Adaptive Image Scanning for Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for object detection and localization in images are time-consuming and suffer from reduced detection performance and accuracy due to the high number of scanning windows required, especially when objects can have arbitrary sizes and locations, leading to inefficient processing and missed detections.
Innovation Solution
An adaptive image scanning strategy is employed, where the scanning window size and direction are adjusted based on the probability of object occurrence, using a cascade of homogeneous classification functions to detect objects with higher precision and accuracy, and shifting the window in all directions with a small step to refine detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If scanning windows of certain scales are shifted utilizing a larger step to decrease the number of inspected windows, then processing time is reduced, but detection performance and localization accuracy are negatively affected
Solution Approach 1:
The patent applies dynamics by making the scanning step size adaptive rather than fixed. The step size changes dynamically based on the detection results - initially using larger steps to cover the image quickly, then switching to smaller steps in regions where objects are detected. This resolves the contradiction by allowing fast processing in most areas while ensuring accurate localization in regions of interest.
Solution Approach 2:
The patent implements local quality by applying different scanning strategies to different regions of the image. Instead of using a uniform scanning approach throughout, the system uses larger steps in regions where no objects are expected and smaller steps in regions where objects have been detected. This localized adaptation resolves the contradiction between fast processing and accurate localization.
2Measurement precision
If a large number of scanning windows are inspected to improve detection performance, then localization accuracy is enhanced, but processing time increases significantly
Solution Approach 1:
The patent applies partial action by inspecting only the necessary number of scanning windows rather than all possible windows. By using adaptive scanning with larger initial steps, the system performs a partial inspection that is sufficient to detect objects while avoiding the excessive processing of all possible window positions, thus resolving the contradiction between detection performance and processing speed.
Solution Approach 2:
The patent segments the scanning process into two phases: an initial phase with larger steps for quick coverage, and a refinement phase with smaller steps in detected regions. This segmentation allows the system to achieve good detection performance without the excessive processing time that would result from uniformly inspecting all possible windows.
3Productivity
If scanning windows are shifted with large steps to speed up processing, then productivity is improved, but detection accuracy deteriorates due to skipped regions
Solution Approach 1:
The patent uses dynamics by making the step size adaptive based on detection results. The system starts with large steps for fast processing, then dynamically reduces the step size in regions where objects are detected, ensuring that no potential objects are missed while maintaining overall processing efficiency.
Solution Approach 2:
The patent implements feedback by using the detection results from initial scanning to adjust the scanning strategy for subsequent processing. The feedback loop identifies regions where objects have been detected and triggers a refinement phase with smaller steps in those specific regions, ensuring detection accuracy is maintained while preserving processing speed in other areas.
Data Source
AI summary
An object detection method and system for detecting an object in an image utilizing an adaptive image scanning strategy is disclosed herein. An initial rough shift can be determined based on the size of a scanning window and the image can be scanned continuously for several detections of similar sizes using the rough shift. The scanning window can be classified with respect to a cascade of homogenous classification functions covering one or more features of the object. The size and scanning direction of the scanning window can be adaptively changed depending on the probability of the object occurrence in accordance with scan acceleration. The object can be detected by an object detector and can be localized with higher precision and accuracy.