Object Detection Using Zone-of-Interest Neural Filtering

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current artificial intelligence object detection mechanisms face challenges in achieving a balance between accuracy and speed, particularly when processing diverse datasets, often requiring different decision-making approaches based on the domain, and are inefficient on edge devices.

Innovation Solution

A two-step neural network process is employed, where pixel groups are analyzed using a first neural network to identify potential zones of interest, followed by a second network to pinpoint specific objects, utilizing domain-specific training datasets to optimize processing on edge devices.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If complex neural networks are used to process diverse input data, then accuracy is improved, but processing speed and execution efficiency deteriorate

Engineering Contradiction:
Improvedetection accuracyVSAvoidprocessing speed
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent divides the image processing task into two distinct neural networks: a first neural network that processes the entire image to identify zones of interest, and a second neural network that processes only those identified zones to detect objects. This segmentation allows the system to maintain high detection accuracy through specialized processing while improving overall processing speed by avoiding redundant computation on entire images.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies partial action by having the first neural network perform a preliminary analysis to identify only the zones of interest, rather than processing the entire image with both networks. The second neural network then performs object detection only on these partial regions. This approach reduces the computational burden while maintaining detection accuracy in the critical areas.

Inventive Principle:
Principle #16Partial or excessive action

2Measurement precision

If large training datasets are used to train neural networks, then detection accuracy is improved, but computational load and power consumption increase

Engineering Contradiction:
Improvedetection accuracyVSAvoidpower consumption
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent segments the processing workload between two neural networks with different computational requirements. The first network uses domain-specific training datasets to identify zones of interest, while the second network processes only those zones. This segmentation reduces the overall computational load and power consumption compared to using a single large network processing entire images, while maintaining detection accuracy through specialized domain-specific training.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies local quality by training each neural network with domain-specific datasets tailored to its function. The first network is trained with datasets appropriate for zone identification, while the second network is trained with datasets optimized for object detection within zones. This localized training approach improves detection accuracy for specific domains while reducing the need for excessively large universal training datasets.

Inventive Principle:
Principle #3Local quality

3Reliability

If neural networks process entire images, then comprehensive object detection is achieved, but processing efficiency on edge devices deteriorates

Engineering Contradiction:
Improvedetection completenessVSAvoidprocessing efficiency
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent segments the image processing task into two stages: first, the entire image is processed to identify zones of interest; second, only those zones are processed by the detection network. This segmentation ensures comprehensive detection coverage while significantly improving processing efficiency on resource-constrained edge devices by avoiding unnecessary computation in non-relevant areas.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies preliminary action by using the first neural network to pre-identify zones of interest before the second network performs detailed object detection. This preliminary filtering step ensures that the detection network only processes relevant regions, maintaining detection completeness while dramatically improving processing efficiency on edge devices.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4062369B1Systems and methods for object detection and recognition
Publication Date: 2026.03.25 XAILIENT
  • EP4062369B1 patent drawingFigure 1
  • EP4062369B1 patent drawingFigure 2
  • EP4062369B1 patent drawingFigure 3

AI summary

Techniques for identifying pixel groups representing objects in an image include using images having multiple groups of pixels, grouped such that each pixel group represents a zone of interest and determining a pixel value for pixels within each pixel group based on a comparison of pixel values for each individual pixel within the group. A probability heat map is derived from the pixel group values using a first neural network using the pixel group values as input and produces the heat map having a set of graded values indicative of the probability that the respective pixel group includes an object of interest. A zone of interest is identified based on whether the groups of graded values meet a determined probability threshold objects of interest are identified within the at least one zone of interest by way of a second neural network.