Looking for breakthrough ideas for innovation challenges? Try Patsnap Eureka!

A Multi-task Image Processing Method Based on Deformable Convolution

An image processing and multi-task technology, applied in the field of image processing, can solve the problems of inability to extract distinguishing task-specific features, limited spatial modeling capabilities, etc., to promote the optimization of difficult tasks, enhance spatial modeling capabilities, and improve performance. Effect

Active Publication Date: 2022-06-24
OCEAN UNIV OF CHINA
View PDF5 Cites 0 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

[0008] Aiming at the problems existing in the above-mentioned prior art, the present invention provides a multi-task image processing method based on deformable convolution, which solves the following two technical problems: (1) For the limited spatial modeling ability in the multi-task network structure, it is impossible to To extract more distinguishable task-specific features, the present invention introduces deformable convolution in the field of multi-task learning to construct a task-specific deformable module, each module includes two parts: a deformable convolution layer and a feature alignment layer

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • A Multi-task Image Processing Method Based on Deformable Convolution
  • A Multi-task Image Processing Method Based on Deformable Convolution
  • A Multi-task Image Processing Method Based on Deformable Convolution

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0042] This embodiment is designed from the aspect of the network structure design of the multi-task learning model, which significantly enhances the spatial modeling transformation capability of the multi-task learning model, and extracts and mines more discriminative specific task features.

[0043] In the multi-task image processing method based on deformable convolution in the present invention, when designing the network structure of the multi-task learning model, deformable convolution is introduced into multi-task learning. like figure 1 As shown, feature sharing network is used to extract different levels of feature sharing blocks; the deformable convolution layer and feature alignment layer are sequentially connected to build task-specific deformable modules; finally, task-specific deformable modules are used to share feature blocks with different levels of features Directly correspond to the connection. In this process, self-adaptive adjustment is realized according ...

Embodiment 2

[0068] During model training, the weight reset method is used for multi-task optimization: in the training process, the difficulty of sub-tasks is ranked in real time according to the relative loss reduction rate, and the weight of the loss function of the easiest sub-task is reset to zero, and the remaining sub-tasks are reset to zero. The task continues training; after two rounds of iterations, the weight parameters that have been set to zero are restored; the above steps are repeated to achieve a dynamic balance between subtasks during the training process, so that all subtasks are in an equally important position.

[0069] Specifically include the following steps:

[0070] Step a: Subtask weight initialization: According to the complexity of the task, set hyperparameters for the loss function of each subtask as the weight of the initial subtask, optimize the network, and obtain the final optimization goal.

[0071]Step b: Real-time determination of task difficulty: Calcula...

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

PUM

No PUM Login to View More

Abstract

The invention discloses a multi-task image processing method based on deformable convolution, which belongs to the technical field of image processing. When designing the network structure of a multi-task learning model, the deformable convolution is introduced into the multi-task learning, and the deformable The convolutional layer and the feature alignment layer are sequentially connected to build a task-specific deformable module; the task-specific deformable module is directly connected to the feature sharing block, adaptively adjusted according to the content characteristics of the sub-task, and distinguishing task-specific feature extraction and feature fusion are formed. The entire multi-task learning network framework; in the multi-task optimization design, the weight reset operation is used to realize the dynamic balance of sub-task optimization.

Description

technical field [0001] The invention belongs to the technical field of image processing, relates to a multi-task learning network model through deformable convolution in deep learning, and in particular relates to a multi-task image processing method based on deformable convolution. Background technique [0002] Most of the current network models are based on single-task design, that is, a network model is only for one specific task. However, in the real world, visual tasks are related to each other, and the single-task network model isolates real problems from each other, ignoring the rich related information between problems, hindering further improvement of performance. The multi-task network trains and learns multiple related tasks together, explores and mines the rich related information contained in multiple related tasks, and helps improve the generalization performance of all related tasks. These characteristics make multi-task learning gradually become one of the r...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

Application Information

Patent Timeline
no application Login to View More
Patent Type & Authority Patents(China)
IPC IPC(8): G06N3/04G06N3/08
CPCG06N3/082G06N3/045
Inventor 黄磊李杰魏志强
Owner OCEAN UNIV OF CHINA
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Patsnap Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Patsnap Eureka Blog
Learn More
PatSnap group products

Browse by: Latest US Patents, China's latest patents, Technical Efficacy Thesaurus, Application Domain, Technology Topic, Popular Technical Reports.

© 2024 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap|About US| Contact US: help@patsnap.com