Generator Network Training for Precise Object Count Image Synthesis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Training image classifiers with a sufficient quantity of varied training images is time-consuming and expensive, especially for rare scenarios, and manual annotation is required to provide true classification scores.
Innovation Solution
A method using generative adversarial networks (GANs) to generate synthetic images with specific object counts, involving a generator network that maps noise samples and target object counts to generated images, and a discriminator network to improve the accuracy of object count prediction and distinguish between real and synthetic images, with adversarial training to enhance realism and object count matching.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If real training images are collected through long test drives, then sufficient quantity and variability of training images are obtained, but time consumption and cost increase significantly
Solution Approach 1:
The patent uses generative adversarial networks to create synthetic copies of real training images. The generator network produces artificial images that mimic the statistical properties and visual characteristics of real images, eliminating the need for extensive physical data collection while providing sufficient quantity and variability for training image classifiers.
Solution Approach 2:
The generator network is trained with conditional inputs that allow control over specific parameters of the generated images, such as the number and types of objects present. By adjusting these conditional parameters, the system can generate diverse training images with controlled characteristics, providing sufficient variability without requiring extensive real-world data collection.
2Measurement precision
If manual annotation is performed to obtain true classification scores, then accurate training labels are obtained, but time consumption and cost increase significantly
Solution Approach 1:
The discriminator network is trained to distinguish between real annotated images and synthetic images. During this process, it learns to recognize the ground truth labels associated with real images. This trained discriminator can then be used to transfer annotation information to synthetic images without requiring manual annotation, significantly reducing time and cost while maintaining accuracy.
Solution Approach 2:
The system uses the discriminator network to automatically annotate synthetic images by transferring label information from real images. This self-annotation process eliminates the need for manual annotation of synthetic data, allowing the system to generate its own training labels automatically.
3Adaptability or versatility
If conditional GANs are used to generate synthetic images of a certain mode, then images belonging to a particular class are produced, but control over specific object counts is not achieved
Solution Approach 1:
The patent modifies the conditional input structure of the generator network to explicitly include object count information. By changing the parameters of the conditional input to include specific counts of different object types, the generator can precisely control the number and types of objects in generated images while maintaining class-specific characteristics.
Solution Approach 2:
The discriminator network provides feedback during training by evaluating both the visual quality and the object count accuracy of generated images. This feedback mechanism guides the generator to improve its control over object counts while maintaining the ability to produce images of particular classes, achieving both adaptability and precision.
Data Source
Figure 1
Figure 2a~3
Figure 4~5
AI summary
A method (100) for training a generator network (1) that is configured to generate images (13) with multiple objects (14a-14d), comprising: • providing (110) a set of training images (11); • providing (115) a generator network (1); • providing (120) a discriminator network (2); • drawing (130) noise samples (17) and target counts (16a-16d) of objects (14a-14d); • mapping (140), by the generator network (1), the noise samples (17) and target counts (16a-16d) of objects (14a-14d) to generated images (13); • randomly drawing (150) images (21) from a pool (P) comprising generated images (13) as well as training images (11); • supplying (160) the randomly drawn images (21) to the discriminator network (2), thereby mapping them to a combination of: a decision (23) whether the respective image (21) is a training image (11) or a generated image (13); • optimizing (170) discriminator parameters (22) that characterize the behavior of the discriminator network (2) with the goal of improving the accuracy with which the discriminator network (2) distinguishes between generated images (13) and training images (11); • optimizing (180) generator parameters (12) that characterize the behavior of the generator network (1) with the goal of deteriorating said accuracy; and • further optimizing (190) both the generator parameters (12) and the discriminator parameters (22) with the goal of improving the match between the predicted count (18a-18d) of objects (14a-14d) on the one hand, and the actual (15a-15d) or target (16a-16d) count of objects (14a-14d) on the other hand.