HDR Image Augmentation for Contrastive Learning on Fine-Grained Objects
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning machine learning models require robust training, which is often hindered by the need for large, annotated datasets and the limitations of current data augmentation methods in contrastive self-supervised learning, leading to poor performance in scenarios with fine-grained boundaries between object categories.
Innovation Solution
Employing High Dynamic Range (HDR) image generation as a data augmentation technique for contrastive self-supervised learning, using multiple exposures or synthetic methods to create diverse image pairs for training, which are then processed through convolutional neural networks and projection heads to enhance feature representation learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional image augmentation is used for training, then the training process is simple and fast, but the model performance deteriorates due to insufficient variability and quality of training images
Solution Approach 1:
The patent applies HDR augmentation by changing the dynamic range parameter of images. This involves transforming standard dynamic range images into high dynamic range images through multi-exposure capture or synthetic processing, thereby enhancing the quality and variability of training images to improve model performance without fundamentally changing the training architecture
Solution Approach 2:
The patent generates multiple synthetic views of the same image through HDR augmentation techniques. By creating multiple augmented versions of training images using different exposure levels and processing methods, the system increases data variability while maintaining the same underlying image content, thus improving model robustness without requiring additional real-world images
2Reliability
If more diverse and high-quality training images are used, then model performance improves, but the quantity and quality of available training data becomes insufficient
Solution Approach 1:
The patent synthesizes multiple augmented views from a single input image using HDR techniques. By generating multiple synthetic images with different exposure levels, contrast adjustments, and dynamic range characteristics from one original image, the system dramatically increases the effective training data quantity while maintaining high quality and diversity
Solution Approach 2:
The patent divides the training data into multiple augmented versions by segmenting the dynamic range of the original image. Through multi-exposure capture or synthetic processing, the image is split into different exposure levels and processing stages, creating multiple distinct but related training samples from a single source
3Measurement precision
If supervised learning with annotated data is used, then model training is accurate, but the process is time-consuming and expensive
Solution Approach 1:
The patent enables self-supervised learning by using the HDR augmented images themselves as the training signal. Instead of requiring external annotations, the system automatically generates supervision signals through the HDR augmentation process, allowing the model to learn from the structural and dynamic range information inherent in the augmented images, thereby eliminating the time-consuming annotation process while maintaining training accuracy
Data Source
Figure 1
Figure 2(a)~2(c)
Figure 3
AI summary
According to an aspect, there is provided a method that includes receiving a first image and a second image as inputs for contrastive self-supervised learning; applying a high dynamic range augmentation to the first image to generate a first pair of views; applying the high dynamic range augmentation to the second image to generate a second pair of views; applying a first convolutional neural network to the first pair of views to output a first pair of encoded representations; applying a second convolutional neural network to the second pair of views to output a second pair of encoded representations; projecting the first pair of encoded representations to form first projected representations; projecting the second pair of encoded representations to form second projected representations; and training a machine learning model using the high dynamic range augmentations and an objective function that provides contrastive self-supervised learning.