Convolutional Neural Network Multi-Scale Feature Fusion for Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current object detection methods, particularly in fields like autonomous driving and medical image analysis, face challenges in achieving high accuracy for object classification and anchor box regression due to the need for precise feature extraction and handling of diverse object sizes, which can be critical for safety but are not adequately addressed by existing AI technologies.
Innovation Solution
A convolutional neural network (CNN) method that extracts multi-scale feature maps, combines spatial and semantic features through scaling and addition operations, and uses an ROI aligner to pool and fuse features for enhanced object classification and anchor box regression, improving performance by providing diversified features and accurate transformations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing AI technologies like R-CNN, Fast R-CNN, Faster R-CNN, and Mask R-CNN are used to enhance object detection performance, then object classification and anchor box regression accuracy are improved, but the precision and reliability of detection results remain insufficient for critical applications like autonomous driving and medical image analysis
Solution Approach 1:
The patent segments the feature extraction process into multiple scales (first feature maps at different resolutions) and combines them through a hierarchical structure. The region of interest (ROI) is divided into multiple grids, and features are extracted and fused at different hierarchical levels to achieve both precision and reliability in detection results.
Solution Approach 2:
The patent implements a nested hierarchical feature extraction architecture where feature maps at different scales are nested within each other. The ROI alignment mechanism nests multiple-scale features within a unified representation, allowing the system to maintain high precision while improving reliability through multi-level feature fusion.
2Measurement precision
If multi-scale feature extraction is implemented to handle diverse object sizes, then object detection accuracy is improved, but computational complexity and processing time increase
Solution Approach 1:
The patent segments the feature extraction into multiple scales but organizes them in a hierarchical structure that reduces overall complexity. The ROI alignment process segments the region of interest into grids, allowing efficient multi-scale processing without proportional increases in computational complexity.
Solution Approach 2:
The patent applies partial feature extraction at different scales rather than exhaustive processing. By selectively extracting and fusing features at hierarchical levels, the system achieves improved accuracy without the full computational burden of processing all possible feature combinations, thus managing complexity effectively.
3Ease of operation
If traditional ROI pooling schemes are used, then processing is simplified, but misalignment occurs that reduces detection precision
Solution Approach 1:
The patent performs preliminary ROI alignment before feature pooling by calculating accurate bounding boxes and adjusting coordinate systems. This preliminary alignment action ensures precise feature mapping while maintaining operational simplicity, as the alignment is performed automatically as part of the standard processing pipeline.
Solution Approach 2:
The patent replaces traditional mechanical ROI pooling approaches with a neural network-based feature fusion mechanism. The neural network automatically learns optimal feature combinations and alignments, eliminating misalignment issues while maintaining simplicity through end-to-end training rather than complex manual alignment procedures.
Data Source
AI summary
Disclosed are an object detection method and a convolution neural network. The method is performed through hierarchical architecture of the CNN and includes extracting groups of augmented feature maps from an input image through a backbone and two other groups of feature maps, identifying positive and negative samples with an IOU-based sampling scheme to be proposals for foreground and background through a proposal-sampling classifier, mapping the proposals to regions on the groups of augmented feature maps through the region proposal module, pooling the regions to fixed scale feature maps based on ROI aligning, fusing the fixed scale feature maps, and flattening the fused feature maps to generate an ROI feature vector through an ROI aligner for object classification and box regression. Because extracted features in the groups of augmented feature maps range from spatially-rich features to semantically-rich features, enhanced performance in object classification and box regression can be secured.