Multi-Segmentation Image Detection for Complex Background Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge in high-speed rail safety inspection is accurately detecting parts and defects in complex tunnel backgrounds, leading to false detections and interference due to background confusion, making anomaly detection difficult.
Innovation Solution
An image detection method using a first segmentation model trained with labels from a second segmentation model, where fused images from simple and complex backgrounds are used to enhance detection accuracy, reducing manual labeling costs and improving segmentation and detection precision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Difficulty of detecting and measuring
If anomaly detection algorithm is applied to detect parts and defects in complex tunnel backgrounds, then detection capability is provided, but false detections increase and detection accuracy deteriorates due to background confusion
Solution Approach 1:
The patent applies segmentation by dividing the complex detection task into multiple dimensions: first segmentation model separates target objects from complex backgrounds, second segmentation model processes fused images with simple backgrounds, and third segmentation model refines detection results. This multi-stage segmentation approach reduces background interference and improves detection accuracy while maintaining detection capability.
Solution Approach 2:
The patent introduces fused images as an intermediary element between the complex original images and the final detection results. These fused images combine information from multiple sources and are processed by the second segmentation model to generate accurate labels, serving as a mediator that reduces background confusion before final detection.
2Measurement precision
If first segmentation model is trained with manually labeled data to improve detection accuracy, then detection precision improves, but labeling cost and time consumption increase significantly
Solution Approach 1:
The patent implements self-service by enabling the system to generate its own training labels automatically. The second segmentation model processes fused images to generate accurate labels without manual intervention, and these labels are used to train the first segmentation model. This self-labeling approach eliminates the need for expensive and time-consuming manual annotation while maintaining high detection precision.
Solution Approach 2:
The patent applies preliminary action by pre-processing images to create fused images with enhanced features before label generation. The second segmentation model performs preliminary segmentation on these fused images to generate accurate labels, which are then used to train the first segmentation model. This preliminary processing step improves label quality without requiring manual labeling.
3Measurement precision
If multiple segmentation models and detection dimensions are employed to reduce false detections, then detection accuracy improves, but system complexity increases
Solution Approach 1:
The patent segments the detection system into specialized components: first segmentation model for complex background processing, second segmentation model for fused image processing, and third segmentation model for result refinement. Each model focuses on a specific aspect of the detection task, which improves accuracy while managing complexity through functional decomposition.
Solution Approach 2:
The patent achieves multi-functionality by designing a unified framework where the same second segmentation model serves multiple purposes: processing fused images, generating accurate labels, and reducing background interference. This multi-functional approach reduces overall system complexity compared to having separate specialized models for each function.
Data Source
AI summary
An image detection method includes processing a to-be-processed image based on a first segmentation model to obtain a to-be-detected image corresponding to a to-be-detected object. The to-be-processed image is obtained when the to-be-detected object is located in a scene. The method further includes detecting the to-be-detected image to obtain a detection result corresponding to the to-be-detected object. The first segmentation model is trained using a first label that is determined based on a second segmentation model. The second segmentation model is obtained by training using at least a fused image carrying a second label. The fused image is generated by fusing a sub-image corresponding to a detection object and a sub-image corresponding to another scene. The sub-image corresponding to the detection object is determined based on the second label.