Neural Network Object Detection via Adaptive Region Refinement

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current object detection systems, such as R-CNN, face inefficiencies and insufficient detection accuracy due to the scanning window technique's inefficiency and the need for improved performance in localizing and classifying objects in images.

Innovation Solution

A method utilizing a neural network with a deep convolutional neural network (DCNN) for feature extraction, a deep recurrent neural network (RNN) for region processing, and a reinforcement learning algorithm to determine bounding boxes and object classes, where feature vectors are extracted and processed through ROI pooling modules to improve detection accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If the scanning window technique is used to convert object detection to object classification, then object detection can be performed, but the detection efficiency deteriorates because classification steps are performed for all potential image regions of various locations, scales, and aspect ratios

Engineering Contradiction:
Improveobject detection capabilityVSAvoiddetection efficiency
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent divides the image processing task into multiple stages: first generating candidate regions using a region proposal generator, then performing detailed classification only on these selected regions. This segmentation approach avoids exhaustive classification of all possible regions while maintaining detection reliability.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary region proposal generation before the actual object classification step. By pre-identifying candidate regions that are likely to contain objects, the system prepares the data in advance, reducing the computational burden of the subsequent classification stage and improving overall detection efficiency.

Inventive Principle:
Principle #10Preliminary action

2Ease of operation

If the region-based convolution neural network (R-CNN) is used to perform two-stage approach with proposal generator, then object detection can be performed, but the detection accuracy is insufficient for some cases

Engineering Contradiction:
Improveobject detection functionalityVSAvoiddetection accuracy
Core Design Contradiction:
Ease of operationVSMeasurement precision

Solution Approach 1:

The patent implements a feedback mechanism where the recurrent neural network iteratively refines region proposals and feature extractions. The system uses the output of each processing stage to improve subsequent stages, allowing continuous optimization of detection accuracy through multiple passes over the data.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent introduces dynamic region refinement where the size and position of candidate regions are adaptively adjusted during processing. The recurrent neural network dynamically modifies region boundaries based on learned features, allowing the system to adapt to different object scales and shapes, thereby improving detection accuracy.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP3488387B1Method for detecting object in image and objection detection system
Publication Date: 2023.02.08 MITSUBISHI ELECTRIC CORP
  • EP3488387B1 patent drawingFigure 1
  • EP3488387B1 patent drawingFigure 2
  • EP3488387B1 patent drawingFigure 3A~3B

AI summary

A method detects an object in an image. The method extracts a first feature vector from a first region of an image using a first subnetwork and determines a second region of the image by processing the first feature vector with a second subnetwork. The method also extracts a second feature vector from the second region of the image using the first subnetwork and detects the object using a third subnetwork on a basis of the first feature vector and the second feature vector to produce a bounding region surrounding the object and a class of the object. The first subnetwork, the second subnetwork, and the third subnetwork form a neural network. Also, a size of the first region differs from a size of the second region.