Object Detection Neural Network with Depthwise Separable Convolutions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection methods using convolutional neural networks (CNNs) face significant slowdowns when detecting objects of different sizes within a single feature map, leading to inefficiencies in computation and memory usage, particularly in low-resource hardware devices.
Innovation Solution
A method and apparatus for high-speed object detection using an artificial neural network that employs a base network with 3×3 convolutional layers and 2×2 maxpooling layers, combined with additional layers featuring depthwise separable convolutions and detection layers with 1×1 convolutional layers to improve detection speed and reduce memory requirements, utilizing preset bounding boxes to streamline the detection process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If region extraction technique with candidate region estimation is used, then object detection accuracy is improved, but computation speed deteriorates
Solution Approach 1:
The network is segmented into base network, additional layers, and detection layers, where each segment performs specific functions. The base network extracts features, additional layers process features at different scales, and detection layers perform object detection, allowing parallel processing and improving speed without sacrificing accuracy.
Solution Approach 2:
The patent introduces multi-scale feature detection by adding detection layers that operate at different spatial dimensions. Instead of estimating candidate regions in a single feature map, the system detects objects at multiple scales simultaneously across different layers, eliminating the need for slow candidate region estimation while maintaining detection accuracy.
2Adaptability or versatility
If multiple bounding boxes are estimated for different object sizes, then detection coverage is improved, but memory usage increases
Solution Approach 1:
The patent uses multiple detection layers operating at different spatial dimensions to detect objects of various sizes. Each detection layer processes features at a specific scale, enabling the system to detect small, medium, and large objects simultaneously without storing multiple bounding box estimates, thus reducing memory consumption while maintaining comprehensive detection coverage.
Solution Approach 2:
The patent extracts features at multiple scales through additional layers that process outputs from the base network at different levels. By extracting and processing features at each scale independently, the system achieves comprehensive detection coverage without needing to maintain all possible bounding box estimates in memory simultaneously.
3Measurement precision
If complex CNN architecture is used for accurate detection, then detection precision is improved, but computation cost increases
Solution Approach 1:
The complex CNN architecture is segmented into modular components: base network for feature extraction, additional layers for multi-scale processing, and detection layers for object identification. This segmentation allows the system to achieve high detection precision through specialized layers while reducing overall computation cost by avoiding redundant operations in each segment.
Solution Approach 2:
The patent achieves accurate detection without excessive computation by operating at multiple spatial dimensions through additional layers. Instead of using a single complex architecture, the system processes features at different scales in parallel, reducing the computational burden on each individual layer while maintaining high detection precision through multi-scale analysis.
Data Source
AI summary
A fast object detection method and a fast object detection apparatus using an artificial neural network. The fast object detection method includes obtaining an input image; inputting the obtained input image into an object detection neural network using a plurality of preset bounding boxes; and detecting an object included in the input image by acquiring output data of the object detection neural network.