Image Segmentation via Text Embedding Convolution

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image segmentation techniques are limited in scope, often misidentifying or failing to identify objects outside a fixed set of annotated classes, leading to decreased segmentation quality and accuracy.

Innovation Solution

A system that uses a learned image model and a text embedding model to generate a common embedding space, allowing for the generation of object-segmented images by convolving the learned image representation with the text embedding, thereby avoiding limitations of fixed classes and improving segmentation accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If a fixed set of annotated classes is used for segmentation, then the segmentation process is simple and fast, but the segmentation scope is limited and objects outside the fixed classes cannot be identified

Engineering Contradiction:
Improvesegmentation speedVSAvoidsegmentation scope
Core Design Contradiction:
ProductivityVSAdaptability or versatility

Solution Approach 1:

The patent transforms the segmentation approach by changing the parameter space from discrete fixed classes to continuous embedding space. Text queries are converted to embeddings that can represent any object concept, allowing the segmentation model to operate in a continuous parameter space rather than being constrained to predefined discrete categories. This enables flexible adaptation to any object while maintaining computational efficiency.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent introduces text embeddings as an intermediary between the image input and segmentation output. Instead of directly mapping images to fixed classes, the system uses text embeddings of user queries as mediators that guide the segmentation process. This intermediary layer enables the system to adapt to any object described by the text query while maintaining a consistent processing framework.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Measurement precision

If image saliency models are used to extract important objects, then the most important object can be identified, but the segmentation scope remains limited to predefined classes

Engineering Contradiction:
Improveobject importance identificationVSAvoidobject class coverage
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent creates a universal segmentation system that can handle any object class through text queries. The same segmentation model can be applied to different objects by simply changing the text query, making the system multi-functional without requiring separate models for each object class. This universality allows the system to identify and segment any important object while maintaining consistent processing.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Solution Approach 2:

The patent makes the segmentation scope dynamic by allowing text queries to define the target object on demand. Instead of being static to predefined classes, the system can adapt its segmentation focus dynamically based on what the user specifies in the text query. This dynamic approach enables the system to identify different important objects in different contexts while maintaining the same underlying model.

Inventive Principle:
Principle #15Dynamics

3Adaptability or versatility

If text embeddings are used to expand segmentation scope, then any object can be identified, but the system complexity increases with trainable components and embedding spaces

Engineering Contradiction:
Improvesegmentation scopeVSAvoidmodel architecture
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent extracts the class definition functionality from the segmentation model itself and places it in the text query input. Instead of having the model learn all possible object classes, the system extracts the specific object of interest through text embedding and uses that to guide segmentation. This separation reduces the burden on the model architecture while maintaining versatile segmentation capability.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent uses text embeddings as a copy or representation of the target object concept. Instead of requiring the model to directly understand and categorize diverse objects, the system creates a textual copy of the target concept through embedding and uses this copy to guide the segmentation process. This approach simplifies the model architecture by replacing complex object recognition with embedding-based guidance.

Inventive Principle:
Principle #26Copying

Data Source

PatentUS12008698B2Image segmentation using text embedding
Publication Date: 2024.06.11 ADOBE INC
  • US12008698B2 patent drawing
  • US12008698B2 patent drawing
  • US12008698B2 patent drawing

AI summary

A non-transitory computer-readable medium includes program code that is stored thereon. The program code is executable by one or more processing devices for performing operations including generating, using a model, a learned image representation of a target image. The operations further include generating, using a text embedding model, a text embedding of a text query. The text embedding and the learned image representation of the target image are in a same embedding space. Additionally, the operations include convolving the learned image representation of the target image with the text embedding of the text query. Moreover, the operations include generating an object-segmented image based on the convolving of the learned image representation of the target image with the text embedding.