UAV Small Object Detection via Multi-Scale CNN Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing solutions for unmanned aerial vehicles (UAVs) lack the capability to efficiently detect and recognize small and ultra-small objects in images in real-time due to limited computing power, leading to inefficiencies and potential loss of information.
Innovation Solution
A system and method utilizing a machine learning algorithm, specifically a convolutional neural network, implemented on a computing module onboard the UAV for real-time detection and recognition of small objects in high-resolution images, enabling classification and immediate communication of results to a ground station.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Difficulty of detecting and measuring
If machine learning algorithms are implemented on board the UAV for real-time image processing, then object detection and recognition capability is improved, but computing power requirements and device complexity increase
Solution Approach 1:
The system segments the image processing task by implementing a multi-scale detection approach where the image is analyzed at different resolution levels. The convolutional neural network processes features at multiple scales, allowing small objects to be detected without requiring the entire high-resolution image to be processed at full detail, thus reducing computational complexity while maintaining detection capability.
Solution Approach 2:
The patent introduces an additional dimension to the detection problem by implementing multi-scale analysis. Instead of processing a single resolution image, the system analyzes the same image space at multiple scale levels, enabling small objects to be detected in higher resolution regions while maintaining overall scene understanding at lower resolutions, thereby reducing total computational load.
2Measurement precision
If high-resolution images are processed to detect small objects, then detection accuracy is improved, but processing time increases
Solution Approach 1:
The processing time is reduced by segmenting the image into multiple scale levels. The convolutional neural network processes different regions at appropriate resolution levels, so that only critical small objects require high-resolution processing while the rest of the scene is processed at lower resolutions, maintaining accuracy for small objects while reducing overall processing time.
Solution Approach 2:
The system applies partial high-resolution processing only where necessary for small object detection. By using multi-scale analysis, the system processes only the portions of the image that contain or may contain small objects at high resolution, while other regions are processed at lower resolutions, thus reducing total processing time while maintaining detection accuracy.
3Productivity
If simplified classification is used to reduce computing load, then processing speed is improved, but classification accuracy deteriorates
Solution Approach 1:
The classification task is segmented into multiple processing stages corresponding to different scale levels. The convolutional neural network performs feature extraction and classification at each scale, allowing complex classification to be distributed across multiple simpler stages rather than requiring a single complex classification pass, thus maintaining accuracy while improving processing speed.
Data Source
AI summary
Disclosed are system and method for detecting small-sized objects based on image analysis using an unmanned aerial vehicle (UAV). The method includes obtaining object search parameters, wherein the search parameters include at least one characteristic of an object of interest; generating, during a flight of the UAV, at least one image containing a high-resolution image; analyzing the generated image using a machine learning algorithm based on the obtained search parameters; recognizing the object of interest using a machine learning algorithm if at least one object fulfilling the search parameters is detected in the image during the analysis; and determining the location of the detected object, in response to recognizing the object as the object of interest.