Abstract
In industrial quality control, maintaining consistent shape perception across varying imaging conditions is critical for automated downstream tasks. This paper introduces SPENet (Shape-Aware Efficient Network), a lightweight, real-time semantic segmentation model optimized for industrial environments. By incorporating a Shape-aware Spatial Path (SSP), a Variable Boundary Domain (VBD) for handling fuzzy edges, and a novel Consistency Mean Square Error (CMSE) metric, SPENet achieves state-of-the-art accuracy and a 50% improvement in segmentation stability. On a specialized PVC coating dataset, the model reaches 95.88% mIoU at 81 FPS with only 3.71M parameters.
Introduction
Automated industrial inspection, such as monitoring PVC coating on vehicle undercarriages, demands extreme reliability. Standard semantic segmentation models often struggle with "shape perception," where the predicted contours of fixed components (e.g., bolts or holes) fluctuate due to lighting changes or sensor noise. Such inconsistencies can disrupt robotic systems that rely on precise, stable measurements.
Furthermore, these environments require real-time performance on edge hardware. While high-accuracy models are often computationally prohibitive, existing lightweight alternatives frequently produce "jagged" edges or inconsistent shapes. SPENet addresses these challenges by prioritizing both efficiency and structural consistency.
Methodology
SPENet utilizes a dual-path architecture designed to balance global semantic context with fine-grained spatial detail.
Dual-Path Structure and Spatial Path
The model features a semantic path based on a streamlined ResNet-18 for global context and a high-resolution Shape-aware Spatial Path (SSP) dedicated to contour extraction. To optimize for the linear nature of object edges, the SSP employs asymmetric convolutions ($1 \times n$ and $n \times 1$) and dilated convolutions. This approach expands the receptive field while minimizing parameter count.
Boundary Supervision and Decoupled Module
A Decoupled Module separates features into low-frequency "body" and high-frequency "edge" components, allowing for specialized supervision. To handle the inherent noise and fuzziness of real-world industrial boundaries, the authors introduced the Variable Boundary Domain (VBD). Instead of a binary pixel line, VBD uses a Gaussian function to create a probability map based on the distance from class transitions, making the model more robust to imaging artifacts.
Consistency Metric
To quantify the reliability of predictions over time or across datasets of fixed components, the Consistency Mean Square Error (CMSE) was developed. CMSE measures the variance in predicted shapes; a lower value indicates higher stability, which is crucial for industrial robotics.
Experimental Results
The model was evaluated on a specialized PVC Coating Dataset (500 image-label pairs) and the standard Cityscapes benchmark.
Performance on PVC Dataset
SPENet demonstrated superior performance, achieving 95.88% mIoU at 81 FPS on a GTX 3060. It significantly outperformed established models like DDRNet and BiseNet. Notably, SPENet is highly efficient, containing only 3.71M parameters compared to DDRNet's 20.1M.
Stability and Generalization
In terms of consistency, SPENet achieved a CMSE of $27.09 \times 10^{-4}$, a reduction of over 50% compared to previous top-performing models. This stability ensures that automated measurements remain reliable across different frames. While optimized for industrial scenes with relatively stable camera positions, the model also showed competitive performance on the dynamic Cityscapes dataset, proving its versatility.
Conclusion
SPENet provides a robust solution for real-time industrial inspection by focusing on shape consistency and boundary awareness. The introduction of the Variable Boundary Domain and the CMSE metric offers new tools for both supervising and evaluating segmentation models in precision-critical applications. Future work will explore the application of VBD to other architectures and broader industrial visual tasks.
Summary updated