Abstract
Surface roughness classification is a critical component of industrial quality control, traditionally reliant on slow and invasive contact-based methods. This paper presents a novel texture descriptor, the Aggregated Discrete Fourier Transform (ADFT), designed for high-speed, non-contact classification of industrial castings. By integrating global frequency-domain analysis with local binary patterns, the proposed hybrid system achieves over 96% accuracy. The method demonstrates robustness across varying optical properties and generalizes well to standard texture benchmarks, offering a viable solution for real-time factory floor inspection.
Introduction
In modern manufacturing, surface roughness ($R_a$) directly influences a component's mechanical performance, including fatigue strength and corrosion resistance. The industry standard for measurement involves contact profilometers, which utilize a diamond stylus to trace surface topography. Despite their accuracy, these instruments are time-consuming, costly, and risk damaging the specimen.
Computer vision presents a non-contact alternative; however, classifying the roughness of industrial castings is inherently challenging. Casting surfaces are typically isotropic and irregular, lacking the structured patterns found in machined or polished surfaces. Furthermore, varying material properties lead to diverse optical behaviors—ranging from diffuse (matte) to specular (shiny) reflections—which significantly alter image data for identical roughness levels.
Methodology
The core contribution of this work is the Aggregated Discrete Fourier Transform (ADFT), a global texture descriptor that mimics the linear sampling of a physical stylus.
The ADFT Pipeline
The proposed architecture combines global frequency information with local spatial patterns to create a comprehensive feature set.
- 1D-DFT Processing: To maintain computational efficiency and simulate mechanical tracing, the algorithm applies a one-dimensional Discrete Fourier Transform (DFT) to every row and column of the input image.
- Feature Aggregation: DFT coefficients are aggregated into two distinct histograms:
- ADFT_M (Amplitude): Represents the energy distribution across frequencies.
- ADFT_GP (Phase Gradient): Captures phase differences between adjacent rows and columns, providing a stable signature for isotropic textures.
- Hybrid Integration: The global ADFT features are concatenated with Completed Local Binary Patterns (CLBP), which capture fine-grained local details.
- Distance Metrics: The researchers utilized the $l^4$ distance for the amplitude component. Unlike standard Euclidean or Chi-square metrics, the $l^4$ norm emphasizes differences in dominant frequency bins, enhancing the discriminative power between roughness classes.
Experimental Results
The system was evaluated on two datasets: a custom Casting Texture Database (253 samples, 8 roughness classes) and the standard CUReT benchmark.
Performance on Industrial Castings
The combined ADFT + CLBP approach achieved a recognition rate of 96.6% on the casting database. This performance significantly exceeds standalone Local Binary Pattern (LBP) methods (86.1%) and specialized filters like VZ-MR8 (81.9%). Notably, the feature extraction process requires only 0.038 seconds per image, meeting the requirements for high-speed online inspection.
Generalization
On the CUReT database, the method reached 97.28% accuracy. This indicates that the ADFT descriptor is a powerful general-purpose tool for texture analysis beyond specific industrial applications.
Discussion and Limitations
The synergy between global frequency analysis and local pattern recognition is central to the method's success. While ADFT captures the broad "rhythm" of the surface, CLBP identifies the "tiny bumps." However, some limitations remain. The ADFT descriptor is currently invariant only to 90-degree rotations due to its row/column sampling strategy. Additionally, while the phase-gradient component (ADFT_GP) is effective for isotropic castings, its performance degrades on anisotropic textures. Future research will focus on enhancing phase-information robustness for a broader range of material textures.
Summary updated