arXiv subjects

Julen Balzategui

Publications and source records attributed to Julen Balzategui.

3 recordsLinked to original sources

arXiv2026

A Machine Learning Framework for Real-Time Personalized Ergonomic Pose Analysis

This paper introduces a new methodology for real-time prediction of ergonomic and non-ergonomic human poses using volumetric video data in three dimensions. Although the methodology was designed for ergonomic assessments, it can be adapted to other applications requiring real-time analysis of human posture. One aspect that makes this system stand out is its ability to analyze 3D point clouds during the assessment, enabling computation from multiple angles. This overcomes a critical limitation of cameras which provide often a fixed viewpoint, thereby restricting the data available for a thorough postural evaluation, especially when occlusions occur. The system continuously and automatically performs pose inference using the chosen perspective on the real-time streaming data; however, only the poses manually selected and labeled by the user are used to train the personalized deep learning classifier. The methodology has been refined through a case study in which RGB-D cameras captured subjects performing load-lifting tasks, enabling real-time skeletal labeling. The model was trained on this data and, following the training phase, performs inference on new streaming data in real time. This research offers a scalable and pragmatic approach for real-time ergonomic evaluation by combining state-of-the-art 3D data technologies and traditional 2D pose estimation algorithms. It addresses the increasing need for safety and health monitoring in workplace environments, marking a notable contribution to the domain.

cs.CV↗

arXiv2022

Few-shot incremental learning in the context of solar cell quality inspection

In industry, Deep Neural Networks have shown high defect detection rates surpassing other more traditional manual feature engineering based proposals. This has been achieved mainly through supervised training where a great amount of data is required in order to learn good classification models. However, such amount of data is sometimes hard to obtain in industrial scenarios, as few defective pieces are produced normally. In addition, certain kinds of defects are very rare and usually just appear from time to time, which makes the generation of a proper dataset for training a classification model even harder. Moreover, the lack of available data limits the adaptation of inspection models to new defect types that appear in production as it might require a model retraining in order to incorporate the detects and detect them. In this work, we have explored the technique of weight imprinting in the context of solar cell quality inspection where we have trained a network on three base defect classes, and then we have incorporated new defect classes using few samples. The results have shown that this technique allows the network to extend its knowledge with regard to defect classes with few samples, which can be interesting for industrial practitioners.

cs.CV↗

arXiv2021

Anomaly detection and automatic labeling for solar cell quality inspection based on Generative Adversarial Network

Quality inspection applications in industry are required to move towards a zero-defect manufacturing scenario, withnon-destructive inspection and traceability of 100 % of produced parts. Developing robust fault detection and classification modelsfrom the start-up of the lines is challenging due to the difficulty in getting enough representative samples of the faulty patternsand the need to manually label them. This work presents a methodology to develop a robust inspection system, targeting thesepeculiarities, in the context of solar cell manufacturing. The methodology is divided into two phases: In the first phase, an anomalydetection model based on a Generative Adversarial Network (GAN) is employed. This model enables the detection and localizationof anomalous patterns within the solar cells from the beginning, using only non-defective samples for training and without anymanual labeling involved. In a second stage, as defective samples arise, the detected anomalies will be used as automaticallygenerated annotations for the supervised training of a Fully Convolutional Network that is capable of detecting multiple types offaults. The experimental results using 1873 EL images of monocrystalline cells show that (a) the anomaly detection scheme can beused to start detecting features with very little available data, (b) the anomaly detection may serve as automatic labeling in order totrain a supervised model, and (c) segmentation and classification results of supervised models trained with automatic labels arecomparable to the ones obtained from the models trained with manual labels.

cs.CV↗