Garment Image Extraction With Edge Smoothing and Depth-Aware Segmentation

The patent segments the image processing task into multiple stages: first performing binary classification to identify foreground/background, then applying bilateral filtering to smooth edges, and finally using quantized matting to refine boundary regions. This multi-stage segmentation approach resolves the contradiction by breaking down the simple binary classification into finer-grained processing steps that progressively improve image quality while maintaining manageable complexity at each stage.

The patent applies different processing qualities to different regions of the image: bilateral filtering is applied globally to smooth edges, while quantized matting with alpha matting is applied specifically to boundary regions where foreground and background meet. This local quality approach ensures that computational resources are concentrated on regions where they are most needed (edge regions) while maintaining efficiency in uniform regions.

The patent uses the binary classification result as a coarse copy or approximation of the final segmentation mask, then refines it through automated bilateral filtering and quantized matting processes. This copying approach allows the system to start with a simple, low-cost binary mask and progressively improve it through automated processing rather than requiring expensive manual labeling at the finest detail level from the start.

The patent performs preliminary binary classification and bilateral filtering before applying the more computationally intensive quantized matting algorithm. This preliminary action prepares the data by providing a good initial estimate and smoothing the input, which reduces the complexity and cost of the subsequent matting process while still achieving high-quality results.