Discriminant Learning-based Colorspace for Blade Segmentation

TL;DR

CSDA tackles suboptimal colorspace representations in segmentation by learning a multidimensional nonlinear colorspace that maximizes inter-class separability while minimizing intra-class variance. It extends Linear Discriminant Analysis with a signed between-class variance and three trainable losses, integrated into an end-to-end segmentation framework using a U-Net. On wind turbine blade data, CSDA yields improved blade-background discrimination and segmentation metrics across multiple windfarms, while maintaining fast inference. This work demonstrates the practical importance of task-specific colorspace preprocessing for robust, domain-adapted image segmentation.

Abstract

Suboptimal color representation often hinders accurate image segmentation, yet many modern algorithms neglect this critical preprocessing step. This work presents a novel multidimensional nonlinear discriminant analysis algorithm, Colorspace Discriminant Analysis (CSDA), for improved segmentation. Extending Linear Discriminant Analysis into a deep learning context, CSDA customizes color representation by maximizing multidimensional signed inter-class separability while minimizing intra-class variability through a generalized discriminative loss. To ensure stable training, we introduce three alternative losses that enable end-to-end optimization of both the discriminative colorspace and segmentation process. Experiments on wind turbine blade data demonstrate significant accuracy gains, emphasizing the importance of tailored preprocessing in domain-specific segmentation.

Figures (6)

• Figure 1: CSDA framework. Joint optimization of colorspace and segmentation via discriminant and probability losses.
• Figure 2: Visualization of colorspace transformations. On the left side, RGB input (top) and ground-truth mask (bottom). On the right, the transformed image at various $d_{CS}$ for CSDA$^{(\Delta)}$ with its corresponding estimated mask on its right; accuracy, F1, and mIoU on top. See \ref{['sec:visual_colorspace']} for details.
• Figure 3: Segmentation performance across model variants and colorspace dimensions. Detailed descriptions of the model variants are provided in \ref{['sec:ablation']}.
• Figure 4: Qualitative comparison on test images for DDA, Focal and CSDA. First two columns: input image and ground-truth mask $\mathbf{M}$. Next: estimated segmentation masks from DDA$^{(\Delta)}$, DDA$^{(\ln)}$, Focal, and CSDA$^{(\Delta)}$ with accuracy, F1, and mIoU shown above.
• Figure 5: Qualitative results for CSDA$^{(\Delta)}$ with $d_{CS}=4$. First two columns: input image and ground-truth mask $\mathbf{M}$. Third to fourth column: transformed colorspace image, see \ref{['sec:visual_colorspace']}. Fifth column: estimated mask. Last columns: individual transformed colorspace channels.