Event-Based Region Detection for Efficient Frame Sequence Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network systems face challenges in efficiently classifying image content in a sequence of high-resolution image frames due to high computational resource requirements, which can be exacerbated by focusing on the entire image rather than specific regions of interest, and determining these regions often requires substantial processing.
Innovation Solution
Converting frame image data into frame event data by identifying changes between frames, processing these events to determine a region of interest, and using downscaled images for classification, thereby reducing the number of pixels processed and memory usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the entire high-resolution image frame is processed for classification, then classification accuracy is improved, but computational resource usage increases
Solution Approach 1:
The patent divides the image processing task into two segments: first, event detection identifies regions of interest by detecting changes between frames; second, classification is applied only to these identified regions using downscaled images. This segmentation allows maintaining classification accuracy for changing regions while reducing overall computational resource usage by excluding static regions from processing.
Solution Approach 2:
The patent applies different processing qualities to different regions of the image. Regions of interest (where changes occur) receive full processing attention with downscaled classification, while static regions are effectively skipped. This local quality approach ensures accurate classification where needed while minimizing resource usage in unchanged areas.
2Productivity
If downscaling is applied to reduce resource usage, then computational efficiency is improved, but information integrity deteriorates
Solution Approach 1:
The patent performs preliminary event detection on full-resolution images to identify regions of interest before downscaling. This preliminary action preserves the spatial information needed to locate important regions, and subsequent downscaling is applied only to these specific regions for classification, minimizing information loss while achieving computational efficiency.
3Use of energy by moving object
If focus is placed on a specific region of interest, then resource usage is reduced, but determining the region location requires substantial processing
Solution Approach 1:
The patent extracts only the essential information needed for region detection by converting full frames to event data that captures only changes between frames. This extraction process identifies regions of interest with minimal processing by focusing solely on pixel changes, thereby reducing both resource usage and the time required to determine region locations.
4Quantity of substance
If event data conversion is performed, then data processing volume is reduced, but conversion processing is required
Solution Approach 1:
The patent applies partial action by converting only the necessary portions of image data to event format - specifically, only the changes between frames are converted to events rather than processing entire frames. This partial conversion significantly reduces data processing volume while keeping the conversion processing itself relatively simple and focused.
Data Source
AI summary
Examples relate to image classification. A method includes converting frame image data into frame event data. The frame image data represents an appearance of image content in a sequence of image frames. The frame event data represents one or more events. A conversion process includes determining, for each event of the one or more events, one or more event parameters including positional coordinates corresponding to a location of a respective pixel in the sequence of image frames. The method can include combining the frame event data of multiple image frames in the sequence of image frames using a weight factor. Event parameters of the frame event data are processed to determine an event-based region of interest in the sequence of image frames. A classification associated with the image content is determined based on the event-based region of interest.