Product Recognition Model Retraining via Selective Error Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing product recognition systems face challenges in improving accuracy while managing the increasing processing load caused by large volumes of image data from actual store operations.
Innovation Solution
A processing system that selectively accumulates only incorrect recognition images as training data, performs relearning when a predetermined threshold of incorrect images is reached, and updates the estimation model to improve product recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If all images from actual operations are accumulated as training data, then product recognition accuracy is improved, but processing load on the computer increases
Solution Approach 1:
The patent extracts only the necessary subset of training data (images with incorrect recognition results) from the entire dataset of images from actual operations. This selective extraction approach maintains the quality of training data while significantly reducing the volume of data that needs to be processed, thereby resolving the contradiction between improving recognition accuracy and reducing processing load.
Solution Approach 2:
The patent segments the training data into two categories: images with correct recognition results and images with incorrect recognition results. By focusing relearning only on the segmented subset of incorrect images, the system achieves improved accuracy without needing to process the entire dataset, thus managing processing load effectively.
2Measurement precision
If relearning is performed frequently to improve accuracy, then product recognition accuracy is improved, but processing load on the computer increases
Solution Approach 1:
The patent implements dynamic control of relearning frequency based on the accumulation of incorrect recognition images. Relearning is triggered only when a predetermined number of incorrect images are accumulated, rather than performing relearning continuously or at fixed intervals. This dynamic approach allows the system to adaptively balance between improving accuracy and managing processing load.
Solution Approach 2:
The patent performs preliminary accumulation of incorrect recognition images before initiating relearning. By accumulating images up to a predetermined threshold, the system prepares a sufficient dataset for effective relearning without triggering frequent small-scale relearning operations, thus reducing overall processing load while maintaining accuracy improvement.
Data Source
AI summary
An image acquisition unit (11) acquires a recognition processing image. A recognition unit (12) recognizes a product in the recognition processing image based on an estimation model. A registration unit (13) registers a result of the recognition in recognized product information. An output unit (14) outputs a result of the recognition. A correction reception unit (16) receives an input for correcting a result of the recognition. A correction unit (17) changes a result of the recognition to a result of the recognition after a correction, and also stores correction information in which a result of the recognition after a correction and the recognition processing image are associated with each other. A learning unit (18) performs relearning by using the recognition processing image stored as the correction information and updates the estimation model, when a number of the recognition processing image stored exceeds a predetermined value.