Dynamic Resolution Reduction in Deep Convolutional Networks

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Deep convolutional networks face high computational complexity and challenging training processes, particularly in reducing image resolution across layers without compromising classification performance, especially in resource-constrained devices like mobile devices.

Innovation Solution

The method involves dynamically selecting and applying reduction factors to image resolutions at various layers of deep convolutional networks, adjusting the network architecture and weights to maintain performance while reducing computational complexity, using a system-on-a-chip architecture with multiple processing units to optimize image processing.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If image resolution is reduced in deep convolutional networks, then computational complexity is lowered, but classification performance deteriorates

Engineering Contradiction:
Improvecomputational complexityVSAvoidclassification performance
Core Design Contradiction:
Device complexityVSReliability

Solution Approach 1:

The patent implements dynamic resolution adjustment where the image resolution is not fixed but adaptively changed at different stages of processing. The system dynamically selects reduction factors and adjusts resolution based on the specific layer and processing stage, allowing the network to maintain high resolution when needed for accuracy while reducing resolution when computational load needs to be reduced.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent applies different resolution reduction strategies to different parts of the network architecture. Specifically, it uses layer-specific reduction factors where early layers may use different reduction factors compared to later layers. This localized approach ensures that each part of the network operates at the optimal resolution for its specific function, maintaining classification performance while reducing overall computational complexity.

Inventive Principle:
Principle #3Local quality

2Speed

If reduction factors are applied to reduce image resolution, then processing speed increases, but training difficulty increases

Engineering Contradiction:
Improveprocessing speedVSAvoidtraining difficulty
Core Design Contradiction:
SpeedVSEase of manufacture

Solution Approach 1:

The patent performs resolution reduction and network adjustment as preliminary actions before the main training and processing stages. By pre-selecting reduction factors and pre-adjusting the network architecture, the system avoids the need to retrain from scratch when changing resolution parameters, significantly reducing training difficulty while maintaining processing speed benefits.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent systematically varies resolution parameters (reduction factors) as controllable variables and observes their impact on both processing speed and training performance. This parameter-based approach allows for methodical optimization where the best reduction factors are selected based on empirical performance data, making the training process more predictable and manageable.

Inventive Principle:
Principle #35Parameter changes

3Reliability

If deep convolutional networks are used for high accuracy, then classification performance improves, but resource consumption increases

Engineering Contradiction:
Improveclassification performanceVSAvoidresource consumption
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent applies partial resolution reduction rather than complete downscaling. By using reduction factors less than 1 (e.g., 0.5 or 0.75 instead of 0.25), the system maintains sufficient image detail for accurate classification while still achieving meaningful reductions in computational load and resource consumption. This partial action approach finds the optimal balance between performance and efficiency.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3289529B1Reducing image resolution in deep convolutional networks
Publication Date: 2024.04.24 QUALCOMM INC
  • EP3289529B1 patent drawingFigure 1
  • EP3289529B1 patent drawingFigure 2
  • EP3289529B1 patent drawingFigure 3A

AI summary

A method of reducing image resolution in a deep convolutional network (DCN) includes dynamically selecting a reduction factor to be applied to an input image. The reduction factor can be selected at each layer of the DCN. The method also includes adjusting the DCN based on the reduction factor selected for each layer.