Target Region Extraction for Small Object Recognition Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current target detection systems face challenges in achieving high detection accuracy for small objects in images, as the size of the target object and the image affect the performance of deep learning neural networks, leading to inefficiencies and loss of features.
Innovation Solution
A recognition system and method that acquires original images and positioning information of target objects, extracts target regions based on this information, and recognizes the objects by comparing feature information with the extracted region features, thereby improving recognition accuracy and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If ultra-high resolution image is input into the neural network to improve small object detection accuracy, then the recognition accuracy of small objects is improved, but the neural network runs slowly
Solution Approach 1:
The patent divides the image processing into multiple stages: first obtaining positioning information to identify target regions, then extracting only those specific regions for neural network processing. This segmentation approach avoids processing the entire ultra-high resolution image, thus maintaining recognition accuracy while improving running speed.
Solution Approach 2:
The patent performs preliminary processing by obtaining positioning information and extracting target regions before inputting images into the neural network. This preliminary action identifies and isolates small objects of interest, allowing the neural network to focus computational resources only on relevant regions rather than processing the complete high-resolution image.
2Loss of information
If down-sampling rate of the image is too low to preserve features, then the features of the target object are preserved, but it is difficult to ensure the running efficiency of the neural network
Solution Approach 1:
The patent extracts only the necessary target regions containing small objects from the original image based on positioning information. By taking out only the relevant portions rather than processing the entire image, the system preserves essential features while significantly reducing the data volume requiring neural network processing, thus improving running efficiency.
3Measurement precision
If the neural network depth increases to improve feature extraction, then the semantic information is enhanced, but the detailed information of small objects disappears completely
Solution Approach 1:
The patent segments the image processing pipeline into positioning-based target region extraction followed by neural network processing. This segmentation allows the neural network to receive preprocessed images containing only target regions, enabling deeper network architectures to extract semantic information without losing detailed information of small objects that would be present in the focused target regions.
Data Source
AI summary
A recognition system and a recognition method are provided. The recognition system includes: an image acquisition module configured to acquire an original image of a target object; an information acquisition module configured to acquire positioning information of the target object; and an image recognition module configured to extract a target region in the original image, and to recognize the target object based on the extracted target region.