Image Data Augmentation Using Random Pixel Displacement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computer vision technologies face challenges in generating a large number of diverse images for training deep learning models, particularly in early stages of new product development or manufacturing processes, and deep learning methods operate as a black box, making it difficult to understand image generation rationality.
Innovation Solution
A data augmentation method involving random displacement of pixels in input images to generate multiple output images with diverse deformations, using a computing device equipped with a storage circuit and processing circuit to execute instructions for pixel displacement and randomization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If deep learning technologies are used to generate images, then image generation capability is improved, but the requirement for vast amounts of training images increases
Solution Approach 1:
The patent applies the copying principle by creating synthetic training images through pixel displacement and randomization operations on existing images. The processing circuit generates multiple output images by copying and transforming input images, thereby replicating diverse image data without requiring actual diverse images for training. This resolves the contradiction by enabling deep learning models to be trained on synthesized data that replicates the diversity needed for effective training while avoiding the need to collect vast amounts of real images.
Solution Approach 2:
The patent applies parameter changes by systematically varying displacement parameters and randomization parameters to generate diverse output images from a single input image. By changing the displacement magnitude, direction, and randomization seeds, the system creates multiple variants of images that simulate different scenarios. This allows the model to learn from parameter variations rather than requiring actual varied images, thus improving image generation capability while reducing the quantity of training images needed.
2Productivity
If deep learning models are used for image generation, then image generation capability is improved, but the interpretability of the generation process deteriorates
Solution Approach 1:
The patent replaces the black-box deep learning mechanism with a transparent pixel manipulation mechanism. Instead of using complex neural networks that operate as black boxes, the system uses direct pixel displacement and randomization operations that are easily interpretable. The processing circuit applies deterministic transformations to input images, making the generation process visible and explainable. This substitution maintains image generation capability while dramatically improving interpretability, as users can trace exactly how input images are transformed into output images through documented pixel-level operations.
3Reliability
If a large number of diverse images are collected for training, then model performance is improved, but the difficulty of data collection increases
Solution Approach 1:
The patent applies the self-service principle by enabling the system to generate its own training data through automated pixel displacement and randomization operations. Instead of requiring manual collection of diverse images, the system takes a single input image and automatically generates multiple diverse output images through computational transformations. This self-service approach maintains high model performance by producing diverse training data while eliminating the difficult manual data collection process, as the system creates its own training dataset through algorithmic manipulation.
Solution Approach 2:
The patent applies preliminary action by pre-processing input images through pixel displacement and randomization before they are used for training. The system performs these transformations in advance to create a diverse set of training images from a single input, preparing the data needed for model training ahead of time. This preliminary action eliminates the need for subsequent manual data collection, as the diverse training data is already generated through automated preliminary processing operations.
Data Source
AI summary
A data augmentation method includes obtaining an input image and creating a plurality of output images corresponding to the input image. At least one first pixel of the input image is displaced to form one output image. The displacement of each first pixel is randomized. This overcomes the challenge when training data is scarce.