Switchable Neural Network Configurations for Consistent Video
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural networks with certain structures produce varying predicted values during training, leading to unnatural video quality when processing temporally consecutive images due to node connections where error back propagation is not performed.
Innovation Solution
An information processing apparatus that includes a determining unit to switch between different neural network configurations, a configuration changing unit to alter node connections and input-output temporal relationships, and a training unit to train the neural network, using a combination of U-Net structures with and without error propagation to stabilize image processing results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a neural network with fixed node connections is used for processing temporally consecutive images, then the network structure remains simple and training is straightforward, but variation in predicted values occurs leading to unnatural video quality
Solution Approach 1:
The patent applies dynamics by making the neural network configuration changeable during training. The determining unit dynamically selects between first and second configurations based on training progress, and the configuration changing unit adjusts node connections and input-output temporal relationships accordingly. This dynamic adaptation allows the network to reduce prediction variations and improve video quality consistency without requiring a permanently complex structure.
Solution Approach 2:
The patent implements parameter changes by modifying the neural network's structural parameters (node connections and input-output temporal relationships) between different configurations. The determining unit monitors training parameters and triggers configuration switches when specific conditions are met, allowing the network to transition between different structural states to optimize performance and reduce variation in predictions.
2Reliability
If dropout technique is used to suppress over-training, then generalization ability improves, but variation in predicted values persists due to node removal
Solution Approach 1:
The patent extends the dynamic approach by introducing multiple configurations that can be switched during training. Unlike static dropout which randomly removes nodes, this invention dynamically switches between different structural configurations (first and second configurations with different node connections and temporal relationships), providing more controlled variation to reduce prediction inconsistencies while maintaining training stability.
3Adaptability or versatility
If different image quality results occur in temporally consecutive image processing, then individual image processing flexibility is maintained, but unnatural video appearance results
Solution Approach 1:
The patent implements feedback through the determining unit that monitors training progress and prediction variations. Based on this feedback, the system determines when to switch between configurations and adjusts the network structure accordingly. This feedback mechanism ensures that the network adapts its configuration to maintain consistent video quality while preserving the flexibility to process individual images with appropriate temporal relationships.
Data Source
AI summary
An information processing apparatus includes: an inference unit to which a neural network is applied, the neural network being configured to use a set of temporally consecutive images and a set of temporally consecutive image features as inputs, and output a set of images and a set of image features for the inputs; a determining unit configured to determine which one of first and second configurations to which the neural network is to be configured, the first configuration having a first input-output temporal relationship and the second configuration having a second input-output temporal relationship different from the first input-output temporal relationship; a configuration changing unit configured to change a node connection and an input-output temporal relationship of the neural network to change the configuration of the neural network according to the result determined by the determining unit; and a training unit configured to train the neural network.