With the rapid development of AI in recent times industries are seeking to streamline repetitive tasks, such as quality inspection in order to eliminate subjective and biased judgments traditionally performed by humans. Due to the demand for implementing a consistent and fast-paced inspection system, this study aimed to perform defect detection on metal surface prints. Within the timeframe of this study, data collection on a variety of print designs was conducted, where it was found that the defects mainly were miscalibrated colors. To handle the mentioned task, experimentation between three separate systems was carried out, following a metric learning-based approach. An XGBoost classifier serving as a baseline, a triplet network processing histograms, and a triplet network processing images were implemented to compare how regular machine-learning systems compare to modern metric learning systems, as well as comparing histogram analysis to image analysis. Metric learning networks follow a template-matching approach and in this study, a pixel-perfect template and a printed paper representing a perfect print were initially available. During the data collection, the disconcerting gap from design files to photographs was seen when comparing the pixel-perfect template and paper template. To address this issue, a Convolutional Autoencoder (CoAE) was built to transform the design files into representations that look closer to a photograph of the design. To replicate camera artifacts a process was built that performs a camera’s process of taking a photograph but in reverse. Since the choice of an anchor within a metric learning network can have a significant impact on the performance, the experimentation also included four separate anchor choices between the pixel-perfect template, the printed paper template, and two CoAE-generated templates, with and without applied camera artifacts that were tested on the three defect detection systems.

The results of the experimentation pointed towards the XGBoost baseline being best in performance based on precision, recall, F1-score, and accuracy. The image triplet network tended to outperform the histogram triplet network, therefore image analysis was deemed superior. The best result was obtained by combining the XGBoost classifier and the CoAE anchor without applied camera artifacts. The pixel-perfect anchor was always better than the paper anchor for the metric learning networks. The CoAE-generated anchor without artifacts was typically better performing than the CoAE anchor with artifacts. The addition of camera artifacts made the anchors look closer to real photographs but they had generally worse performance. Although the performance of the different systems additionally can be improved, this study can provide usable information for further research in the area of defect detection on color variations.