.png){kind=icon}
.png){kind=icon}
Return
LEARNABLE BOUNDARY ENHANCED SOFTMAX LAYER BASED ON GRAY MORPHOLOGY FOR DATA-DRIVEN IMAGE SEGMENTATION
.png){kind=icon}
OA
AI
DOI：10.3934/ipi.2025041.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The data-driven segmentation model is fundamentally a pixel-level image classification task in which it generally neglects the overall geometric structure of the segmentation object. The commonly used encoder-decoder architecture can inadvertently overlook boundary information during down-sampling, leading to a subpar segmentation. Consequently, the development of a data-driven image segmentation model that effectively preserves boundaries is of paramount importance. This can be accomplished by integrating the boundary constraints of variational segmentation models, which aid in preserving edges, with the proficiency of data-driven methods in extracting scenario-specific features. Thus, we leverage a traditional morphological algorithm to unroll a learnable boundary extraction module and integrate the boundary prior into the boundary detection function of the soft threshold dynamics (STD) model to prevent the smoothness of boundary pixels. Because the STD model serves as the variational explanation of the Softmax activation function, it can be unrolled into a network module through its iterative format for final classification, thereby creating a data-driven image segmentation model with boundary prior. Experiments conducted on three open-access medical datasets demonstrate that the model effectively preserves the boundary structure and enhances the results of segmentation. Ultimately, the feasibility and robustness of the model are validated.
Keywords:
Image segmentation
morphology
image boundary
variational method
deep Learning
Journal
I
IF:
1.5
Papers:
59
Citations:
0