FCN Image Processing for Small Object Detection Accuracy

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image processing techniques struggle with detecting small detection targets, such as white canes, due to low detection accuracy when the target occupies a small region of the image and is easily confused with other features.

Innovation Solution

An image processing system utilizing a Fully Convolutional Network (FCN) with convolutional layers that maintain spatial features, performing binarization and expansion processing to enhance detection accuracy, and employing a teacher image-based training method to differentiate between detection targets and background.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional image processing or CNN is used for detection, then the system is simple to implement, but detection accuracy for small targets deteriorates

Engineering Contradiction:
Improvedetection accuracyVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The detection process is segmented into three distinct modules: region proposal network for candidate region generation, classification network for target identification, and visualization network for reasoning process display. This segmentation allows each module to specialize in specific tasks, improving overall detection accuracy for small targets while maintaining manageable system complexity through modular architecture.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a new dimension of visualization by adding the visualization network that displays the reasoning process of detection. This goes beyond traditional binary classification by providing intermediate information about detection confidence and region proposals, effectively adding a dimensional layer to the output that enhances interpretability without compromising detection accuracy.

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

2Measurement precision

If CNN with fully connected layers is used, then processing speed is fast, but classification accuracy for small coexisting subjects deteriorates

Engineering Contradiction:
Improveclassification accuracyVSAvoidprocessing speed
Core Design Contradiction:
Measurement precisionVSSpeed

Solution Approach 1:

The patent extracts and removes the fully connected layers from the traditional CNN architecture, retaining only the convolutional layers that are better suited for spatial feature extraction. This extraction of problematic components allows the system to maintain fast processing speeds while achieving superior classification accuracy for small targets through the region proposal mechanism.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The region proposal network acts as an intermediary between image input and final classification. It generates candidate regions that are then fed to the classification network, serving as a mediating step that improves accuracy by focusing computational resources on relevant regions rather than processing the entire image through fully connected layers.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Reliability

If traditional detection methods are used, then computational resources are conserved, but detection reliability for inconspicuous targets deteriorates

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

Solution Approach 1:

The region proposal network performs preliminary action by generating and filtering candidate regions before the main classification process. This preliminary step identifies promising areas in the image, allowing the subsequent classification network to focus computational resources only on these regions rather than processing the entire image, thereby improving detection reliability while managing computational consumption.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3901900B1Image processing system, image processing method, and image processing program
Publication Date: 2024.05.01 HITACHI KOKUSAI ELECTRIC INC
  • EP3901900B1 patent drawingFigure 1
  • EP3901900B1 patent drawingFigure 2
  • EP3901900B1 patent drawingFigure 3

AI summary

The present invention provides a technique suitable for the detection of small objects such as white canes. In order to solve this problem, one representative example of this image processing system comprises a video acquisition unit configured to input an input image acquired as a video, an estimation unit including a Fully Convolutional Network (FCN) learning model trained with respect to a region of a predetermined detection target, and configured to generate a region estimation image for the detection target by processing the input image with the learning model, and an image processing unit configured to detect the region of the detection target based on the region estimation image.