Mask RNN Target Recognition for UAV Small Objects
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current target recognition methods for unmanned aerial vehicles (UAVs) using box selection are prone to errors, especially when selecting small targets, which can be mixed with background information, leading to inaccurate marking and recognition.
Innovation Solution
A target recognition method based on a MASK RCNN network model that constructs a network model using feature maps of different sizes from intermediate and end layers, trained with a dataset until a preset condition is met, allowing for automatic target recognition without manual selection, thereby reducing human error and improving accuracy for small targets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If box selection method is used for target marking, then operator can manually select targets, but marking errors occur and recognition accuracy deteriorates
Solution Approach 1:
The system uses automated algorithms (RCNN, YOLO, SSD) to perform target detection and marking without requiring manual operator intervention. The algorithm independently identifies and marks targets, eliminating human error in the marking process while maintaining ease of operation through automated workflow.
Solution Approach 2:
The patent replaces the manual mechanical box-drawing operation with automated computer vision algorithms. Instead of operators manually dragging fingers to create selection boxes, the system uses neural networks to automatically detect and mark targets, substituting human mechanical action with computational processing.
2Ease of operation
If box selection is used for small targets, then target can be selected, but background information mixing occurs and recognition accuracy decreases
Solution Approach 1:
The patent employs multiple detection algorithms (RCNN, YOLO, SSD) that segment the image into different feature maps at various processing stages. This multi-stage segmentation allows the system to analyze targets at different levels of detail, improving small target detection by processing features at multiple resolutions and combining results.
Solution Approach 2:
The system changes detection parameters dynamically by using multiple algorithms with different characteristics. RCNN provides high precision for small targets, YOLO offers fast detection, and SSD balances both. The system adjusts which algorithm and parameter settings to use based on target characteristics, improving small target recognition accuracy.
3Device complexity
If single-stage network is used, then network structure is simple, but feature extraction capability is limited and recognition accuracy is insufficient
Solution Approach 1:
The patent uses multi-stage network architectures (RCNN, FPN) that divide feature extraction into multiple sequential stages. Each stage processes features at different resolutions and extracts different levels of semantic information. This staged approach improves recognition accuracy by progressively refining feature representations while managing computational complexity.
Solution Approach 2:
The system adds dimensional complexity to the network by incorporating multiple processing stages with different feature map resolutions. Instead of a single-stage flat architecture, the network operates across multiple dimensions of feature space, combining low-level detailed features with high-level semantic features to improve recognition accuracy.
4Device complexity
If feature maps of same size are used, then network construction is simple, but small targets are omitted and recognition completeness decreases
Solution Approach 1:
The patent introduces multiple feature map sizes corresponding to different processing stages of the network. Each stage produces feature maps at different resolutions, allowing the system to detect targets at various scales. This multi-dimensional feature representation ensures small targets are not omitted while maintaining manageable network construction complexity.
Solution Approach 2:
The system applies different feature map resolutions to different spatial regions and target sizes. Larger feature maps are used for detecting small targets where detailed local information is critical, while smaller feature maps handle larger targets. This local adaptation of feature quality improves recognition completeness without uniformly increasing network complexity.
Data Source
AI summary
A target recognition method and device based on a MASK RCNN network model are disclosed. The method comprises: determining a multi-stage network as a basic network; selecting at least one intermediate layer capable of extracting a feature map from the basic network, and inputting respectively a feature map output by the intermediate layer and a feature map output by an end layer of the basic network to corresponding MASK RCNN recognition networks to construct a network model based on the MASK RCNN, wherein the feature map output by the intermediate layer and the feature map output by the end layer have different sizes; training the MASK RCNN recognition networks with a data set and stopping training until a preset training end condition is satisfied; and recognizing the target using the MASK RCNN recognition networks after trained. This solution is very suitable for small target recognition of a flying UAV.