Continuous-Image Object Detection With Bounding-Box Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning networks for object detection struggle with inconsistent results across continuous images, leading to inaccuracies in object detection due to objects being obscured or distorted.
Innovation Solution
A post-processing method for deep learning networks that uses rectifying and tracking keys to correct object detection results across continuous images by comparing bounding boxes and generating key vectors to merge, add, or delete bounding boxes, improving precision without altering the network structure.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning networks are used for object detection, then detection capability is improved, but computation expenses increase
Solution Approach 1:
The patent divides the object detection task into two parts: the deep learning network performs initial detection to provide bounding boxes and classification, while a separate post-processing module refines the results by comparing continuous images. This segmentation allows the computationally intensive deep learning part to be optimized independently from the refinement process.
Solution Approach 2:
The deep learning network performs preliminary detection on each image individually, providing initial bounding box predictions. The post-processing then uses these preliminary results across continuous images to correct and refine the detections, avoiding the need for the network to re-process everything from scratch.
2Speed
If deep learning networks detect objects from single images, then detection speed is maintained, but detection consistency across continuous images deteriorates
Solution Approach 1:
The patent processes images continuously in the order they are captured, maintaining a queue of previously detected objects and their bounding boxes. The post-processing module continuously compares the current image with previous images to track object consistency, ensuring reliable detection across continuous image streams without sacrificing speed.
Solution Approach 2:
The post-processing module uses feedback from comparing continuous images to correct detection results. By analyzing whether objects detected in one image appear consistently in subsequent images, the system refines bounding box predictions and maintains detection reliability across the image sequence.
3Device complexity
If objects are detected from continuous images independently, then processing simplicity is maintained, but detection accuracy deteriorates due to obscuration and distortion
Solution Approach 1:
The patent merges the results of independent image processing with a post-processing refinement stage. The deep learning network processes each image independently to maintain simplicity, while the post-processing module combines information from continuous images to correct inaccuracies caused by obscuration or distortion, achieving both simplicity and accuracy.
Solution Approach 2:
The post-processing module acts as an intermediary between the independent image processing stage and the final detection output. It receives bounding boxes from individual image processing, compares them across continuous images, and mediates the final results to correct accuracy issues while preserving the simplicity of the original processing approach.
Data Source
AI summary
An object detection method and apparatus are disclosed. The object detection method may include acquiring identification information of an object detected from each of N (N being a preset integer of 3 or more) images continuous in order of capture from the output of a classification layer of an object detection deep learning network and correcting, based on the identification information of the object detected from each of at least two images that have been captured early in order of time, among the N images, the identification information of the object detected from an image that has been captured later in order of time than the at least two images that have been captured early in order of time, among the N images.