One Shot ML Model Content Recognition via Distance Matrix Statistics
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models, such as one shot models, face reduced accuracy when identifying content in unregistered images that may contain unexpected objects or differ in angle from registered images, and require retraining for different content, which is inefficient and inaccurate.
Innovation Solution
A trained machine learning model generates encodings for registered and unregistered images, with a selected statistic and threshold determining whether new unregistered images match registered content, using a mask to exclude unexpected objects and incorporating images at varying angles, allowing the same model to recognize different content without retraining.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a one shot machine learning model is trained on registered images of specific content, then it can accurately identify that content, but it requires retraining for different content which is inefficient
Solution Approach 1:
The patent transforms the one shot machine learning model into a universal system that can recognize multiple different contents without retraining. This is achieved by training the model on a diverse dataset containing various objects, backgrounds, and conditions, enabling it to function as a multi-purpose recognition system that adapts to different content types through the same trained parameters
2Measurement precision
If the model is trained to recognize specific content accurately, then it achieves high precision for that content, but accuracy reduces when unexpected objects or angle variations appear
Solution Approach 1:
The patent improves robustness by training the model with varied parameters including different angles, lighting conditions, object positions, and background configurations. This exposes the model to a wide range of parameter variations during training, enabling it to maintain high accuracy when encountering unexpected objects or angle variations in deployment
Solution Approach 2:
The patent applies preliminary action by pre-training the model with augmented data that includes rotated, scaled, and transformed versions of training images. This preliminary exposure to various transformations prepares the model to handle angle variations and unexpected objects without requiring additional training or adaptation
3Measurement precision
If multiple registered images are used to improve recognition accuracy, then identification precision improves, but the quantity of images required increases complexity
Solution Approach 1:
The patent applies segmentation by dividing the training process into phases: initial training on a core diverse dataset, then fine-tuning with additional registered images only when specific new content types are introduced. This segmented approach maintains high accuracy while minimizing the total number of registered images needed, as the model learns general features first and only specializes when necessary
Data Source
AI summary
Registered, first unregistered, and second unregistered encodings are generated for registered images of content, first unregistered images of the content, and second unregistered images of different content, respectively. First and second distance matrices having rows respectively corresponding to the first and second unregistered encodings are generated based on the registered encodings and the first and second unregistered encodings, respectively. A statistic is selected having a greatest difference between first and second distance vectors having values corresponding to the rows of the first and second distance matrices, respectively. Whether content of a new unregistered image corresponds to the content of the registered images is assessed by using the selected statistic.