Object Detection Region Clipping for Small-Target Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection methods struggle with low detection accuracy for small objects in images, leading to false negatives due to low detection scores and increased computational load when reducing image size to reduce processing, which further reduces the size of the detection target and increases non-detections.
Innovation Solution
An information processing apparatus that includes a first detection unit to estimate detection regions and scores, a candidate generation unit to generate candidates, a region calculation unit to clip images based on low-scoring candidates, and a second candidate generation unit to enhance detection by processing clipped images, improving detection accuracy for small objects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the input image size is reduced to decrease calculation amount, then processing efficiency is improved, but the image size of the detection target becomes smaller and detection accuracy deteriorates
Solution Approach 1:
The patent divides the detection process into two stages: first performing object detection on the reduced-size input image to identify candidate regions, then performing detailed analysis on clipped images of those regions. This segmentation allows efficient initial filtering followed by precise local analysis, resolving the contradiction between processing efficiency and detection accuracy.
Solution Approach 2:
The patent introduces a hierarchical dimension to the detection process by operating at two different image resolution levels. The first detection operates on the downsampled image for efficiency, while the second detection on clipped images provides accuracy, effectively adding a processing-stage dimension to resolve the resolution-accuracy tradeoff.
2Speed
If the image size of the detection target is small in the image, then processing speed is improved, but the detection score becomes lower and false negatives increase
Solution Approach 1:
The patent performs a preliminary detection on the reduced image to identify candidate regions before performing detailed analysis. This preliminary action allows the system to quickly filter potential targets and then apply more rigorous analysis only where needed, improving both speed and reliability.
Solution Approach 2:
The clipped images serve as an intermediary between the reduced input image and the final detection result. By creating intermediate high-resolution views of candidate regions, the system can make more reliable detection decisions without permanently increasing the size of the entire input image.
3Measurement precision
If object detection processing is performed on the original image without size reduction, then detection accuracy is improved, but calculation amount increases
Solution Approach 1:
The patent applies partial action by performing full-resolution detection only on selected candidate regions rather than the entire image. The first detection on the reduced image uses less computation, and the second detection on clipped images applies higher computational effort only where necessary, optimizing the balance between accuracy and computational load.
Solution Approach 2:
The patent applies different processing qualities to different parts of the image: reduced resolution for the overall image during initial detection, and full resolution only for clipped candidate regions during secondary detection. This local quality differentiation reduces overall computational load while maintaining accuracy where it matters most.
Data Source
AI summary
Feature(s) improve detection accuracy in object detection that detects a subject set as a detection target from an image. An information processing apparatus may estimate a detection region and a detection score regarding a subject set as a detection target with respect to an input image, generate a first detection candidate regarding detection of the subject set as the detection target with respect to a first image, calculate a region to be clipped based on the first detection candidate in a case where the first detection candidate does not include a detection candidate for which the detection score is equal to or higher than a first threshold value, generate a second image by clipping from the first image based on the region to be clipped, and generate a second detection candidate regarding the detection of the subject set as the detection target with respect to the second image.