Deep Learning-Based Flange Dimension Measurement for Industrial Parts by Integrating Self-Supervised Binary Image Segmentation and Contour Detection

This is the authors' abstract. We don't add key points for this paper.

This study proposes a deep learning-based flange dimension measurement framework (DL-FDM) for industrial metrology, integrating a self-supervised binary image segmentation network (BIS-Net) with contour detection. The DL-FDM framework addresses the limitations of manual measurement methods (e.g., inefficiency and inaccuracy) and mitigates the environmental sensitivity of traditional machine vision techniques. First, a self-supervised BIS-Net is developed to segment industrial components from complex backgrounds, generating high-precision binary images while eliminating preprocessing steps such as filtering, denoising, and manual threshold adjustment. Contour detection algorithms are then applied to extract dimensional outlines, with least-squares fitting employed to calculate precise geometric measurements. Experimental validation employed a CMOS area-array camera and a GTX1050 workstation to evaluate flange outer and inner diameters under varying environmental conditions. Each dimension underwent ten repeated measurements, with experimental analysis (referenced to caliper measurements) yielding a maximum standard deviation of 0.046 mm and a mean absolute error of 0.053 mm under normal conditions. Compared to state-of-the-art methods, the proposed DL-FDM achieves superior performance, exhibiting a maximum relative error of 0.384%. These results confirm that the framework satisfies the high-precision demands of intelligent manufacturing systems. By bridging self-supervised learning with industrial metrology, this work presents a reproducible solution for robust, high-accuracy measurement in varying environments, advancing the integration of AI-driven approaches in precision manufacturing.