Pseudo Label Estimation for Unlabeled Image Attribute Identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing technologies face challenges in accurately estimating attributes from unlabeled training data, leading to deteriorated identification accuracy of learning models, as they often rely on the same identification target region for learning, which can be difficult for certain images.
Innovation Solution
An image processing apparatus that includes a pseudo label estimation unit to estimate pseudo labels based on the identification target region according to the attribute type, allowing the learning unit to learn a model using these pseudo labels, thereby improving attribute identification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If the attribute is estimated from the same identification target region as the learning model, then the learning process is simplified, but the attribute estimation fails for certain images leading to deteriorated identification accuracy
Solution Approach 1:
The patent divides the attribute estimation process into two independent stages: (1) detecting the identification target region using a region detection model, and (2) estimating the attribute using a attribute estimation model. This segmentation allows each model to specialize in its function, with the region detection model handling diverse image types and the attribute estimation model focusing on accurate attribute prediction from the detected region.
Solution Approach 2:
The patent introduces an intermediary region detection model that acts as a mediator between the image input and the attribute estimation model. This intermediary component detects and localizes the identification target region first, then passes this information to the attribute estimation model, enabling reliable attribute estimation even when the target region has specific characteristics that would otherwise prevent direct estimation.
2Quantity of substance
If unlabeled training data is used to expand training coverage, then the quantity of training data increases, but the attribute cannot be estimated from the same region leading to reduced learning quality
Solution Approach 1:
The patent applies preliminary action by first detecting the identification target region in unlabeled images before performing attribute estimation. The region detection model processes unlabeled training data to identify and localize relevant regions, which are then used as input for the attribute estimation model. This preliminary region detection enables effective utilization of unlabeled data that would otherwise be unusable for training.
Data Source
AI summary
According to one embodiment, an image processing apparatus 1 includes one or more hardware processors configured to function as an acquisition unit 20A, a pseudo label estimation unit 20B, and a learning unit 20C. The acquisition unit 20A acquires unlabeled training data including an image to which a correct label of an attribute is unassigned. The pseudo-label estimation unit 20B estimates a pseudo-label, which is an estimation result of the attribute of the image of the unlabeled training data, based on an identification target region according to a type of the attribute to be identified by a first learning model 30 to be learned in the image of the unlabeled training data. The learning unit 20C learns the first learning model 30 identifying the attribute of the image by using first labeled training data with the pseudo-label being assigned to the image of the unlabeled training data.