Image Description Generation With Adversarial CNN-RNN Training

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional image description generation methods using encoder-decoder structures, such as CNN-RNN, suffer from low generation quality, resulting in sentences that are not smooth or conform to daily oral expression habits.

Innovation Solution

Implement adversarial training of an image description information generation network using a discriminative network to alternately train a CNN-RNN structure, optimizing evaluation indexes through reinforcement learning to improve generation quality.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If encoder-decoder structure (CNN-RNN) is used to generate image description information, then image content recognition is achieved, but generation quality is relatively low with sentences not smooth or conforming to daily oral expression habits

Engineering Contradiction:
Improveimage content recognition accuracyVSAvoidgeneration quality
Core Design Contradiction:
Measurement precisionVSManufacturing precision

Solution Approach 1:

The patent introduces a discriminative network that provides feedback to the generative network. The discriminative network evaluates the generated image description information and provides gradient feedback to guide the generative network's parameter updates, thereby improving generation quality while maintaining content recognition accuracy.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent introduces a discriminative network as an intermediary between the generative network and the evaluation criteria. This intermediary network translates the quality assessment into actionable gradient feedback, mediating the optimization process and enabling quality improvement without directly modifying the generative model's architecture.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Manufacturing precision

If adversarial training is implemented with discriminative network, then generation quality is improved, but device complexity increases

Engineering Contradiction:
Improvegeneration qualityVSAvoidnetwork structure complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent merges the generative and discriminative networks into a unified adversarial training framework. By combining these two networks with complementary functions (generation and evaluation), the system achieves quality improvement while sharing computational resources and training data processing pipelines, thereby managing complexity through functional integration.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The discriminative network serves multiple functions: it evaluates generation quality, provides gradient feedback for optimization, and implicitly learns the distribution of high-quality image descriptions. This multi-functionality reduces the need for separate evaluation and training components, managing overall system complexity while achieving quality improvement.

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

Data Source

PatentEP3889836B1Image description information generation method and device, and electronic device
Publication Date: 2025.10.01 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP3889836B1 patent drawingFigure 1~2
  • EP3889836B1 patent drawingFigure 3~4
  • EP3889836B1 patent drawingFigure 5~6

AI summary

An image description information generation method and device, and an electronic device. The method comprises: acquiring a target image to be processed (S202); inputting the target image to a target-image description information generation network, wherein the target-image description information generation network is acquired after adversarial training is performed by using multiple sample images, and is for generating image description information (S204), the adversarial training being performed alternately on an initialized image description information generation network and an initialized determination network, the determination network being used to determine an output result of an image description information generation network; and generating, according to an output result of the target-image description information generation network, target-image description information used to describe the target image (S206). The method solves the technical problem of poor quality of information generated by an image description information generation method provided in the prior art.