Dual-Subnetwork Neural Network for Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current convolutional neural networks for object detection face challenges in effectively utilizing high-level features from upper layers for processing information, leading to insufficient representativeness and robustness, particularly in detecting small objects.
Innovation Solution
A deep convolutional neural network architecture with a first sub-network having down-sampling layers and a second sub-network with up-sampling layers, forming a symmetric hourglass structure, fuses feature maps to obtain fusion feature maps that represent both high-level semantic and low-level detail features, enhancing object detection accuracy and robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If pooling layers are used to reduce feature map size in convolutional neural network, then detection speed is improved, but representativeness of features deteriorates
Solution Approach 1:
The network is divided into two sub-networks: a first sub-network for extracting high-level semantic features and a second sub-network for preserving low-level detail features. This segmentation allows each sub-network to specialize in different feature types, maintaining representativeness while enabling efficient processing.
Solution Approach 2:
Feature maps from both sub-networks are fused together to create comprehensive fusion feature maps. The fusion process combines high-level semantic information from the first sub-network with low-level detail information from the second sub-network, achieving both speed and representativeness.
2Device complexity
If feature map size is reduced through pooling layers, then computational complexity is decreased, but detection accuracy for small objects deteriorates
Solution Approach 1:
The patent introduces a new dimensional aspect by creating parallel processing paths (two sub-networks) that operate simultaneously. The first sub-network processes for semantic understanding while the second preserves spatial details, and their fusion in a new dimension combines both advantages without compromising accuracy.
Solution Approach 2:
Different parts of the network are assigned different functions: the first sub-network focuses on global semantic features while the second sub-network maintains local detail features. This local quality differentiation ensures that small object detection requirements are met in specific regions while overall computational complexity remains manageable.
3Device complexity
If conventional zoom-out structure is used, then network processing is simplified, but robustness of detection deteriorates
Solution Approach 1:
The patent employs an asymmetric dual-subnetwork architecture where the first sub-network uses down-sampling operations for semantic feature extraction while the second sub-network uses up-sampling operations for detail feature preservation. This asymmetric design enhances robustness by capturing diverse feature types that a symmetric conventional structure would miss.
Solution Approach 2:
The fusion feature maps produced by the network serve multiple functions: they provide both high-level semantic understanding and low-level detail information, making the detection system universally applicable to various object sizes and types while maintaining robustness across different detection scenarios.
Data Source
AI summary
An object detection method, a neural network training method, an apparatus, and an electronic device include: obtaining, through prediction, multiple fused feature graphs from images to be processed, through a deep convolution neural network for target region frame detection, obtaining multiple first feature graphs from a first subnet having at least one lower sampling layer, obtaining multiple second feature graphs from a second subnet having at least one upper sampling layer, and obtaining fused graph by fusing multiple first feature graphs and multiple second feature graphs respectively; and obtaining target region frame data according to the multiple fused feature graphs. Because the fused feature graphs better represent semantic features on high levels and detail features on low levels in images, target region frame data of big and small objects in images can be effectively extracted according to the fused feature graphs, thereby improving accuracy and robustness of object detection.