Consumer Product Recognition Using Ranked Tags and Two-Stage Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current AI systems lack a uniform process for creating consistently accurate and adaptive object recognition training sets, struggle with speed and accuracy in object detection, and are ineffective in recognizing partially obscured objects, leading to inefficiencies in consumer product identification.
Innovation Solution
A consumer product identification system utilizing a camera system, neural network, recognition application, and training module to create high-quality training sets, improve classification methodology, and apply a two-step process for precise object recognition, combining YOLO for broad classification with local template matching for detailed identification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing methods are used for object recognition, then a relatively good degree of accuracy can be achieved, but the process becomes cumbersome and tedious from a data processing perspective
Solution Approach 1:
The patent segments the object recognition process into distinct stages: data collection, data processing, model training, and inference. By dividing the training data into multiple datasets with different levels of abstraction and detail, the system can process information more efficiently at each stage while maintaining high accuracy in the final recognition results.
Solution Approach 2:
The patent applies preliminary action by pre-processing and curating training data before the actual model training begins. This includes creating standardized datasets, pre-labeling objects, and preparing reference images in advance, which significantly reduces the computational burden and time required during the actual recognition process.
2Productivity
If fast object recognition methods are used, then processing speed is improved, but accuracy decreases and only very broad results are produced
Solution Approach 1:
The patent implements a multi-level recognition architecture where broad categorization is performed first using fast processing methods, then progressively refined through subsequent stages that apply more detailed analysis only where needed. This segmentation allows the system to achieve both speed and accuracy by applying computational resources selectively.
Solution Approach 2:
The patent applies local quality by providing different levels of recognition detail to different objects or regions based on their importance or characteristics. High-priority or ambiguous detections receive more detailed analysis while clear, confident detections are processed quickly, optimizing the balance between speed and accuracy across the entire system.
3Adaptability or versatility
If AI systems are trained with standard data sets, then basic object recognition can be achieved, but the systems struggle to recognize partially obscured objects and lack adaptive thinking
Solution Approach 1:
The patent applies preliminary action by pre-training the AI system with specialized datasets that include partially obscured objects, occluded views, and varied lighting conditions. This preparatory training enables the system to develop robust recognition capabilities for challenging scenarios while maintaining consistent performance through standardized evaluation protocols.
Solution Approach 2:
The patent implements feedback mechanisms where the system continuously learns from its recognition attempts, particularly from cases involving partially obscured objects. User corrections and system self-evaluation are fed back into the training process, progressively improving the system's ability to handle occlusion while maintaining reliable and consistent identification through validated performance metrics.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A machine learning based system and method for automated recognition of consumer products in images and video using a camera system, a neural network using a ranked tagging system, a two-stage recognition application, and a training module. Training image sets of items captured by the camera system are assigned identification tags through template matches to training sets within the neural network. Tags are assigned from various levels of specificity to identify exact product matches. A user recognition application captures images and generates bounding boxes for detected objects and assigns a general classification within the image using a single, fast, convolutional neural network (CNN) layer. General classification narrows subsets for each generated bounding box and multi-scale template matching is applied to achieve detailed identification of single or multiple items detected within single images or video. The training module adjusts smart camera systems and the neural network based on accuracy feedback.