Multi-rate Dilated Convolution for Complete Object Region Location

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional image semantic segmentation methods struggle to locate all object regions in an image, as they typically focus on the most discriminative parts, leading to incomplete segmentation.

Innovation Solution

A training method using a multi-rate dilated convolutional neural network model performs full-image classification annotation at different dilation rates to obtain global object location maps, which are then used as supervision information for an image semantic segmentation network, enabling the identification of all object regions through dilated convolution processing at various dispersion levels.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional convolution operation is used to locate object regions, then the network can be trained with full-image classification annotation, but only one or more most discriminative parts of the entire object can be located, making it difficult to locate all object regions

Engineering Contradiction:
Improveobject region location accuracyVSAvoidincomplete object region coverage
Core Design Contradiction:
Measurement precisionVSLoss of information

Solution Approach 1:

The patent applies segmentation by dividing the feature map processing into multiple parallel branches, each using dilated convolution with different dilation rates (e.g., rates of 1, 2, 4, 8). This segmentation allows each branch to capture object regions at different scales and dispersion levels, collectively covering the entire object rather than just discriminative parts.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent changes the parameter of dilation rate in the convolution operation to solve the problem. By using multiple dilation rates (1, 2, 4, 8), the network can capture features at different dispersion levels, enabling comprehensive location of all object regions while maintaining the ability to train with full-image classification annotation.

Inventive Principle:
Principle #35Parameter changes

2Loss of information

If dilated convolution at multiple dilation rates is used to locate all object regions, then complete object coverage is achieved, but the network model complexity increases

Engineering Contradiction:
Improveobject region coverage completenessVSAvoidnetwork model complexity
Core Design Contradiction:
Loss of informationVSDevice complexity

Solution Approach 1:

The patent merges multiple feature maps obtained from dilated convolution operations with different dilation rates through element-wise addition or concatenation. This combining approach integrates information from all scales into a unified feature representation, achieving complete object coverage while managing model complexity through efficient fusion operations.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The multi-rate dilated convolution network serves multiple functions: it performs full-image classification, locates all object regions at different scales, and generates comprehensive feature maps for semantic segmentation. This multi-functionality reduces the need for separate networks for different tasks, thereby managing overall system complexity.

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

3Device complexity

If conventional deep convolutional neural network is used for image semantic segmentation, then the network structure is simple, but the segmentation quality is insufficient because not all object regions can be located

Engineering Contradiction:
Improvenetwork structure simplicityVSAvoidsegmentation quality
Core Design Contradiction:
Device complexityVSManufacturing precision

Solution Approach 1:

The patent introduces dynamic dilation rates into the convolution operation, allowing the network to adaptively capture features at multiple scales. This dynamic approach enables the network to adjust its receptive field dynamically, improving segmentation quality by capturing both local and global object regions without requiring a completely complex network architecture.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP3779774B1Training method for image semantic segmentation model and server
Publication Date: 2024.05.08 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP3779774B1 patent drawingFigure 1~2
  • EP3779774B1 patent drawingFigure 3~4
  • EP3779774B1 patent drawingFigure 5~6-a

AI summary

Disclosed are a training method for an image semantic segmentation model and a server, for use in locating all object areas in a raw image, increasing the segmentation quality of an image semantic segmentation. Provided in the embodiments of the present application, the training method for the image semantic segmentation model comprises: acquiring a raw image for use in model training; using a multi-magnification diluted convolutional neural network model under different dilution magnifications for a full image classification annotation with respect to the raw image, thus producing global object location images of the raw image in different degrees of dispersion, any degree of dispersion being used for indicating the distribution on a target object of the object areas located in the dilution magnification corresponding to the degree of dispersion via the multi-magnification diluted convolutional neural network model; using the global object location images as supervision information for an image semantic segmentation network model, training the image semantic segmentation network model via the supervision information.