Detecção e segmentação de modificações em placas de circuito impresso em ambientes não controlados através de métodos de uma classe e autoencoders

Abstract

The visual inspection of printed circuit boards (PCBs) is an important problem in areas such as quality control in production lines and fraud detection. Inspection involves detecting assembly errors and identifying the replacement or addition of new components that were not originally specified. Such changes can interfere with the correct behavior of the circuit. Detecting alterations through automated visual inspection is a viable alternative that accelerates the production process or the detection of fraud by non-specialists. However, anomaly detection presents challenges, especially when there is an imbalance between data classes and few anomaly samples available. Additionally, visual inspection in uncontrolled environments, such as anomaly detection in PCBs, is complex due to external variations, such as lighting, angle, and perspective. This work proposes a deep learning-based approach using convolutional autoencoders for the detection and segmentation of anomalies in PCB images in uncontrolled environments. The main objective is to develop a method that can operate in real conditions, trained only with images without anomalies. This eliminates the need to collect a large volume of images with anomalies for training, which can be impractical or costly. The contributions of this work include defining an effective cost function and comparison function for training and evaluating convolutional autoencoders. These functions enable the detection and segmentation of anomalies based on high-level features rather than individual pixels, making the method more robust in uncontrolled environments. Additionally, the labeled dataset MPI-PCB was created, containing PCB images for training, evaluation, and testing of the proposed method. This dataset was made publicly available, along with anomaly masks for comparison with other methods. Validation experiments of the method were performed against comparative methods from the literature PaDiM, SPADE, DFR and SPTM, using the proposed image dataset. The proposed method achieved the best results in all analyzed regions of the image, obtaining an Intersection over Union of 0.677 compared to the best comparative method’s 0.507, the PaDiM. Furthermore, experiments were conducted on the MVTec-AD anomaly detection image dataset, demonstrating the method’s effectiveness in other applications. Overall, this work provides a robust solution to the described problem and contributes to the goal of developing methods capable of detecting anomalies in uncontrolled environments.