Object Detection Feature Fusion for Small-Object Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional object detection using deep CNNs struggles with spatial information loss, particularly for small objects, leading to difficulty in accurate detection, and increasing feature map size to address this issue results in excessive computational demands.
Innovation Solution
An object detection device that generates mid-level and high-level feature maps through a combination of image data acquisition, first and second feature map extraction, and object detection, utilizing mid-level features for objectness and high-level features to enhance detection accuracy without excessive computational load.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the size of feature map is increased to detect small objects, then detection precision is improved, but calculation amount increases explosively
Solution Approach 1:
The patent segments the feature extraction process into two distinct pathways: a first feature map extraction path that preserves spatial information with smaller pooling rates, and a second feature map extraction path that reduces computation with larger pooling rates. This segmentation allows the system to process different feature representations separately, combining their advantages without requiring a single large feature map that would demand excessive computation.
Solution Approach 2:
The patent introduces a new dimension to the feature representation by generating a third feature map through element-wise multiplication of the first and second feature maps. This dimensional transformation creates a fused feature space that integrates both spatial details and computational efficiency, enabling small object detection without proportionally increasing computational load in the traditional feature map size dimension.
2Difficulty of detecting and measuring
If deep CNN is used for object detection, then detection capability is improved, but spatial information is lost due to pooling
Solution Approach 1:
The patent divides the feature extraction into parallel streams: one stream (first feature map) uses smaller pooling operations to preserve spatial information, while another stream (second feature map) uses larger pooling operations for computational efficiency. By segmenting the feature extraction process, the system maintains spatial details that would otherwise be lost in a single deep CNN pathway.
Solution Approach 2:
The patent merges the first and second feature maps through element-wise multiplication to create a third feature map that combines the advantages of both pathways. This merging operation integrates spatially-rich features with computationally-efficient features, allowing the model to detect small objects while maintaining reasonable computational complexity.
Data Source
AI summary
An object detection device includes an image data acquiring unit that acquires image data indicating an image captured by a camera, a first feature amount extracting unit that generates a first feature map using the image data, a second feature amount extracting unit that generates a second feature map using the image data, and generates a third feature map by performing addition or multiplication of the second feature map using the first feature map and weighting the second feature map, and the object detection unit that detects an object in the captured image using the third feature map. A first feature amount in the first feature map uses a mid-level feature corresponding to objectness, and a second feature amount in the second feature map uses a high-level feature.