Two-Stage GAN Data Generation for Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection algorithms using machine learning often misrecognize classes with insufficient data, such as a North Korean tank, due to inadequate training data, and struggle to generate high-resolution images effectively.
Innovation Solution
A GANs-based data generation apparatus and method that includes multiple generators and discriminators performing machine learning using DNNs, CNNs, and RNNs, generating images in two steps to produce higher resolution images and improve object recognition by increasing the size of fake images relative to target images, thereby reducing misrecognition caused by insufficient data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If a single GAN model is used to generate images for object detection, then the generation process is simple and fast, but the image resolution is insufficient leading to poor object recognition performance
Solution Approach 1:
The patent divides the single GAN model into two separate GAN models: a first GAN model for generating initial fake images and a second GAN model for generating high-resolution fake images. This segmentation allows each model to specialize in different resolution levels, thereby improving overall image resolution without requiring a single overly complex model to handle all resolution requirements simultaneously.
Solution Approach 2:
The patent introduces an additional dimensional aspect to the image generation process by creating a two-stage generation pipeline. The first GAN operates in one dimension (standard resolution) and the second GAN operates in another dimension (high resolution), effectively using dimensional progression to achieve higher resolution outputs while managing model complexity through staged processing.
2Reliability
If more training data is collected to improve object recognition for rare classes, then recognition accuracy improves, but data collection time and cost increase
Solution Approach 1:
The patent applies preliminary action by using the first GAN model to generate initial fake images that serve as training data for the second GAN model. This preliminary generation of training data allows the second model to learn and produce high-resolution images without requiring extensive manual data collection, thereby reducing time loss while improving recognition accuracy for rare object classes.
Solution Approach 2:
The patent uses synthetic copying through GAN-generated fake images to create additional training data for rare object classes. Instead of collecting more real images (which would be time-consuming), the system copies and synthesizes additional training examples from available data, improving recognition accuracy without the time penalty of extensive data collection.
3Manufacturing precision
If high-resolution images are generated directly in one step, then the process is efficient, but the image quality and resolution are insufficient for accurate object detection
Solution Approach 1:
The patent segments the high-resolution image generation process into two distinct stages: the first GAN generates intermediate-resolution images, and the second GAN processes these to produce final high-resolution images. This segmentation improves image quality by allowing each model to optimize for its specific resolution target, while maintaining productivity through the streamlined two-stage pipeline rather than attempting single-step generation.
Solution Approach 2:
The first GAN performs preliminary action by generating intermediate fake images that serve as input for the second GAN. This preliminary generation step prepares the data in a format suitable for the second model, enabling efficient high-resolution generation without requiring the first model to directly produce final high-resolution images, thus balancing quality and efficiency.
Data Source
AI summary
An apparatus for generating data based on generative adversarial networks (GANs) is provided. The apparatus includes a first generator configured to receive input data and generate a first fake image, a first discriminator configured to receive the first fake image generated by the first generator and a first real image and verify whether an image is fake or real, a second generator configured to receive the first fake image and generate a second fake image; and a second discriminator configured to receive the second fake image and a second real image and verify whether an image received by the second discriminator is fake or real. The apparatus improves an object recognition performance as compared with a typical object recognition algorithm by generating an image of a class of which a sufficient database is not secured so as to reduce an object misrecognition caused by an insufficient amount of data.