Image synthesis method and device, electronic equipment and computer readable storage medium
An image synthesis and sub-network technology, applied in the field of image processing, can solve the problems of inconsistent color of left and right eyebrows, asymmetry of left and right eyes, affecting the display effect of natural images, etc.
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
preparation example Construction
[0049] First, please refer to figure 1 , the embodiment of the present application provides an image synthesis method, which may include the following steps:
[0050]Step S110: Obtain a semantic segmentation map including a plurality of local regions and a sampling noise map of the same scale as the semantic segmentation map, each of the local regions has a class label for representing its category.
[0051] Step S120: Convert the semantic segmentation map into semantic and appearance-related semantic vectors according to the pre-acquired similarity information between objects represented by each class label.
[0052] Step S130: Obtain a composite image by processing the semantic vector and the sampled noise map.
[0053] Among them, when the method can be applied to image synthesis network.
[0054] like figure 2 As shown, the image synthesis network 200 may include a semantic vector generation sub-network 210 and a semantic rendering sub-network 220 .
[0055] Among the...
Embodiment approach
[0073] As mentioned above, the semantic vector generation sub-network 210 of the image synthesis network 200 may include a feature extraction unit 211 , and correspondingly, the semantic rendering sub-network 220 of the image synthesis network 200 also includes a residual calculation unit 221 . In this embodiment, the target semantic segmentation map can be directly input into the image synthesis network 200 (equivalent to downsampling the target semantic segmentation map with a preset multiple of 1), at this time, the feature extraction unit 211 obtains The first input information of is the target semantic segmentation map.
[0074] Furthermore, in some implementations, when the semantic vector generation sub-network 210 includes N (for example, 3) feature extraction units 211 , correspondingly, the semantic rendering sub-network 220 also includes 3 residual calculation units 221 . In this embodiment, scale conversion can be performed on the target semantic segmentation map i...
PUM
Login to View More Abstract
Description
Claims
Application Information
Login to View More 


