Image Processing for Large and Small Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional object detection models based on deep learning struggle to accurately detect both relatively large and small objects in high-definition images.
Innovation Solution
An image processing technique that divides a high-definition image into multiple smaller images, performs object detection on each divided image and the overall image, and combines the metadata to generate comprehensive metadata for the input image, ensuring both large and small objects are detected.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a high-definition image is shrunk to meet input size restrictions, then the image can be processed by the object detection model, but characteristic portions of objects are shrunk and small objects become difficult to detect
Solution Approach 1:
The input image is divided into multiple divided images (e.g., four quadrants), and object detection is performed separately on each divided image after appropriate size adjustment. This segmentation allows small objects to be detected with higher precision in each subset while maintaining overall processing efficiency.
Solution Approach 2:
The patent introduces a dual-processing dimension by simultaneously performing object detection on both the overall shrunk image and multiple divided images. This multi-dimensional approach ensures that both large objects (detectable in the overall image) and small objects (detectable in divided images) are captured.
2Measurement precision
If an input image is divided into multiple images for detection, then small objects can be detected more accurately, but large objects that straddle divided images have their characteristic portions divided and become difficult to detect
Solution Approach 1:
The patent merges the detection results from multiple divided images with the detection results from the overall shrunk image. By combining these results and eliminating duplicates, the system recovers complete information about large objects that may have been split across divided images, while retaining the enhanced small object detection capability.
Solution Approach 2:
The system uses feedback by comparing detection results from divided images with results from the overall image. Duplicate detections are identified and eliminated, ensuring that large objects spanning multiple divisions are correctly reconstructed while maintaining detection accuracy.
3Measurement precision
If both division processing and overall processing are performed, then both large and small objects can be detected, but the processing time and computational load increase
Solution Approach 1:
The patent applies partial action by performing division processing only when the input image contains regions with small objects that would be lost in overall processing. For images where only large objects are present, only the faster overall processing is performed, thus avoiding unnecessary computational overhead while maintaining detection accuracy when needed.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An image processing device (1) includes: a division unit (111) that divides the input image Iin and outputs a plurality of divided images; a first processing unit (110) configured to, by providing each of the plurality of divided images as input to an object detection model, perform computation of the object detection model, and acquire a set of pieces of attribute information including an attribute value of an object included in each of the plurality of divided images and a quadrangular frame BB1 surrounding the object, as metadata MD1 of the divided image; a scaling unit (121) outputs an overall image obtained by shrinking the input image Iin; a second processing unit (120) that performs computation of the object detection model by providing the overall image as input to the object detection model, and acquires a set of pieces of attribute information including an attribute value of the object included in the overall image and a quadrangular frame BB2 as metadata MD2 of the overall image; and a third processing unit (130) that generates metadata MD of the input image Iin by combining a set of pieces of the attribute information of the metadata MD2 and a set of pieces of attribute information that are not held in common by the metadata MD2 and the metadata MD1.