Object Detection via Pixel Correlation and Temporal Feature Updates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing salient object detection algorithms in images and videos often neglect time continuity between frames, leading to low computational efficiency and accuracy due to high computational complexity, especially when using recurrent neural networks.
Innovation Solution
A method that determines the correlation between pixels in an image and across frames of a video, updating image features based on this correlation to improve object detection accuracy and speed by integrating semantic and correlation information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If recurrent neural networks are used to consider time continuity of video frames, then object detection accuracy is improved, but computational complexity increases significantly
Solution Approach 1:
The patent segments the video processing task into two distinct parts: (1) extracting image features from individual frames using a convolutional neural network, and (2) determining temporal correlation between frames separately. This segmentation allows each component to be optimized independently, reducing overall computational complexity while maintaining accuracy.
Solution Approach 2:
The patent extracts only the essential temporal correlation information from video frames rather than processing entire frames through complex recurrent neural networks. By extracting key feature correlations between frames, the method reduces computational burden while preserving the time continuity information needed for accurate detection.
2Productivity
If existing salient object detection algorithms process each frame independently, then computational speed is improved, but time continuity information is neglected reducing accuracy
Solution Approach 1:
The patent performs preliminary extraction of image features from each frame before combining them with temporal correlation information. This preliminary processing allows fast independent frame analysis while preparing data for subsequent correlation-based temporal integration, maintaining both speed and accuracy.
Solution Approach 2:
The patent introduces temporal correlation as an intermediary mechanism that connects independently processed frames. Rather than directly combining full frames through complex RNNs, the method uses extracted feature correlations as intermediaries to preserve time continuity information while maintaining computational efficiency.
3Measurement precision
If temporal correlation between frames is considered, then object detection accuracy is improved, but detection time increases
Solution Approach 1:
The patent segments temporal correlation computation to operate only on extracted feature representations rather than full video frames. This segmentation dramatically reduces the computational burden of temporal analysis, allowing accurate multi-frame correlation to be computed in feasible time.
Solution Approach 2:
The patent changes the parameter representation from raw pixel data to compressed feature vectors before computing temporal correlations. This parameter transformation reduces the dimensionality of temporal analysis, enabling accurate multi-frame processing without excessive detection time.
Data Source
AI summary
This application discloses a method and apparatus for object detection in an image, a vehicle, and a robot. The method for object detection in an image is performed by a computing device. The method includes determining an image feature of an image; determining a correlation of pixels in the image based on the image feature; updating the image feature of the image based on the correlation to obtain an updated image feature; and determining an object detection result in the image according to the updated image feature.