Vision-Based Item Detection Using Occlusion-Trained Recognition Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current item checkout methods are inefficient and inaccurate when dealing with multiple items, especially when they are occluded, as existing techniques require extensive training on various images and manual annotation, making them time-consuming and error-prone.
Innovation Solution
A vision-based detection system that uses two machine-learning algorithms to identify multiple items with occluded views by extracting features and assigning bounding boxes, reducing the need for exhaustive training data and manual annotation, and enabling detection of multiple items with improved accuracy using a smaller dataset.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If multiple items are placed on checkout counter for recognition, then item checkout efficiency is improved, but item recognition accuracy deteriorates due to occlusion
Solution Approach 1:
The system segments the item recognition problem into two distinct processing stages: a detection model that identifies item locations and generates bounding boxes, and a recognition model that classifies items within those bounding boxes. This segmentation allows each model to specialize, with the detection model handling occluded items by identifying their locations even when partially hidden, and the recognition model focusing on classifying items based on visible features within the detected regions.
Solution Approach 2:
Bounding boxes serve as an intermediary between the detection and recognition processes. The detection model generates bounding boxes that encapsulate items including occluded ones, and these bounding boxes are passed to the recognition model which then performs classification. This intermediary structure allows the system to maintain item identification capability even when full item views are not available.
2Measurement precision
If exhaustive training data with all possible item combinations is collected, then item recognition accuracy is improved, but training time and computational resources increase significantly
Solution Approach 1:
The training process is segmented into two independent model training pipelines: training data is used to train both the detection model and the recognition model separately. This eliminates the need to train on all possible item combinations simultaneously, as each model learns from its specialized training set. The detection model learns to identify item locations and generate bounding boxes, while the recognition model learns to classify items within those boxes.
Solution Approach 2:
The system uses partial action by training the detection and recognition models on separate, specialized datasets rather than requiring complete coverage of all possible item combinations. The detection model is trained to identify item locations even with partial visibility, and the recognition model is trained on item classification, allowing the system to achieve high accuracy without exhaustive training data.
3Reliability
If manual annotation of training images is performed to ensure quality, then training data quality is improved, but the effort and cost of data preparation increases
Solution Approach 1:
The annotation process is segmented into two distinct tasks: the detection model is trained with annotations focused on item locations and bounding boxes, while the recognition model is trained with annotations focused on item identities and categories. This segmentation allows each model to receive targeted, high-quality annotations relevant to its specific function, improving overall system reliability without requiring a single exhaustive annotation process.
Solution Approach 2:
The system enables self-service in data preparation by using the detection model to automatically generate bounding boxes that are then used by the recognition model. This automated bounding box generation reduces the manual annotation effort required, as the detection model's output serves as input for the recognition model, eliminating the need for separate manual bounding box drawing for each item.
4Measurement precision
If a single item is scanned at a time during checkout, then item identification accuracy is improved, but checkout process time increases
Solution Approach 1:
The system merges the detection and recognition functions into a unified two-stage processing pipeline that operates on all items in the field of view simultaneously. The detection model processes the entire image to identify all item locations, and the recognition model then processes all detected items in parallel, allowing multiple items to be identified at once rather than requiring sequential scanning of each item.
Solution Approach 2:
The system replaces the mechanical sequential scanning process with an automated vision-based detection and recognition system. Instead of requiring physical handling or sequential pointing of each item, the camera captures an image of all items, and the machine learning models automatically detect and recognize multiple items simultaneously, eliminating the need for manual item-by-item processing.
Data Source
AI summary
Item recognition of a given item is trained on a single item from different views. The item recognition is then trained on images of the given item partially occluded by a second item having same, similar, or different shapes and features to that of the given item. General features of the item are noted and used to detect the given item when the given item is presented with multiple different items having multiple different occluded views.