Identificação automática de adulterações em placas de circuito impresso controladoras de bombas de combustível baseada em imagens

Abstract

Frauds involving illegal modifications to fuel pumps harm customers and collaborate with crimes such as money laundering and tax evasion. Currently, the search for adulterations is performed by a human inspector through visual analysis, a tiring and unreliable practice due to the miniaturization of components. This paper presents an image-based approach to support the human inspector, detecting suspect regions on printed circuit boards for fuel pump controllers. The proposed approach begins with the alignment of a dataset of images, which may contain perspective distortions, noise, and illumination variations, to a reference image, using an image registration technique based on a planar projective transformation (homography), estimated from matches obtained by the Scale-Invariant Feature Transform (SIFT) algorithm. Aligned images are partitioned into regions, which are manipulated independently. Descriptors are extracted from each region, considering two algorithms: Histograms of Oriented Gradients (HOG) and a variation of the SIFT algorithm. An SVM classifier (Support Vector Machine) is trained for each region, using samples with and without tampering. Samples were extracted from perfect condition plates, with adulterated samples having been generated artificially. During inspection, the classifiers look at the input images by regions that differ significantly from the regions corresponding to the training samples. That allows the detection of unknown adulterations, not just those observed during training. The results can be integrated and presented as a heatmap to a human inspector, who has the final decision on the presence of adulterations. Experiments were performed on a dataset containing 649 images of a plate model, including images containing modification manually added. Using SIFT descriptors, was achieved a precision of 0.7739, a recall of 0.9638 and an F-measure of 0.8503 when we try to maximize the F-measure, indicating that our approach can effectively identify suspicious regions, providing invaluable help to the human inspector.