Dish image segmentation and removal method based on instance segmentation and spatial prior clustering

By combining YOLO instance segmentation, GrabCut, and K-means color clustering, plates are automatically removed from food images, solving the interference problem in food recognition and nutritional estimation, and achieving efficient and accurate food image processing.

CN122289690APending Publication Date: 2026-06-26NANJING UNIV OF AERONAUTICS & ASTRONAUTICS
0 Cites 0 Cited by

Patent Information

Application Number
CN202610475421.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-04-10
Publication Date
2026-06-26

AI Technical Summary

Technical Problem

Existing technologies struggle to automatically and accurately remove diverse plates from food images, especially plates of different shapes and colors, which interferes with food recognition and nutritional estimation tasks.

Method used

The YOLO instance segmentation model is used to extract images of individual dishes. The GrabCut algorithm generates a transparent background. Combined with K-means color clustering with spatial prior constraints, the plate region is automatically identified and removed. By using the prior constraint that the outer edge of the image must be a plate and the center must be a dish, K-means clustering and connectivity verification are used to prevent misjudgment.

Benefits of technology

It enables automatic removal of plates of different colors and shapes, improves batch processing efficiency, ensures the purity and robustness of food images, and supports food analysis in diverse dining scenarios.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention discloses a dish image segmentation and removal method based on instance segmentation and spatial prior clustering, and the method comprises the three steps: 1, carrying out the reasoning of a picture through a YOLO instance segmentation model, extracting a segmentation mask, and cutting the segmentation mask into a black background image; 2, foreground extraction is carried out on the image through a GrabCut algorithm, a transparent background is converted, and an RGBA format image is generated; and step 3, performing spatial partitioning based on Euclidean distance transformation on the transparent background image, performing K-means clustering in a Lab color space by taking the periphery as dish priori and the center as dish priori, automatically judging dish and dish clusters according to the spatial enrichment degree ratio of each cluster at the periphery and the center, removing dish pixels by combining external reachable connectivity verification, and obtaining dish and dish data. And outputting the transparent background image only containing the dishes. The method does not need to manually mark areas or set color threshold values, can adaptively process dishes with different colors, shapes and sizes, is suitable for diversified meal scenes, and supports tasks such as dish identification and nutrition estimation.
Need to check novelty before this filing date? Find Prior Art