Perturbation-Based Sample Acquisition for Target Detection Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing target detection methods in computer vision face challenges in recognition accuracy due to limited and poorly selected positive samples, which often exclude backgrounds, restricting the number of extractable samples and impairing the classifier's ability to recognize targets with backgrounds.
Innovation Solution
A sample acquisition method that introduces perturbations to pre-marked sample boxes in images to create diverse positive samples, including backgrounds, and a target detection model generation method using these samples along with negative samples to train weak classifiers, enhancing recognition accuracy and robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If positive samples are selected by artificially marking identification boxes that exactly frame complete targets without background, then the sample selection process is simple and clear, but the number of extractable samples is limited and recognition accuracy deteriorates
Solution Approach 1:
The patent applies dynamics by transforming static identification boxes into dynamic sampling windows through perturbation operations. The sampling window dynamically adjusts its position and size by adding random perturbations to the original identification box coordinates, enabling multiple sample extractions from single annotated data while maintaining target inclusion.
Solution Approach 2:
The patent introduces dimensional transformation by extending the 2D identification box into a 3D sampling space through the addition of perturbation dimensions. The sampling process operates in an expanded parameter space including position offsets and size variations, allowing diverse sample generation from limited annotated data.
2Ease of operation
If positive samples exclude background to simplify sample selection, then the sampling process is easier, but the classifier's ability to recognize targets with backgrounds is impaired
Solution Approach 1:
The patent applies preliminary action by pre-defining the sampling window based on the identification box before actual sample extraction. The perturbation parameters are predetermined within reasonable ranges, allowing the system to automatically generate diverse samples with backgrounds without requiring complex manual annotation or post-processing adjustments.
3Quantity of substance
If multiple samples are extracted from single annotated data through perturbation, then the number of positive samples increases, but the complexity of the sampling process increases
Solution Approach 1:
The patent implements self-service by enabling the sampling process to automatically generate diverse positive samples from single annotated data through self-contained perturbation operations. The system autonomously performs coordinate transformation, sample extraction, and validation without requiring external intervention or complex processing pipelines, making the increased sample generation capacity manageable.
Data Source
AI summary
The present disclosure discloses a sample acquisition method, a target detection model generation method, a target detection method, a computing device, and a computer readable medium. The sample acquisition method includes: adding a perturbation to a pre-marked sample original box in an original image to obtain a sample selection box, wherein an image framed by the sample original box contains a target; and extracting an image framed by the sample selection box as a sample. The technical solutions of the present disclosure can effectively increase the number of the samples that can be acquired in the original image, and adding a background to the samples can effectively improve the recognition accuracy of the trained target detection model.