CNN Inference Optimization via Overlapping Sub-Image Caching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional Neural Networks (CNNs) face challenges in efficiently processing high-resolution images due to high computational costs and bandwidth requirements, especially when performing fine-grained analysis on limited computational devices like edge devices, which leads to increased resource consumption and inference time.
Innovation Solution
The method involves dividing high-resolution images into partially overlapping sub-images, caching inference results for overlapping areas, and reusing these results to reduce computational workload and inference time, thereby optimizing CNN inference for full resolution images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-resolution images are processed using CNN on edge devices, then fine-grained analysis capability is improved, but computational cost and inference time increase significantly
Solution Approach 1:
The patent divides a high-resolution image into multiple smaller patches, allowing the CNN to process each patch independently and in parallel. This segmentation reduces the computational burden on edge devices while maintaining fine-grained analysis capability, as each patch can be analyzed separately and results aggregated to form the complete image understanding.
Solution Approach 2:
The patent applies CNN inference only to selected patches rather than the entire high-resolution image, representing a partial action approach. By identifying and processing only relevant or representative patches, the system achieves fine-grained analysis of critical regions without the prohibitive computational cost of processing every pixel at full resolution.
2Measurement precision
If high-resolution images are processed using CNN on edge devices, then fine-grained analysis capability is improved, but resource consumption increases
Solution Approach 1:
By segmenting the high-resolution image into smaller patches, the patent reduces the total computational workload required for fine-grained analysis. Each patch requires fewer computational resources and less energy to process, and the parallel processing of patches further optimizes energy efficiency on edge devices compared to processing the entire image sequentially.
Solution Approach 2:
The patent implements partial processing by applying CNN inference only to selected patches rather than the complete high-resolution image. This partial action approach significantly reduces energy consumption while still achieving the desired fine-grained analysis capability in critical regions, making the system feasible for resource-constrained edge devices.
3Reliability
If overlapping sub-images are processed separately, then comprehensive object recognition is improved, but computational redundancy increases
Solution Approach 1:
The patent merges inference results from overlapping patches by aggregating predictions for objects that appear across multiple patches. This combining approach maintains comprehensive object recognition accuracy while reducing computational redundancy, as the system intelligently consolidates results from overlapping regions rather than treating each patch independently and duplicating effort.
Data Source
AI summary
A method identifies a type of object in a digital image. A user and/or one or more processors selects, from a plurality of partially overlapping sub-images of a digital image, a first sub-image and a second sub-image that overlap one another. The user/processors input the first sub-image into a neural network to create a first inference result that includes an overlapping inference result, for the overlapping area, that recognizes a partial portion of a specific type of object based on the overlapping area. The user/processors infer that the second sub-image creates a second inference result that recognizes a second portion of the specific type of object in the second sub-image based on the second sub-image and the overlapping inference result. The neural network identifies the specific type of object in the digital image based on the first and second sub-images being sub-images of a same type of object.