Divided-Image Object Detection With Full-Image Metadata Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional object detection models struggle to accurately detect both relatively large and small objects in high-definition images using deep learning, as resizing such images can obscure characteristic features.
Innovation Solution
An image processing technique that divides the input image into multiple segments, performs object detection on each segment, and combines the results with detection on a resized version of the entire image to generate comprehensive metadata, ensuring both large and small objects are detected.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If a high-definition image is resized to fit model input constraints, then processing speed and model compatibility are improved, but detection accuracy of small objects deteriorates due to loss of characteristic features
Solution Approach 1:
The input high-definition image is divided into multiple divided images, each processed separately by the object detection model. This segmentation allows the model to maintain higher detection accuracy for small objects by examining smaller regions in detail, while still processing the overall scene efficiently through parallel processing of multiple segments.
Solution Approach 2:
The patent processes images at multiple scales by creating divided images from the original high-definition image. This multi-scale processing approach adds a spatial dimension to the detection process, allowing the system to detect objects at various sizes by examining the image at different granularities and combining the results.
2Measurement precision
If the input image is divided into multiple images for detection, then detection accuracy of small objects is improved, but detection accuracy of large objects deteriorates due to division of characteristic portions
Solution Approach 1:
The patent segments the original image into multiple divided images for processing. Each divided image is processed by the object detection model to identify objects within that segment. This segmentation strategy improves small object detection by providing focused examination of smaller regions while maintaining overall scene context through the collection of all segments.
Solution Approach 2:
The patent merges the detection results from multiple divided images to generate comprehensive metadata for the original image. By combining the attribute information and coordinate data from all divided image detections, the system reconstructs complete object information for large objects that may span multiple segments, ensuring reliable detection across the entire scene.
3Adaptability or versatility
If multiple processing paths (division and scaling) are used to detect both large and small objects, then comprehensive object detection is improved, but system complexity increases
Solution Approach 1:
The patent segments the detection process into two main paths: a division path that splits the image into multiple divided images for detailed small object detection, and a scaling path that processes a scaled-down version of the original image for efficient large object detection. This segmented approach enables comprehensive detection capability while organizing complexity into manageable, modular processing streams.
Solution Approach 2:
The patent employs a single object detection model that serves multiple functions by processing both divided images and scaled images. The same model architecture is used across different processing paths, allowing it to detect both large and small objects effectively. This multi-functional use of the model reduces system complexity compared to having separate specialized models for different object sizes.
Data Source
AI summary
An image processing device includes a division circuit that divides an input image and outputs a plurality of divided images, a first processor that performs computation of an object detection model and acquires attribute information including an attribute value of an object included in each of the divided images and a first quadrangular frame surrounding the object as first metadata, a scaling circuit that output an overall image obtained by shrinking the input image, a second processor that performs computation of the object detection model and acquires attribute information including an attribute value of an object included in the overall image and a second quadrangular frame as second metadata, and a third processor that generates third metadata of the input image by combining pieces of the attribute information of the second metadata and pieces of attribute information that are not held in common by the first and second metadata.