Neural Network Object Detection Using Shared Convolutional Layers

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current object recognition systems face inefficiencies in processing images of varying sizes, leading to increased computational load and reduced performance, as they often require separate convolutional processing for each scale, which is time-consuming and resource-intensive.

Innovation Solution

The method involves reusing a set of convolutional layers for multiple scales within an artificial neural network, generating feature map patches of different sizes from a single input image, and feeding these patches to respective classifiers for object recognition, thereby performing convolutional processing only once.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If separate convolutional processing is performed for each scale, then object recognition accuracy is maintained, but computational load and processing time increase

Engineering Contradiction:
Improveobject recognition accuracyVSAvoidprocessing speed
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent applies universality by designing a single convolutional processing pipeline that serves multiple scales simultaneously. The convolutional layers process the input image once and generate feature maps that are then scaled to different resolutions using interpolation, allowing the same convolutional features to be used for detecting objects at various sizes without requiring separate processing for each scale.

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

Solution Approach 2:

The patent segments the object detection task into two distinct stages: (1) a shared convolutional processing stage that extracts features at a base scale, and (2) a scaling stage where feature maps are interpolated to different resolutions for multi-scale detection. This segmentation allows the computationally intensive convolutional operations to be performed only once, while the lighter interpolation operations handle the multi-scale requirements.

Inventive Principle:
Principle #1Segmentation

2Reliability

If separate convolutional processing is performed for each scale, then detection accuracy for varying object sizes is maintained, but resource consumption increases

Engineering Contradiction:
Improvedetection accuracyVSAvoidcomputational resource consumption
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The convolutional layers are designed to be universal and scale-independent, processing the input image once to generate feature maps that can be adaptively scaled to any resolution. This eliminates the need for multiple separate convolutional processing pipelines, significantly reducing computational resource consumption while maintaining the ability to detect objects at various sizes accurately.

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

Solution Approach 2:

The patent performs preliminary convolutional processing at a base scale before scaling operations. By extracting features once and then scaling the feature maps, the computationally expensive convolutional operations are performed in advance, reducing the overall computational burden compared to performing separate convolutional processing for each scale.

Inventive Principle:
Principle #10Preliminary action

3Adaptability or versatility

If multiple convolutional processing pipelines are used for different scales, then comprehensive object detection coverage is achieved, but device complexity increases

Engineering Contradiction:
Improvemulti-scale detection capabilityVSAvoidnetwork structure complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The network architecture uses a universal convolutional processing pipeline that can handle multiple scales through feature map scaling rather than requiring separate processing pipelines for each scale. This significantly simplifies the network structure while maintaining comprehensive multi-scale detection capability.

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

Solution Approach 2:

The patent merges the multi-scale processing requirements into a single unified pipeline by combining convolutional feature extraction with subsequent feature map scaling operations. Instead of having separate parallel pipelines for different scales, the approach merges these functions into one streamlined process that reduces structural complexity.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP3329424B1Object detection with neural network
Publication Date: 2023.07.12 NOKIA TECHNOLOGIES OY
  • EP3329424B1 patent drawingFigure 1
  • EP3329424B1 patent drawingFigure 2A
  • EP3329424B1 patent drawingFigure 2B

AI summary

There is provided an apparatus comprising at least one processing core, at least one memory including computer program code, the at least one memory and the computer program code being configured to, with the at least one processing core, cause the apparatus at least to provide an input data item to a first convolutional layer of an artificial neural network comprising a set of convolutional layers (610), process the input data item in the set of convolutional layers (620), define, in a feature map output from a last convolutional layer of the set of convolutional layers, a first feature map patch and a second feature map patch, and provide the first feature map patch to a first classifier and the second feature map patch to a second classifier (630).