Deep Learning-Based Segmentation for Defect Detection in Metal Additive Manufacturing: A Custom Neural Network Approach

This paper presents an approach to defect detection in metal additive manufacturing using deep learning-based object detection. This is an implementation of a custom neural network architecture with a simplified convolutional backbone to identify and localize defects in manufactured metal components. This model employs dual output heads for bounding box coordinate prediction and defect type identification. An exploration of various optimization strategies including network architecture modifications, training procedure enhancements, and detection quality improvements were done. Experimental results demonstrate that this approach achieves a mean average precision of 0.236 in defect detection, with significantly better performance for workpiece defects compared to nozzle defects. The model generates a fixed set of 100 potential detections per image, with an overall precision of 0.109, recall of 0.128, and F1-score of 0.118. Despite modest performance metrics, the proposed method establishes a baseline approach for automated defect detection in metal additive manufacturing