Top-Down CNN for Small Object Detection Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing one-stage object detection schemes using bottom-up network structures struggle to accurately detect small objects due to insufficient feature extraction and contextual information utilization.
Innovation Solution
The apparatus employs a hidden state top-down (HSTD) structure and a recursive prediction module (RPM) within a convolutional neural network (CNN) framework, which extracts hierarchical feature information and contextual data, and uses an attention mask to correct detection results, enabling accurate detection of small objects by propagating contextual information from upper layers to lower layers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a bottom-up network structure is used for object detection, then the detection process is simplified, but the accuracy of small object detection deteriorates
Solution Approach 1:
The patent inverts the traditional bottom-up detection approach by implementing a top-down detection framework. Instead of detecting objects from fine-grained features upward, the system first detects large objects using contextual information from upper layers, then progressively refines detection downward through hierarchical layers, thereby improving small object detection accuracy while maintaining process simplicity
Solution Approach 2:
The patent introduces a hierarchical dimensional structure with multiple detection layers (upper, middle, lower layers) that operate at different scales. By adding this vertical dimensional aspect to the detection process, the system can simultaneously handle both large and small objects across different hierarchical levels, resolving the trade-off between complexity and accuracy
2Productivity
If contextual information is not utilized, then the detection process is faster, but the accuracy of small object detection deteriorates
Solution Approach 1:
The patent performs preliminary extraction of contextual information at the upper layers before actual object detection occurs. This pre-processing of contextual data allows the detection process to proceed efficiently while already having the necessary contextual cues available, thus maintaining both speed and accuracy
Solution Approach 2:
The patent introduces contextual information as an intermediary element that mediates between the image input and object detection. This contextual mediator provides additional discriminative features that help distinguish small objects from background, improving accuracy without significantly impacting detection speed
3Device complexity
If feature extraction is insufficient, then the detection process is simpler, but the accuracy of small object detection deteriorates
Solution Approach 1:
The patent segments the feature extraction process into multiple hierarchical layers, each responsible for extracting features at different scales. The upper layer extracts contextual features, while lower layers extract detailed object features. This segmentation allows comprehensive feature extraction without overwhelming complexity at any single stage
Data Source
AI summary
An apparatus for detecting an object includes a processor that extracts information for object detection from image data frames based on a hierarchical structure of a convolutional neural network (CNN) and transmits information for object detection extracted from an uppermost layer of the hierarchical structure to a lower layer to detect an object based on information received at each layer, and storage that stores the information for object detection and detected object information.