Image Recognition Prioritizing Ambience Change Features
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image recognition systems for vehicles face challenges in achieving high accuracy for target recognition in time-series frame images, particularly due to low positional detection and identification accuracy, which can lead to inadequate control system operation.
Innovation Solution
An image recognition apparatus that prioritizes targets with 'ambience change features', such as positional and size changes, when photographed from a moving object, using a neural network trained through deep learning to extract and output target information accurately.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If image processing is performed on time-series frame images to extract targets, then target information can be output for control systems, but positional detection accuracy and identification accuracy remain low
Solution Approach 1:
The patent applies dynamics by making the target extraction process adaptive through deep learning. The neural network dynamically adjusts its recognition criteria based on learned patterns from training data, enabling it to distinguish true targets from false detections. This dynamic adaptation resolves the contradiction by improving measurement precision (target recognition accuracy) while maintaining reliability (control system operation) through learned decision boundaries.
Solution Approach 2:
The patent changes parameters by transforming the target extraction problem into a learning problem with adjustable parameters (weights and biases in the neural network). Through training, the system optimizes these parameters to maximize recognition accuracy while maintaining reliable operation. The parameter changes enable the system to adapt to different scenarios and improve both accuracy and reliability simultaneously.
2Measurement precision
If all detected targets are processed equally, then processing is simple, but targets with incorrect positional detection or identification cannot be distinguished from valid targets
Solution Approach 1:
The patent segments the target processing into distinct stages: initial detection, deep learning-based verification, and final extraction. This segmentation allows the system to apply different processing levels to different targets, improving identification accuracy for critical targets while keeping the overall system manageable. The segmentation resolves the contradiction by enabling precise differentiation without overwhelming complexity.
Solution Approach 2:
The patent introduces deep learning verification as an intermediary step between initial detection and final target extraction. This intermediary layer filters out false detections and validates true targets through learned patterns, improving identification accuracy without requiring complete redesign of the processing pipeline. The intermediary resolves the contradiction by adding necessary complexity only where needed.
3Measurement precision
If traditional image processing methods are used, then system complexity is low, but target recognition accuracy is insufficient for safe control system operation
Solution Approach 1:
The patent replaces traditional mechanical image processing methods with a deep learning-based neural network system. This substitution enables significantly improved positional detection accuracy and target identification capability, overcoming the limitations of conventional methods. The increased complexity is justified by the substantial improvement in measurement precision required for safe control system operation.
Solution Approach 2:
The patent applies preliminary action by training the neural network offline before deployment. The extensive training process prepares the system in advance with learned patterns and knowledge, enabling it to achieve high accuracy during actual operation without requiring complex real-time processing. This preliminary preparation resolves the contradiction by shifting complexity from runtime to training time.
Data Source
AI summary
An image recognition apparatus includes a controller. The controller is configured to perform positional detection and identification for the target in each of the frame images, and extract a first target having an ambience change feature with priority over a second target that does not have the ambience change feature. The ambience change feature is a feature about a positional change of the target that is exhibited when the ambience is photographed from a moving object. The positional change is a positional change of the target identified in common among the frame images.