Marcelo Ricardo Stemmer

A Printed Circuit Board (PCB) is a fundamental component in the manufacturing process of electronic devices. When talking about the process of fabrication of these devices, the quality control stage represents, in general, the biggest cost in production. This stage is usually realized by humans, which are prone to fail, as well as being slower and less consistent when compared to computers. This study presents the utlization of six distincts Artificial Intelligence archtiectures trained for PCB defect detection. The results of the experiments highlighted key differences in training time, accuracy, and computational complexity across various models. YOGA-s achieved a good balance with the second-best mAP 50 accuracy (83.2%) and a moderate training time of 50 minutes. YOLOv11n, despite its shorter training time (40 minutes), outperformed YOGA-s in mAP 50-95 (42.7%). DETR demonstrated the highest accuracy (92.4% mAP 50), but at the cost of significantly longer training times and higher complexity. LW-DETR, while efficient, showed limited performance improvements as its scale increased. YOLOv11n emerged as the most cost-effective option for PCB inspection tasks.

Autonomous navigation in dynamic environments presents a significant challenge for mobile robotic systems. This paper proposes a novel approach utilizing Convolutional Neural Networks (CNNs) for multi-object detection in 3D space and 2D segmentation using bird’s eye view (BEV) maps derived from 3D Light Detection and Ranging (LiDAR) data. Our method aims to enable mobile robots to localize movable objects and their occupancy, which is crucial for safe and efficient navigation. To address the scarcity of labeled real-world datasets, a synthetic dataset based on a simulation environment is generated to train and evaluate our model. Additionally, we employ a subset of the NVIDIA r2b dataset for evaluation in the real world. Furthermore, we integrate our CNN-based detection and segmentation model into a Robot Operating System 2 (ROS2) framework, facilitating communication between mobile robots and a centralized node for data aggregation and map creation. Our experimental results demonstrate promising performance, showcasing the potential applicability of our approach in future assembly systems. While further validation with real-world data is warranted, our work contributes to advancing perception systems by proposing a solution for multi-source, multi-object tracking and mapping.

Autonomous navigation in dynamic environments presents a significant challenge for mobile robotic systems. In this paper, we propose a novel approach utilizing Convolutional Neural Networks (CNNs) for multi-object detection in 3D space and 2D segmentation using Bird’s Eye View (BEV) maps derived from 3D LiDAR data. Our method aims to enable mobile robots to localize movable objects and their occupancy, crucial for safe and efficient navigation. To address the scarcity of labeled real-world datasets, a synthetic dataset based on simulation environment is generated to train and evaluate our model. Additionally, we employ a subset of the NVIDIA r2b dataset for evaluation in real-world. Furthermore, we integrate our CNN-based detection and segmentation model into a ROS2-based framework, facilitating communication between mobile robots and a centralized node for data aggregation and map creation. Our experimental results demonstrate promising performance, showcasing the potential applicability of our approach in future assembly systems. While further validation with real-world data is warranted, our work contributes to the advancement of perception systems by proposing a solution for multi-source multi-object tracking and mapping.

The semantic segmentation of log faces constitutes the initial step towards subsequent quality analyses of timber,such as quantifying properties like mechanical strength, durability, and the aesthetic attributes of growth rings. In the literature, works based on both classical and machine learning approaches for this purpose can be found. However, more recent architectures and techniques, such as ViTs or even the latest CNNs, have not yet been thoroughly evaluated. This study presents a comparison of modern deep neural network architectures for cross-section segmentation in images of log ends. The results obtained indicate that the networks using the ViTs considered in this work outperformed those previously evaluated in terms of both accuracy and processing time.

Computer vision systems are increasingly used in industry for inspection or process control. The more demanding requirements observed today make the implementation of this type of systems a technological challenge. Many of the computational architectures available allow us to meet the main functional requirements related to the use case and also the non-functional ones, such as processing time restrictions and connectivity. However, the requirement for adaptability so that such systems can be easily modified to meet different use cases or even accommodate environmental changes remains a challenge. This work proposes a flexible architecture for computer vision systems using FPGA. This architecture combines components of the processing flow of the vision system implemented in hardware and software to obtain advantages associated with both approaches. The proposed solution is validated against different real use cases existing in the industry. The results obtained allow us to affirm that such architecture brings an interesting advantages since it meets the operational requirements present in industrial applications, demands less development effort and can be easily adapted to new usage scenarios.

In this paper we present an incremental variant of the Twin Support Vector Machine (TWSVM) called Fuzzy Bounded Twin Support Vector Machine (FBTWSVM) to deal with large datasets and learning from data streams. We combine the TWSVM with a fuzzy membership function, so that each input has a different contribution to each hyperplane in a binary classifier. To solve the pair of quadratic programming problems (QPPs) we use a dual coordinate descent algorithm with a shrinking strategy, and to obtain a robust classification with a fast training we propose the use of a Fourier Gaussian approximation function with our linear FBTWSVM. Inspired by the shrinking technique, the incremental algorithm re-utilizes part of the training method with some heuristics, while the decremental procedure is based on a scored window. The FBTWSVM is also extended for multi-class problems by combining binary classifiers using a Directed Acyclic Graph (DAG) approach. Moreover, we analyzed the theoretical foundations properties of the proposed approach and its extension, and the experimental results on benchmark datasets indicate that the FBTWSVM has a fast training and retraining process while maintaining a robust classification performance.