Dynamic Anchor Generation for Small Target Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Anchor-based deep learning target detection algorithms face challenges in capturing small-sized targets and intensive targets due to fixed anchor settings, leading to missed matches and errors in classification and regression.
Innovation Solution
A method and apparatus that dynamically determine target anchors by extracting feature maps, performing feature fusion using a feature pyramid network, and generating candidate anchors based on offset parameters, allowing for effective anchor allocation across different directions and sizes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If fixed anchor settings are used in anchor-based deep learning target detection algorithms, then the algorithm structure remains simple and computationally efficient, but the ability to capture small-sized targets and intensive targets deteriorates, leading to missed matches and classification errors
Solution Approach 1:
The patent transforms the static anchor setting into a dynamic one by introducing a feature pyramid network that generates multiple feature maps at different scales. Each feature map produces anchors of different sizes, allowing the system to adapt to targets of various scales. The anchor size is no longer fixed but dynamically selected based on the scale of the target being detected, resolving the contradiction between simple structure and accurate detection of small targets.
Solution Approach 2:
The patent changes the parameter of anchor size from a fixed value to a variable that can take multiple values corresponding to different feature map scales. By introducing multiple anchor sizes through the feature pyramid network, the system can adjust the anchor parameters to match the scale of detected targets, thereby improving detection accuracy without significantly complicating the overall algorithm structure.
2Productivity
If the feature map size is reduced through downsampling (e.g., 8 times downsampling), then the computational load is reduced and processing speed is improved, but the resolution deteriorates, causing small-sized targets to be missed
Solution Approach 1:
The patent segments the feature extraction process into multiple parallel pathways, each producing feature maps at different downsampling levels. Instead of a single heavily downsampled feature map, the system creates multiple feature maps with varying resolutions (e.g., 1/8, 1/16, 1/32 of original image size). This segmentation allows small targets to be detected in higher-resolution feature maps while large targets are detected in lower-resolution maps, maintaining both processing efficiency and detection precision.
Solution Approach 2:
The patent adds a scale dimension to the feature representation by constructing a feature pyramid with multiple levels. Each level of the pyramid represents a different scale dimension, allowing the system to detect targets across multiple size dimensions simultaneously. This dimensional expansion enables the system to maintain high processing speed through downsampling while preserving the ability to detect small targets through higher-resolution feature maps in the pyramid structure.
Data Source
AI summary
Embodiments of the present disclosure disclose a method and apparatus for determining a target anchor, a device and a storage medium. The method may include: extracting a plurality of feature maps of an original image using a feature extraction network; inputting the plurality of feature maps into a feature pyramid network to perform feature fusion, to obtain a plurality of fused feature maps; and using a region proposal network to implement operations as follows: determining an initial anchor of a network head using the fused feature map, based on a size of each fused feature map, and determining an offset parameter of the initial anchor, based on a ratio of the size of the fused feature map to the original image, and generating a plurality of candidate anchors in different directions, based on the offset parameter of the initial anchor.