Feature Map Generator for Semantic Segmentation Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing methods, such as semantic segmentation using neural networks, face challenges in training models effectively due to the need for large annotated datasets and the limitations of using synthetic images, which often result in decreased performance due to visual artifacts affecting low-level convolutional layers.
Innovation Solution
The method involves training a generator to produce feature maps instead of images, which are then used to improve the training of a second portion of the model, allowing for the generation of dense features that encode location information and details, thereby enhancing the model's performance. This approach uses a generator and discriminator in an adversarial manner, with the generator producing feature maps that can have a higher number of channels and lower resolution than input images, and includes the use of atrous spatial pyramid pooling to aggregate multi-scale information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If synthetic images are used for training, then the quantity of training data is increased, but the quality of semantic segmentation deteriorates due to visual artifacts
Solution Approach 1:
The patent introduces feature maps as an intermediary representation between real images and the segmentation model. Instead of directly using synthetic images that contain visual artifacts, the generator creates feature maps that serve as a mediating training data source, preserving the benefits of synthetic data generation while avoiding the harmful artifacts that would otherwise degrade segmentation quality
Solution Approach 2:
The patent creates synthetic feature maps as copies of the essential information contained in real image features, without copying the visual artifacts present in synthetic images. The generator learns to replicate the meaningful feature representations from real images while omitting the detrimental visual noise, thus providing abundant training data with preserved quality
2Measurement precision
If high resolution synthetic images are used for training, then the detail information is improved, but visual artifacts still affect low-level convolutional layers
Solution Approach 1:
The patent extracts only the useful feature information from real images to train the generator, while deliberately excluding the visual artifacts. The generator is trained to produce feature maps that contain the essential detail information needed for high-quality segmentation, but without the harmful artifacts that would interfere with low-level convolutional processing
Solution Approach 2:
Feature maps serve as an intermediary that decouples the detail information from the visual artifacts. The generator creates these intermediate representations that contain high-detail information suitable for training, while the harmful artifacts are filtered out in the feature map domain rather than being present in the final training data
3Manufacturing precision
If more annotated images are collected for training, then the model accuracy is improved, but the time and cost for annotation increase
Solution Approach 1:
Instead of manually annotating more images, the patent uses the generator to synthesize feature maps that copy the essential training signal from existing annotated data. This synthetic feature map generation provides abundant diverse training examples without requiring additional manual annotation time, thus maintaining model accuracy improvement while avoiding the time cost of annotating more images
Solution Approach 2:
The system uses the existing annotated images to train the generator, which then automatically generates additional training feature maps without requiring further human annotation effort. The generator serves itself to create the training data, eliminating the need for continuous manual annotation while still providing diverse training examples for improving model accuracy
Data Source
AI summary
A method for training a model to be used for processing images, wherein the model comprises: —a first portion (101) configured to receive images as input and configured to output a feature map, —a second portion (102) configured to receive the feature map outputted by the first portion as input and configured to output a semantic segmentation, the method comprising: —training a generator (201) so that the generator is configured to generate a feature map configured to be used as input to the second portion, —generating a plurality of feature maps using the generator, —training the second portion using the feature maps generated by the generator.