Parallel Neural Network Paths for Image Processing Latency

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing computer vision systems face inefficiencies in processing image data to detect objects and their characteristics simultaneously, leading to increased latency and reduced processing speed, especially when dealing with real-time video streams.

Innovation Solution

A neural network architecture with parallel paths processes image data to detect objects and their characteristics in parallel, sharing initial layers and using scale-specific groups of layers for efficient detection and recognition, allowing for simultaneous object and characteristic identification, reducing latency and improving processing efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Speed

If sequential processing of objects and characteristics is used, then processing accuracy is maintained, but processing speed decreases and latency increases

Engineering Contradiction:
Improveprocessing speedVSAvoidlatency
Core Design Contradiction:
SpeedVSLoss of time

Solution Approach 1:

The neural network is segmented into multiple parallel paths (first path for object detection, second path for characteristic detection) that share initial layers. This segmentation allows simultaneous processing of objects and characteristics without sequential delays, resolving the contradiction between processing speed and latency.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The invention transitions from sequential (one-dimensional) processing to parallel (multi-dimensional) processing by creating separate processing paths for objects and characteristics. This dimensional change enables both tasks to occur simultaneously, improving speed while reducing latency.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Productivity

If parallel processing paths are used, then processing efficiency increases, but device complexity increases

Engineering Contradiction:
Improveprocessing efficiencyVSAvoidneural network complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The first and second processing paths share initial layers of the neural network, merging common processing functions into a single set of layers. This reduces the overall complexity of the parallel architecture while maintaining the productivity benefits of simultaneous object and characteristic detection.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The shared initial layers serve multiple functions by processing features for both object detection and characteristic detection simultaneously. This multi-functionality reduces redundancy and simplifies the overall device complexity while maintaining high processing efficiency.

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

Data Source

PatentEP3467712B1Methods and systems for processing image data
Publication Date: 2023.04.26 CAMBRIDGE ENTERPRISE LTD
  • EP3467712B1 patent drawingFigure 1
  • EP3467712B1 patent drawingFigure 2
  • EP3467712B1 patent drawingFigure 3A

AI summary

There is described a computer-implemented system for processing image data, the system comprising: a sensor operable to capture image data comprising an image of an environment of the sensor; and a processing circuit comprising: a processor; and a computer-readable storage medium comprising computer-readable instructions which, when executed, cause the processor to: receive the image data from the sensor; process the image data using a first path through a neural network to obtain first data, the first path being configured to indicate a presence in the environment of one or more objects of a predetermined object type; process the image data using a second path through a neural network to obtain second data, the second path being configured to indicate a presence in the environment of one or more object characteristics corresponding to the predetermined object type; and generate output data using the first and the second data, wherein the first path and the second path are arranged to be enable the first data and the second data to be obtained in parallel.