Neural Network Feature Map Frequency Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks struggle to effectively distinguish between features of different frequencies, leading to unstable training processes due to discontinuous information transmission.
Innovation Solution
An electronic device and method that generate multiple feature maps representing features of different frequencies by performing convolution, upsampling, or downsampling to match sizes, and concatenating high-frequency and low-frequency feature maps for stable information transmission to the neural network model's output layer.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If convolutional neural network uses standard convolution operation, then the network structure is simple, but it cannot effectively distinguish between features of different frequencies
Solution Approach 1:
The patent segments the feature extraction process by separating high-frequency and low-frequency feature maps into distinct processing streams. The convolutional layer generates both high-frequency and low-frequency feature maps independently, which are then processed through different paths (upsampling for low-frequency, downsampling for high-frequency) before being concatenated. This segmentation enables frequency-specific processing that improves discrimination capability while maintaining manageable complexity through modular architecture.
Solution Approach 2:
The patent introduces a frequency dimension by applying different sampling operations to the same spatial feature maps. Low-frequency feature maps undergo upsampling to expand spatial resolution, while high-frequency feature maps undergo downsampling to reduce spatial resolution. This dimensional transformation in the frequency domain allows the network to capture and distinguish different frequency characteristics effectively.
2Reliability
If the neural network training process uses discontinuous information transmission, then the computational load is reduced, but the training process becomes unstable
Solution Approach 1:
The patent ensures continuous information transmission by maintaining consistent spatial dimensions across different frequency feature maps through upsampling and downsampling operations. Both high-frequency and low-frequency feature maps are transformed to match spatial dimensions before concatenation, ensuring that information flows continuously through the network without discontinuities. This continuity stabilizes the training process while the parallel processing of different frequency streams maintains computational efficiency.
Data Source
AI summary
An electronic device and a method for training or applying a neural network model are provided. The method includes the following steps. An input data is received. Convolution is performed on the input data to generate a high-frequency feature map and a low-frequency feature map. One of upsampling and downsampling is performed to match a first size of the high-frequency feature map and a second size of the low-frequency feature map. The high-frequency feature map and the low-frequency feature map are concatenated to generate a concatenated data. The concatenated data is inputted to an output layer of the neural network model.