Multi-Scale Image Feature Fusion for Occluded Vehicle Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning-based vehicle recognition algorithms face challenges in complex real-world traffic scenarios due to variations in vehicle shapes and scales, reduced visibility under adverse weather conditions, and occlusions, leading to low accuracy and false alarms.
Innovation Solution
An image recognition method utilizing a pre-constructed recognition model with spatial, channel, and cross-channel attention mechanisms, combined with deformable convolution, to extract and fuse features from images, enhancing detection capability for multi-scale features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing deep learning-based recognition algorithms are used, then the system can process traffic scenarios, but the recognition accuracy decreases due to vehicle shape and scale variations
Solution Approach 1:
The patent divides the feature extraction process into multiple parallel branches, each specialized for detecting vehicles at different scales (small, medium, large targets). This segmentation allows the system to maintain high recognition accuracy across varying vehicle sizes by assigning dedicated processing paths to different scale ranges.
Solution Approach 2:
The patent introduces multi-scale receptive fields that operate across different spatial dimensions simultaneously. By constructing feature pyramids and using dilated convolutions with varying rates, the system captures vehicle features at multiple scales without sacrificing position precision, thus resolving the contradiction between scale adaptability and detection accuracy.
2Measurement precision
If existing algorithms are used under adverse weather conditions, then processing can continue, but visibility reduction makes vehicle location and distinction difficult
Solution Approach 1:
The patent applies local quality enhancement by using attention mechanisms that dynamically adjust feature weights based on local image characteristics. In adverse weather conditions, the model automatically emphasizes regions with higher vehicle presence probability and suppresses noisy backgrounds, maintaining location precision despite reduced overall visibility.
Solution Approach 2:
The system performs preliminary feature enhancement and denoising operations before main detection. By pre-processing features to amplify vehicle-related signals and suppress weather-induced noise, the model prepares optimized input for subsequent detection stages, improving location accuracy under poor visibility conditions.
3Reliability
If existing algorithms are used with dense traffic flow and occlusion, then detection can proceed, but key features are obscured leading to missed detections and false alarms
Solution Approach 1:
The patent implements multi-functionality through a unified detection framework that simultaneously performs vehicle detection, occlusion detection, and context reasoning. The model can identify vehicles even when partially occluded by using features from visible portions and inferring occluded regions based on contextual information, thereby maintaining detection reliability under occlusion.
Solution Approach 2:
The system employs feedback mechanisms where detection results from different scales and channels are aggregated and refined iteratively. Features from multiple sources are fused with confidence weighting, allowing the model to correct individual detection errors and reduce false alarms through cross-validation of multiple detection hypotheses.
Data Source
AI summary
Disclosed is an image recognition method, an image recognition device, an electronic device, and a storage medium. In the image recognition method, an image is acquired and inputted into a pre-built recognition model to determine an initial feature corresponding to the image via the recognition model. Then, the initial feature is divided into multiple groups according to a predetermined channel dimension. Further, feature extractions are performed on the multiple groups of the initial feature to obtain intermediate features corresponding to the multiple groups. Then the intermediate features are fused to obtain a final feature. Finally, a recognition result is output corresponding to the image via the recognition model based on the final feature.