Transposed Convolution Upsampling for Pixel-Aligned Feature Maps

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current image processing methods, particularly in CNNs, face inefficiencies when enlarging feature maps due to the need for external interpolation, which slows processing speed and increases overhead, especially when performing transposed convolution operations.

Innovation Solution

The method involves obtaining a kernel deterministically defined based on an extension ratio to perform a transposed convolution operation, allowing for up-sampling of feature maps without external interpolation, using a neural network layer configured with kernel parameters such as size, stride, and padding, which are determined by the extension ratio, and optionally using an external information provider for interpolation if pixel alignment is not met.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Volume of moving object

If external interpolation is used to enlarge feature maps, then the feature map size is increased, but processing speed decreases and overhead increases

Engineering Contradiction:
Improvefeature map sizeVSAvoidprocessing speed
Core Design Contradiction:
Volume of moving objectVSProductivity

Solution Approach 1:

The patent merges the interpolation operation with the transposed convolution operation into a single unified operation. By integrating the up-sampling function directly into the convolution layer, the system eliminates the need for separate external interpolation steps, thereby maintaining processing speed while achieving feature map enlargement.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The transposed convolution layer is designed to perform its own interpolation function internally without requiring external assistance. The layer uses learnable parameters to automatically perform the up-sampling operation as part of its core functionality, making the system self-sufficient and eliminating dependency on external interpolation modules.

Inventive Principle:
Principle #25Self-service

2Volume of moving object

If external interpolation is used to enlarge feature maps, then the feature map size is increased, but processing overhead increases

Engineering Contradiction:
Improvefeature map sizeVSAvoidprocessing overhead
Core Design Contradiction:
Volume of moving objectVSDevice complexity

Solution Approach 1:

The patent combines multiple operations (convolution and interpolation) into a single integrated layer. This merging reduces the number of separate processing steps and eliminates the overhead associated with coordinating multiple independent operations, simplifying the overall system architecture.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The transposed convolution layer is designed to serve multiple functions simultaneously: it performs feature extraction through convolution while also performing up-sampling through its inherent interpolation capability. This multi-functionality eliminates the need for separate dedicated interpolation modules, reducing system complexity.

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

3Productivity

If transposed convolution operation is performed without external interpolation, then processing speed is improved, but pixel alignment accuracy may be compromised

Engineering Contradiction:
Improveprocessing speedVSAvoidpixel alignment accuracy
Core Design Contradiction:
ProductivityVSManufacturing precision

Solution Approach 1:

The patent adjusts the parameters of the transposed convolution operation (such as kernel size, stride, and padding) to optimize both processing speed and pixel alignment accuracy. By carefully selecting these parameters, the system achieves accurate pixel alignment while maintaining high processing speed without requiring external interpolation.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3965014A1Method and apparatus with image processing
Publication Date: 2022.03.09 SAMSUNG ELECTRONICS CO LTD
  • EP3965014A1 patent drawingFigure 1
  • EP3965014A1 patent drawingFigure 2
  • EP3965014A1 patent drawingFigure 3

AI summary

An image processing method and apparatus are provided. The image processing method includes obtaining a kernel that is deterministically defined based on an extension ratio of a first feature map, up-sampling the first feature map to a second feature map by performing a transposed convolution operation between the first feature map and the kernel, and outputting the second feature map.