CVPR 2022 𞓜 a creative and aesthetic text generation method! Support any input

This article briefly introduces CVPR 2022 Papers hired “Aesthetic Text Logo Synthesis via Content-aware Layout Inferring” . This paper aims to explore the process of character logo image design Automatic layout generation in . This paper is based on conditional countermeasure generation network （conditional-GAN）, A dual discriminator structure and a differentiable splicing module are proposed , According to the visual and semantic information of the input text , The layout geometric parameters of each glyph are deduced , So as to synthesize the text identification image . This method can assist graphic design and other visual tasks related to text . The data set and code related to this work have been open source （ See the end of the article ）.

The paper : https://arxiv.org/abs/2204.02701
Data sets and code : https://github.com/yizhiwang96/TextLogoLayout

One 、 Research background

Text mark （Text Logo） The design of depends very much on the creativity and experience of the designer , among , How to arrange the layout of each text element is a core problem . Layout design needs to consider many factors , Such as glyph 、 Literal semantics 、 Theme, etc . Pictured 1 Shown , There is usually no shape overlap between different words ; Line breaks or column breaks in Chinese signs are usually in the word element （Token） after ; Emphasize the meaning of words , Larger sizes are usually used ; Geometric transformations such as oblique cutting and rotation can respectively embody themes such as a sense of strength and a sense of joy . Most of the existing schemes in the industry are to design a set of rules that are easy to implement , Design the layout according to some preset templates , However, the results are often monotonous and lack creativity and beauty . In response to this question , This paper proposes a content aware text logo image generation model , From a large number of existing words Logo Learn layout design rules implicitly , Thus, a new font can be generated for any input font Logo.

chart 1 Common layout types in text logo images

 

Two 、 Data sets

Training AI Models usually require a lot of data , However, there is no data set for this task in the industry . To solve this problem , This paper proposes TextLogo3K Data sets , With the help of Tencent video platform , collect 、 Marking the 3,470 Zhang's carefully selected words Logo chart , these Logo From movies 、 Cover of TV series and animation . Pictured 2 Sum graph 3 Shown , The dataset accurately annotates the glyphs at the pixel level , Also marked with a font bounding box 、 Character category .

chart 2 TextLogo3K in Logo Image annotation

 

meanwhile , Their position and segmentation information in the original poster picture are also provided ：

chart 3 TextLogo3K Annotation of poster image in

The data set is free for users to do academic research （ No commercial use ）. Except for the words Logo Generate , This data set can also be applied to Text detection and recognition 、 Artistic font generation 、 Texture effect migration 、 Scene text editing Etc .

 

3、 ... and 、 Model design

The flow diagram of this model is shown in the figure below ：

chart 4 The flow chart of this model

This model is based on Conditional GAN To generate text Logo, Innovative use of the dual discriminator structure （ Sequence discriminator and image discriminator ）, On the trajectory sequence and the whole of the glyph Logo The images are judged separately ; At the same time, with the help of differentiable splicing (Differentiable Composition), Construction position coordinates to Logo The differential rendering process of an image . Its main processes include ：

• First, the bimodal characteristics of the input elements are used （ That is, the visual features of the font and the semantic features of the text ）, Encode it as a conditional feature .

• Coordinate generator A conditional feature and a random noise are used as input , Predict position coordinates for each character , That is, the coordinate of the center point of the font circumscribed frame , Width and height .

• The position coordinates of each character form a track sequence , Therefore, a Sequence discriminator To judge whether the sequence and are true or false according to the conditions . Note that the coordinate values in this task are continuous , It ensures that the sequence discriminator can propagate the gradient .

• adopt Differentiable splicing , Merge each glyph to get Logo Images .
• introduce Image discriminator , As Sequence discriminator A supplement to , The purpose is to further capture the details of the logo image , Ensure that there is no large overlap between different glyphs , The space between glyphs is reasonable .