Bounding Box Auto Segmentation for Low-Label Food Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current food identification technologies using deep learning for image recognition face challenges due to the need for large amounts of manually labeled data, which is time-consuming and costly, and the low accuracy of food type identification, leading to errors in calorie counting and inefficient food management.
Innovation Solution
A method and apparatus for auto segmentation using a bounding box through weakly-supervised learning, where a segmentation model is trained using bounding box labels and color similarity maps to classify objects and backgrounds, with auto-labeling through user inspection for prediction errors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual labeling is used to obtain training data for deep learning models, then data quality can be ensured, but time consumption and cost increase significantly
Solution Approach 1:
The system uses automatically generated bounding box labels from detector models to train segmentation models, which then perform auto-labeling on new images. This self-service mechanism eliminates the need for manual pixel-level annotation while maintaining data quality through iterative model improvement and edge-case focused learning.
Solution Approach 2:
The approach performs preliminary bounding box labeling using detector models before segmentation labeling. This preliminary action provides a foundation that reduces subsequent manual work, as the bounding boxes serve as pre-processing that guides the segmentation model training and reduces the need for complete manual annotation.
2Measurement precision
If segmentation models are trained with pixel-level labels, then segmentation accuracy improves, but labeling cost and complexity increase dramatically
Solution Approach 1:
The methodology segments the labeling process into two stages: first generating bounding box labels using detector models, then using these bounding boxes to train segmentation models for pixel-level segmentation. This segmentation of the labeling task reduces complexity by breaking down the difficult pixel-level annotation into easier bounding box annotation followed by automated segmentation.
Solution Approach 2:
Bounding box labels serve as an intermediary between simple object detection and complex pixel-level segmentation. The detector model generates bounding boxes that act as intermediate training data, which then trains the segmentation model to achieve pixel-level accuracy without requiring direct manual pixel annotation.
3Reliability
If more training data is collected to improve model performance, then model accuracy increases, but data processing time and storage requirements increase
Solution Approach 1:
The system focuses on learning from edge-case data where the model does not provide satisfactory predictions rather than processing all available data uniformly. This partial action approach improves model accuracy by concentrating computational resources on difficult cases that benefit most from additional learning, rather than processing excessive amounts of already-well-handled data.
Solution Approach 2:
The methodology changes the parameter of data selection from random or uniform sampling to edge-case focused sampling. By identifying and prioritizing difficult prediction cases, the system optimizes data processing efficiency while maintaining or improving model accuracy through targeted learning on challenging examples.
Data Source
AI summary
Provided are a method and an apparatus for auto segmentation using a bounding box. A method for auto segmentation using a bounding box according to one embodiment of the present disclosure comprises receiving a first object image including an object labeled with a bounding box, which is a pre-learning target, learning a segmentation model by classifying an object and a background from the bounding box of the received first object image, and segmenting an object from a second object image, which is an identification target, using the learned segmentation model.