Best Image Crop Selection for Neural Network Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for selecting images from object tracks for identification or re-identification in video surveillance are inefficient due to varying image quality, motion blur, angular changes, and unpredictable object movement, making it difficult to determine the best image for accurate identification or re-identification.
Innovation Solution
A method for creating a quality annotated training data set for a neural network that automatically selects a 'best' image from a series of images by comparing probe images to a gallery of images with varying qualities, generating match scores, and determining quality values, which enables the network to learn what constitutes a high-quality image for identification or re-identification purposes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If a fixed delay is assigned before selecting an image crop, then the selection process becomes simple, but the accuracy of identification deteriorates because it cannot adapt to unpredictable object movement
Solution Approach 1:
The patent implements a dynamic selection mechanism that adapts to varying object movement patterns. Instead of using a fixed delay, the system evaluates multiple candidate images at different time points and selects the one with the highest quality score, allowing the selection process to respond dynamically to unpredictable object motion while maintaining identification accuracy
Solution Approach 2:
The system changes the selection criterion from a fixed temporal parameter (delay time) to a variable quality parameter. By introducing image quality assessment metrics and selecting images based on their quality scores rather than a predetermined time delay, the system achieves both operational flexibility and identification precision
2Measurement precision
If manual creation of training data set is performed, then subjectivity can be reduced through human judgment, but time consumption and labor increase significantly
Solution Approach 1:
The system implements self-service by automatically generating quality annotations for training images through algorithmic assessment. The neural network and image quality evaluation metrics autonomously determine image quality without requiring manual human judgment, thereby eliminating time-consuming manual annotation processes while maintaining consistent and objective quality assessment
Solution Approach 2:
The patent replaces the mechanical process of manual image quality assessment with an automated computational system. By substituting human manual evaluation with algorithm-based quality metrics and neural network processing, the system achieves rapid automated annotation that is both time-efficient and free from human subjectivity
3Quantity of substance
If images with varying quality are used for re-identification, then more image data is available, but the reliability of identification deteriorates due to poor quality images such as those with motion blur or angular changes
Solution Approach 1:
The patent applies local quality assessment by evaluating specific quality attributes of individual images (such as sharpness, lighting, angle, and motion blur) rather than treating all images uniformly. This allows the system to identify and select high-quality images from a larger set, ensuring that only reliable images with adequate quality metrics are used for re-identification, thereby maintaining identification reliability while utilizing available image data
Data Source
AI summary
Methods and apparatus, including computer program products, for creating a quality annotated training data set of images for training a quality estimating neural network. A set of images depicting a same object is received. The images in the set of images have varying image quality. A probe image whose quality is to be estimated is selected from the set of images. A gallery of images is selected from the set of images. The gallery of images does not include the probe image. The probe image is compared to each image in the gallery and a match score is generated for each image comparison. Based on the match scores, a quality value is determined for the probe image. The probe image and its associated quality value are added to a quality annotated training data set for the neural network.