Neural Network Object Recognition with Limited Input Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning algorithms require a large number of input images to properly train neural networks for object recognition, which is inefficient for objects that are similar in appearance, leading to increased processing complexity and maintenance needs.
Innovation Solution
A computer-implemented method that updates a neural network trained to recognize a first object rendition by applying predetermined image mappings, such as geometric, illumination, and obscuration transformations, to a limited set of model images of a second object rendition, generating training images for updated training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large number of input images are used to train a neural network for object recognition, then recognition accuracy is improved, but processing complexity and maintenance requirements increase
Solution Approach 1:
The patent creates synthetic training images by applying geometric transformations (rotations, flips, scaling) to a small set of original training images. These copied and transformed images serve as additional training data, enabling the neural network to achieve high recognition accuracy without collecting and processing a large number of actual input images, thereby reducing processing complexity
Solution Approach 2:
The patent systematically varies parameters such as rotation angles, flip directions, and scaling factors to generate diverse training images from a single original image. By changing these transformation parameters, the system creates multiple unique training samples that improve recognition accuracy across different object orientations and appearances without increasing the base number of required images
2Measurement precision
If training is performed on each variant of an object separately, then recognition accuracy for each variant is improved, but processing time and maintenance effort increase
Solution Approach 1:
The patent develops a universal image transformation framework that can be applied to any object variant using the same set of geometric transformations. Instead of creating separate training pipelines for each object variant, the system uses a single versatile approach that works across different products, designs, and versions, significantly reducing maintenance effort and processing time while maintaining recognition accuracy
3Measurement precision
If a large set of training images is required for each new object, then the neural network can recognize the object accurately, but system maintenance complexity increases
Solution Approach 1:
When a new object needs to be added to the recognition system, the patent applies the same image transformation process to generate training images from a small set of new object photos. This copying approach eliminates the need to manually collect large datasets for each new product, simplifying the onboarding process and reducing maintenance complexity while ensuring the neural network achieves accurate recognition for the new object
Data Source
AI summary
Systems and methods for object recognition with limited input are disclosed herein. An example method includes updating a neural network trained to perform object recognition on a first rendition of an object, so that the neural network performs object recognition on a second rendition of the object, using a limited set of input images. The method includes receiving a limited set of model images of the second rendition of the object, accessing a corresponding image mapping, and generating a large number of training images from the limited set, where image mappings include geometric, illumination, and/or obscuration transformations. The neural network is then trained, from this initial small set, to classify the second rendition of the object.