Traffic Sign Detection (TSD) is a complex and fundamental task for developing autonomous vehicles; it is one of the most critical visual perception problems since failing in this task may cause accidents. This task is fundamental in decision-making and involves different internal conditions such as the internal processing system or external conditions such as weather, illumination, and complex backgrounds. At present, several works are focused on the development of algorithms based on deep learning; however, there is no information on a methodology based on descriptive statistical analysis with results from a solid experimental framework, which helps to make decisions to choose the appropriate algorithms and hardware. This work intends to cover that gap. We have implemented some combinations of deep learning models (MobileNet v1 and ResNet50 v1) in a combination of the Single Shot Multibox Detector (SSD) algorithm and the Feature Pyramid Network (FPN) component for TSD in a standardized dataset (LISA), and we have tested it on different hardware architectures (CPU, GPU, TPU, and Embedded System). We propose a methodology and the evaluation method to measure two types of performance. The results show that the use of TPU allows achieving a processing training time 16.3 times faster than GPU and better results in terms of precision detection for one combination.

Automating industrial processes, particularly quality inspection, is a key objective in manufacturing. While welding tasks are frequently automated, inspection processes remain largely manual. Advances in computer vision and AI, especially ViTs, now enable more effective defect detection and classification, offering opportunities to automate these workflows. This study evaluates ViTs for identifying defects in aluminum welding using the Aluminum 5083 TIG dataset. The analysis spans binary classification (detecting defects) and multiclass categorization (Good Weld, Burn Through, Contamination, Lack of Fusion, Misalignment, and Lack of Penetration). ViTs achieved 98% to 99% accuracy across both tasks, significantly outperforming prior models such as dense and CNNs, which struggled to surpass 80% accuracy in binary and 70% in multiclass tasks. These results, achieved with datasets of 2400 to 8000 images, highlight ViTs’ efficiency even with limited data. The findings underline the potential of ViTs to enhance manufacturing inspection processes by enabling faster, more reliable, and cost-effective automated solutions, reducing reliance on manual inspection methods.

Traffic sign detection is a crucial task in autonomous driving systems. Due to its importance, several techniques have been used to solve this problem. In this work, the three more common approaches are evaluated. The first approach uses a model of the traffic sign which is based in color and shape. The second one enhances the image model of the first approach using K-means for color clustering. The last approach uses convolutional neural networks designed for image detection. The LISA Traffic Sign Dataset was used which it was divided into three superclasses: prohibition, mandatory, and warning signs. The evaluation was done using objective metrics used in the state-of-the-art.

Detecting defects in Tungsten Inert Gas (TIG) aluminium welding is crucial for industries such as marine, aerospace and automotive, where weld quality impacts safety and performance. This study investigates the performance of Vision Transformer (ViT) models for detecting welding defects in the aluminium 5083 TIG dataset. In addition, this study evaluates ViTs on embedded Graphics Processing Unit (GPU) boards using benchmarks for resource-constrained devices. Although many studies have explored Machine Learning (ML) or Deep Learning (DL) techniques for welding defect detection, they often overlook performance evaluation metrics, such as inference speed, on embedded hardware. We provide a comparative analysis of ViT models, including Swin Transformers and traditional ViTs, evaluating defect detection performance using metrics like F1-score, precision, recall, accuracy, Frames-Per-Second (FPS), and memory consumption on embedded GPU hardware. Results from the TIG Aluminium 5083 dataset show that the DeiT-Base model achieved an accuracy of 99.83%, outperforming existing models. These results confirm the effectiveness of ViTs for automated welding defect detection on both conventional computers and resource-constrained embedded systems. Finally, we successfully implemented ViT-based models on embedded hardware, showing robust performance even under limited computational resources.