Image Data Augmentation for Night-Vision Recognition Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Training recognition models for images captured by night-vision devices is challenging due to the lack of diverse and complete training data, leading to poor recognition accuracy.
Innovation Solution
An image data augmentation device and method that automatically or semi-automatically crops object images from pre-stored images and randomly collages them onto images obtained by night-vision devices, generating additional training sample images to enhance model training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a model trained on color images is used to identify night-vision images, then the recognition accuracy deteriorates
Solution Approach 1:
The patent creates synthetic night-vision training images by copying and transforming existing color images. Specifically, it converts color images to grayscale, adds noise, adjusts brightness, and combines them with object images to generate synthetic night-vision-like images. This allows the model to be trained on realistic night-vision data without requiring actual night-vision photographs, thereby improving recognition accuracy for night-vision images.
Solution Approach 2:
The patent applies multiple parameter transformations to color images to simulate night-vision conditions. These include converting color to grayscale, adding Gaussian noise, adjusting brightness and contrast, and modifying image resolution. By changing these image parameters, the system generates training data that closely resembles actual night-vision images, enabling the model to adapt to low-light conditions.
2Reliability
If the quantity of images including specific objects in the database is insufficient, then the recognition accuracy deteriorates
Solution Approach 1:
The patent uses object images (such as human figures, vehicles, or other objects) and combines them with processed background images to generate synthetic training images. This copying approach allows the system to create multiple variations of the same object in different night-vision-like backgrounds, significantly increasing the quantity of training images available for specific objects without requiring actual photographs of each object in night conditions.
Solution Approach 2:
The patent segments the image generation process into separate components: object images and background images are processed independently, then combined. The background images undergo transformation (grayscale conversion, noise addition, brightness adjustment) separately from the object images. This segmentation allows for flexible combination of objects with various background conditions, increasing diversity and quantity of training data.
3Quantity of substance
If data augmentation methods are used to generate training data, then the quantity and diversity of training data improve
Solution Approach 1:
The patent performs preliminary processing of images during the training data generation phase. Color images are pre-processed to convert them to grayscale, add noise, and adjust brightness before being combined with object images. This preliminary action ensures that the synthetic images closely resemble actual night-vision images, improving data quality without requiring complex real-time processing during model inference.
Solution Approach 2:
The patent uses an intermediary processing pipeline that includes image conversion tools, noise generation modules, and image combination algorithms. These intermediary components facilitate the transformation of color images into synthetic night-vision-like images. The intermediary processing layer abstracts the complexity of image transformation, providing a manageable interface between the input color images and the output synthetic training data.
Data Source
AI summary
An image data augmentation device includes a memory and a processor. The memory is configured to store multiple commands and multiple images; the processor is connected to the memory to read the multiple images and execute multiple commands to: identify at least one object-related image including at least one object from the multiple images; capture at least one object image from the at least one object related image; and superimpose the at least one object image on multiple arbitrary positions in the multiple images to generate multiple training sample images, and using the multiple training sample images to execute machine learning. In addition, an image data augmentation method is also disclosed here.