Vertex Interpolation for Synthetic Training Data Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for training convolutional neural networks to recognize retail items from single images face challenges due to data sparsity and the need for diverse object positions and perspectives, with conventional augmentation techniques often degrading image quality and failing to replicate real-world object warping.
Innovation Solution
The method generates synthetic training data by overlaying a grid on a source image, perturbing vertex coordinates, and interpolating pixel values to create multiple augmented images that reflect realistic object variations, preserving image integrity and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If conventional image augmentation methods (resizing or cropping) are used to modify a single image of the object, then the quantity of training data is increased, but the image quality deteriorates and the augmentation fails to accurately reflect natural object warping
Solution Approach 1:
The patent applies parameter changes by systematically varying multiple image parameters simultaneously (scaling factors, rotation angles, translation offsets, and warping coefficients) to generate diverse training samples. This allows creating realistic object variations that maintain image quality while reflecting natural warping effects that occur in real-world scenarios.
Solution Approach 2:
The patent introduces an intermediary warping model that acts as a mediator between the original image and augmented versions. This model applies physics-based warping transformations that accurately represent natural object deformations, serving as a bridge that preserves image quality while generating realistic variations for training data augmentation.
2Quantity of substance
If a single image of a retail item is provided, then the data sparsity is reduced for the distributor, but the convolutional neural network struggles to recognize the item due to insufficient diverse training examples
Solution Approach 1:
The patent applies preliminary action by pre-generating a comprehensive set of augmented training images from the single provided item image before training the convolutional neural network. This preliminary data preparation ensures that the network is exposed to diverse object positions, scales, and warping variations in advance, improving recognition reliability without requiring multiple original images from the distributor.
Solution Approach 2:
The patent uses copying by creating multiple synthetic copies of the original item image through various geometric transformations and warping operations. These copied and transformed images serve as additional training examples, allowing the neural network to learn from diverse variations while the distributor only needs to provide a single original image.
3Reliability
If magnitudes of data are required to achieve high generalization in object recognition, then the recognition accuracy is improved, but the requirement for data collection and processing becomes impractical
Solution Approach 1:
The patent applies segmentation by decomposing the complex task of generating diverse training data into systematic geometric transformations applied to a single source image. By segmenting the augmentation process into distinct operations (scaling, rotation, translation, warping), the system achieves high generalization capability without requiring manual collection and processing of magnitudes of diverse images.
Data Source
AI summary
A computer implemented method for generating synthetic training data to train a convolutional neural network is described. The method consists of steps including receiving a source image depicting an object for identification. The type and shape of the depicted object is determined. The source image is overlayed with a N×M grid of vertices, the grid including horizontal and vertical edges and being fit to the shape of the depicted object. For each vertex in the grid, perturbations are added to the (x,y) coordinates of the vertex and the pixel values in a range between the original and final (x,y) coordinates are interpolated, resulting in the generation of an item of synthetic training data. The method is repeated to generate multiple items of synthetic training data which are then used to train a neural network to identify the object in an image.