Attention-Based Configurable CNN for Visual Question Answering

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current systems for automatically answering questions related to images lack accuracy and effectiveness, as they fail to adequately integrate visual and semantic information, often relying on global visual features and insufficient attention mechanisms.

Innovation Solution

The proposed Attention-Based Configurable Convolutional Neural Networks (ABC-CNN) framework extracts image feature maps and generates question-guided attention maps by projecting question embeddings into a visual space, allowing for spatial weighting of image features to focus on relevant regions, thereby improving answer generation through a unified integration of visual and semantic information.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If global visual features are used for image question answering, then the system can process images efficiently, but the accuracy of answering region-specific questions deteriorates

Engineering Contradiction:
Improveprocessing efficiencyVSAvoidanswer accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent segments the image into multiple spatial regions and processes each region independently using separate convolutional neural network branches. This allows the system to maintain global processing efficiency while capturing region-specific details needed for accurate answer generation.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies different processing qualities to different parts of the image by using attention mechanisms that adaptively weight regional features. Regions containing question-relevant information receive higher attention weights, improving answer accuracy without sacrificing overall processing efficiency.

Inventive Principle:
Principle #3Local quality

2Measurement precision

If attention mechanisms are added to focus on relevant regions, then the accuracy of visual question answering improves, but the system complexity increases

Engineering Contradiction:
Improveanswer accuracyVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent implements dynamic attention mechanisms that adaptively adjust attention weights based on the specific question being asked. The attention coefficients are dynamically computed through neural network layers that process both image features and question embeddings, allowing the system to focus computational resources on relevant regions without requiring a fixed complex architecture.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent introduces attention coefficient maps as intermediary elements that mediate between the image feature maps and the answer generation process. These attention maps serve as intermediate representations that highlight relevant regions, simplifying the overall system architecture by providing a clear intermediate step rather than requiring complex direct interactions between all components.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Reliability

If visual and semantic information are integrated in a unified framework, then the performance of visual question answering improves, but the difficulty of system design and implementation increases

Engineering Contradiction:
Improvesystem performanceVSAvoiddesign complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent merges visual information processing and semantic information processing into a unified convolutional neural network framework. Both image features and question embeddings are processed through shared and specialized convolutional layers that work together to generate answers, creating an integrated system that improves performance while maintaining manageable design complexity through modular architecture.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP3166049B1Systems and methods for attention-based configurable convolutional neural networks (abc-CNN) for visual question answering
Publication Date: 2020.09.02 BAIDU USA LLC
  • EP3166049B1 patent drawingFigure 1
  • EP3166049B1 patent drawingFigure 2
  • EP3166049B1 patent drawingFigure 3

AI summary

Described herein are systems and methods for generating and using attention-based deep learning architectures for visual question answering task (VQA) to automatically generate answers for image-related (still or video images) questions. To generate the correct answers, it is important for a model's attention to focus on the relevant regions of an image according to the question because different questions may ask about the attributes of different image regions. In embodiments, such question-guided attention is learned with a configurable convolutional neural network (ABC-CNN). Embodiments of the ABC-CNN models determine the attention maps by convolving image feature map with the configurable convolutional kernels determined by the questions semantics. In embodiments, the question-guided attention maps focus on the question-related regions and filters out noise in the unrelated regions.