Abstract
This work presents a novel methodology for the quality assessment of material extrusion parts through AI-based Computer Vision. To this end, different techniques are integrated using inspection methods that are applied to other areas in additive manufacturing field. The system is divided into four main points: (1) pre-processing, (2) color analysis, (3) shape analysis, and (4) defect location. The color analysis is performed in CIELAB color space, and the color distance between the part under analysis and the reference surface is calculated using the color difference formula CIE2000. The shape analysis consists of the binarization of the image using the Canny edge detector. Then, the Hu moments are calculated for images from the part under analysis and the results are compared with those from the reference part. To locate defects, the image of the part to be analyzed is first processed with a median filter, and both the original and filtered image are subtracted. Then, the resulting image is binarized, and the defects are located through a blob detector. In the training phase, a subset of parts was used to evaluate the performance of different methods and to set the values of parameters. Later, in a testing and validation phase, the performance of the system was evaluated using a different set of parts. The results show that the proposed system is able to classify parts produced by additive manufacturing, with an overall accuracy of 86.5%, and to locate defects on their surfaces in a more effective manner.

Abstract
Inspecting reflective parts is challenging due to strong specular reflections that conceal small porosities and reduce defect visibility. This work presents a framework that combines augmented reality with a deep learning detector. An augmented reality headset is used to capture multi-view images under natural illumination, enabling the operator to adjust the viewpoint and obtain angles that reduce glare. The collected data form a 640 × 480 dataset used to train a yolov8 detection model, integrated into a Robot Operating System 2 architecture for real-time processing. Testing on an independent set of unseen parts yields a precision of 86.70 %, a recall of 87.26 %, and an F1-score of 86.97 %. Additional qualitative examples confirm that the model can identify low-contrast porosities despite reflective surfaces. The results demonstrate the feasibility of AR-assisted acquisition combined with deep learning for real-time inspection of machined aluminum components in a laboratory case study. © 2026 IEEE.