Neural Network Inference with Integrated DWConv Image Downsampling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Inference processing using deep neural networks on large images increases processing load and time due to the need for separate image size reduction processing, which places a high load on the processor and delays the output of results.
Innovation Solution
An information processing device that integrates depthwise convolution layers with a stride of two or more closest to the input layer in the neural network model, allowing simultaneous size reduction and inference processing without separate preprocessing, thereby reducing the overall processing load.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If image size reduction processing is performed separately before inference processing, then the neural network can process images at optimal size, but the processing load and time increase due to separate processing steps
Solution Approach 1:
The patent merges the image size reduction function into the neural network model itself by configuring convolutional layers with stride greater than 1. This eliminates the need for separate pre-processing steps, as the network automatically performs downsampling during its normal operation, thereby reducing total processing time while maintaining inference accuracy.
Solution Approach 2:
The neural network model is pre-configured with convolutional layers that have stride greater than 1, which inherently performs image size reduction as part of its structure. This preliminary design allows the model to automatically handle resizing without requiring additional processing steps before inference.
2Speed
If image size is reduced before inference processing, then processing speed improves, but separate size adjustment processing increases overall processing load
Solution Approach 1:
The patent combines multiple functions (feature extraction and image size reduction) into a single integrated neural network model. By configuring convolutional layers with stride > 1, the model simultaneously performs downsampling and feature extraction in one operation, eliminating redundant processing steps and reducing overall computational load.
Solution Approach 2:
The neural network model is designed to serve multiple purposes: it performs both feature extraction and image size reduction through its convolutional layers with stride greater than 1. This multi-functionality eliminates the need for separate dedicated resizing operations, thereby reducing total processing load while maintaining processing speed.
3Device complexity
If conventional neural network architecture is used without DWConv layers with stride > 1, then model structure remains simple, but separate image size adjustment increases processing complexity
Solution Approach 1:
The patent integrates image size reduction functionality directly into the existing neural network model by configuring convolutional layers with stride greater than 1. This approach maintains relative structural simplicity while eliminating the need for separate preprocessing pipelines, thereby improving processing efficiency without significantly increasing model complexity.
Data Source
AI summary
An information processing device includes: an inferencer that receives an image as an input and performs inference processing using a neural network model on the image received; and a postprocessor that processes result information indicating a result of the inference processing performed by the inferencer to output an output value. The inferencer includes one or more depthwise convolution (DWConv) layers each with a stride of two or more as one or more layers located closest to an input layer among a plurality of layers included in the neural network model.