTSC.Figure 9 compares accuracy against parameter count.Source: §4.2 · p. 10
4.3 Observation And Analysis
- 99.75% accuracy resulted from combining all four preprocessing stages on GTSRB, compared with 99.61% for contrast enhancement alone.Retinex with histogram equalization reached 99.62%, while Retinex with edge enhancement reached 99.6%; the authors identify contrast and brightness as the main GTSRB image problems.Source: §4.3 · p. 11
- Feature maps after four-stage preprocessing highlight finer distinctive details than maps after contrast enhancement alone.The authors report that conv1 focuses on shape and content, while conv2 focuses more on edges and invariant features.Source: §4.3 · pp. 11–12
- RIECNN correctly classifies challenging test images across GTSRB, rMASTIF, and BTSC, while some misclassified images are difficult for humans to distinguish.In Figure 15, green marks correct classifications and purple marks misclassifications.Source: §4.3 · p. 12
5 Conclusion
RIECNN combines image enhancement with a CNN and reports the highest recognition accuracy across three traffic-sign benchmarks. Its stated limitation is lower accuracy on the GTSRB EndOf and Speed subsets than other state-of-the-art architectures.
5 Conclusion
- RIECNN first preprocesses images to enhance their quality, then feeds them to a CNN; the authors report that Retinex combined with image-enhancement algorithms contributes to competitive recognition accuracy.Source: §5 · p. 12
- 99.75% recognition accuracy in GTSRB, 99.25% in BTSC, and 99.55% in rMASTIF; RIECNN reports the highest rank across all three benchmarks.Figure 9 plots accuracy against parameter count for GTSRB and BTSC.Source: §5 · p. 13, Figure 9 · p. 13
- Recognition accuracy on the GTSRB EndOf and Speed subsets is lower than that of other state-of-the-art architectures.Source: §5 · p. 13