Incremental Model Training for Multiview Checkout Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine-Learning Models (MLMs) for multiview item recognition face challenges in accurately classifying items due to seasonal changes, new items, and occlusions, requiring tedious manual labeling and leading to inaccuracies, which can cause friction during checkouts.
Innovation Solution
An incremental training method where operators at checkouts visually identify unknown items, scan their barcodes, and label image patches, with additional images captured and scored for subsequent training sessions, allowing the MLM to improve accuracy over time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the MLM is trained with a large amount of labeled data to improve accuracy, then the classification precision improves, but the training time and manual labeling effort increase significantly
Solution Approach 1:
The system performs preliminary actions by capturing multiple images of items during the normal checkout process before training is needed. These images are stored and later used for incremental training, eliminating the need for time-consuming manual data collection and labeling during the training phase.
Solution Approach 2:
The system enables self-service by automatically capturing images during checkout, automatically labeling them with item information from barcode scanners or RFID readers, and automatically incorporating them into the training dataset. This eliminates manual intervention in the data collection and labeling process.
2Adaptability or versatility
If the MLM is trained extensively to recognize all possible items, then the adaptability improves, but the training complexity and resource requirements increase
Solution Approach 1:
The system implements dynamic adaptability through incremental training, where the model continuously learns from new items encountered during checkouts. Rather than requiring comprehensive pre-training, the model adapts dynamically as new items are introduced to the system, maintaining simplicity while improving coverage over time.
Solution Approach 2:
The system prepares training data in advance by capturing images during normal operations. When new items are encountered, their images are immediately captured and labeled, creating a ready-to-use training dataset that simplifies future training processes and reduces complexity.
3Reliability
If the MLM is updated frequently with new data to maintain accuracy, then the reliability improves, but the processing time and operational friction increase
Solution Approach 1:
The system ensures continuous improvement of model reliability through incremental training that occurs in the background without interrupting checkout operations. Images captured during normal operations are continuously added to the training dataset, and model updates are performed periodically without stopping the checkout process, maintaining both reliability and productivity.
Solution Approach 2:
The system performs preliminary data collection during normal operations, so that when training is needed, the data is already prepared. This eliminates the need for time-consuming data collection during training updates, allowing faster model retraining and maintaining high checkout processing speeds.
Data Source
AI summary
Multiple images of multiple items are captured of a transaction area during a checkout. The Red-Green-Blue (RGB) data associated with each item image patch is collected across the images and provided as input by a Machine-Learning Model (MLM), which returns an item code for the item. When the given MLM is unable to satisfactorily predict an item code for a given set of image patches, the patches associated with the images are presented to an operator of a checkout and the operator is asked to scan an item barcode for that item. The patches are labeled within the images with the item code and additional images of the item are captured and labeled with the item code when the barcode is scanned by the operator. The labeled images are used in a subsequent training session with the MLM to improve its item recognition accuracy for the item.