Endoscope Object Detection Using Grid-Based Multi-Bounding Boxes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection methods using machine learning, such as YOLO, U-Net, and BodyPix, struggle to simultaneously maintain real-time processing and accurately represent the contour shape of objects, particularly in complex or fluctuating positions within images.
Innovation Solution
An information processing system that divides images into grid cells and generates multiple bounding boxes around overlapping areas to represent object positions and shapes, using a trained model to ensure real-time processing and detailed object representation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single bounding box is used to detect an object, then the processing speed is maintained, but the precision of object contour representation deteriorates
Solution Approach 1:
The detection target image is divided into multiple grid cells, and multiple bounding boxes are generated for objects that span across grid boundaries. This segmentation approach allows precise representation of object contours by using multiple smaller bounding boxes instead of one large box, thereby improving measurement precision without requiring complex additional hardware or systems.
2Measurement precision
If multiple bounding boxes are generated for each grid cell containing an object, then the object contour representation is improved, but the processing time increases
Solution Approach 1:
By dividing the image into grid cells and generating bounding boxes only for objects that overlap with grid boundaries, the method efficiently processes only relevant regions. This segmentation strategy maintains high detection precision while avoiding unnecessary computation in empty or non-critical areas, thus reducing overall processing time.
Solution Approach 2:
Different processing approaches are applied to different regions: grid cells containing objects of interest receive detailed multi-bounding box processing, while other regions use simpler detection methods. This local quality differentiation ensures high precision where needed while maintaining fast processing speed in less critical areas.
3Measurement precision
If traditional object detection methods are used, then real-time processing is maintained, but the ability to represent complex object shapes deteriorates
Solution Approach 1:
The image is segmented into grid cells, allowing complex objects to be represented by multiple bounding boxes across different grid cells. This segmentation enables accurate representation of complex shapes while maintaining real-time processing by using efficient grid-based organization and only processing relevant grid cells.
Solution Approach 2:
The detection approach transitions from a single global bounding box to multiple local bounding boxes distributed across the spatial dimension of grid cells. This dimensional transformation allows complex 2D object shapes to be accurately represented through the coordinated arrangement of multiple 1D bounding box segments.
Data Source
AI summary
An information processing system includes a processor that performs an object detection to detect an object from a detection target image. The processor divides the detection target image into a group of first grid cells. The object is positioned to overlap a group of second grid cells included in the group of the first grid cells. At this time, the processor generates a bounding box in a respective second grid cell included in the group of the second grid cells. The processor surrounds a portion of the object positioned in the respective second grid cell with the bounding box generated in the respective second grid cell and displays, on a display, a position and shape of the object by a collection of a plurality of bounding boxes superimposedly on the detection target image.