Streaming Deep Learning Models for Computer Vision Tasks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning neural networks face significant computational expenses and network overhead when analyzing images in a distributed configuration, leading to increased latency and inefficiency in processing large images across multiple computers.
Innovation Solution
Implementing a patch-based streaming protocol that transfers lower-resolution versions of an image sequentially from one computer to another, allowing the receiving computer to start processing the image before the full resolution is completely downloaded, and continuing to transfer incrementally higher resolution versions until the image is identified, thereby reducing network traffic and latency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the full-resolution image is transferred over the network before processing, then the image processing can be performed with complete data, but the network overhead and latency increase significantly
Solution Approach 1:
The patent segments the image into multiple patches and transfers them sequentially over the network. The receiving computer processes each patch as it arrives, performing local analysis on smaller data units rather than waiting for the entire high-resolution image to be transferred. This segmentation approach reduces network latency while maintaining processing accuracy through progressive refinement.
Solution Approach 2:
The patent applies preliminary processing to generate lower-resolution versions of the image before transfer. These pre-processed patches are transferred first, allowing the receiving computer to begin processing immediately. The full-resolution patches are transferred subsequently to refine the results, eliminating the need to wait for complete image transfer before starting processing.
2Power
If multiple trained neural networks are implemented in a distributed configuration, then the computational resources are sufficient to handle complex image analysis, but the network overhead between computers increases
Solution Approach 1:
The patent divides the image processing task into multiple patches and distributes the computational workload across different neural networks in the distributed system. Each computer processes specific patches rather than receiving and processing complete high-resolution images, reducing the amount of data transmitted over the network and thereby reducing network overhead and energy consumption.
Solution Approach 2:
The patent extracts and transfers only the necessary image patches to the specific computing nodes that need them for processing. Instead of transferring complete images across the distributed network, only relevant patch data is moved, minimizing network traffic and associated energy costs while maintaining the computational capacity of the distributed system.
3Measurement precision
If the entire image is analyzed to identify patterns, then comprehensive object classification is achieved, but the computational expense becomes significant
Solution Approach 1:
The patent segments the image into multiple patches and processes them through the neural network in a distributed manner. Each patch is analyzed independently for patterns and features, and the results are aggregated to achieve comprehensive object classification. This segmentation reduces the computational expense of processing entire high-resolution images while maintaining classification accuracy through collective analysis of all patches.
Solution Approach 2:
The patent applies partial processing by analyzing image patches at different resolution levels. Lower-resolution patches are processed first to identify obvious patterns and features, and only regions requiring detailed analysis are processed at full resolution. This partial action approach reduces overall computational expense while maintaining comprehensive object classification accuracy.
Data Source
AI summary
A method and system for processing images by a neural network are provided. A first computer can sequentially transfer at least two versions of an image having increasingly greater resolution to a second computer. The second computer performs an image identification process on each of the sequentially transferred at least two versions until the image is identified.