A Domain Adaptive Semantic Segmentation Approach to Enhanced Adversarial Learning in Feature Spaces
A feature space and domain adaptive technology, applied in neural learning methods, instruments, biological neural network models, etc., can solve problems such as distorted feature distribution, unbalanced confrontation training, and inability to predict the structured output of the target domain. Adaptive, Enhanced Feature Encoder Effects
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
Embodiment 1
[0078] This embodiment proposes a domain-adaptive semantic segmentation method that enhances feature space adversarial learning, such as figure 1 , figure 2 As shown, the domain-adaptive semantic segmentation system based on enhanced feature space adversarial learning includes the following steps:
[0079] Step 1: Use a feature encoder to perform feature extraction on source domain images and target domain images to generate source domain image features and target domain image features;
[0080] Step 2: Perform feature space adversarial learning on source domain image features and target domain image features using a classification constraint discriminator;
[0081] Step 3: Using a combination of confrontational learning and pseudo-label self-training, the source domain image features and the target domain image features are segmented through the shared classifier, and the source domain image feature segmentation map and the target domain image feature segmentation map are o...
Embodiment 2
[0091] In this embodiment, on the basis of the above-mentioned embodiment 1, in order to better realize the present invention, further, in the step 2, in the classification constraint discriminator, the classifier component is used as a constraint, and the source domain characteristic of the discriminator is given Structural information, forcing the feature encoder to extract domain-invariant features containing structural information from the target domain to confuse the discriminator; in the process, after inputting the source domain image features and target domain image features into the classification constraint discriminator, the classification constraint discriminator is used loss For training, the specific formula is as follows:
[0092] ;
[0093] In the formula, the first two items on the right side of the equal sign are the loss function of the discriminator, and the third item is the auxiliary segmentation loss for the source domain output of the classifier; C ...
Embodiment 3
[0099] In this embodiment, on the basis of any one of the above-mentioned embodiments 1-2, in order to better realize the present invention, further, in the step 3:
[0100] For source domain data marked with : using forecast and the ground-truth label Between the cross entropy loss function to train the segmentation network, the specific loss function is as follows:
[0101] ;
[0102] where G1 represents the segmentation network, h and w represent the height and width of the input image, C2 represents the predefined category, and c represents the category.
[0103] In order to better realize the present invention, further, in the step 3:
[0104] For unlabeled target domain image features: Generate pseudo-labels for target domain image features during self-training, and use pseudo-labels to guide network updates. The specific operations are:
[0105] First, from the entire target domain data through the pre-trained model Select pixels with high prediction confide...
PUM
Login to View More Abstract
Description
Claims
Application Information
Login to View More 


