Neural Network Video Decomposition for Reflection Removal

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing video processing technologies fail to effectively remove unwanted features like reflections, shadows, and occlusions from video data, especially in complex scenarios with motion and varying lighting conditions, leading to suboptimal video quality and increased processing requirements.

Innovation Solution

A neural network system comprising 3D spatio-temporal encoder and decoder networks, along with a corrector network, decomposes video data into layered representations, using techniques like optical flow and permutation invariant loss functions to separate desired and undesired features, and applies successive corrections to enhance decomposition quality.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If traditional video processing methods are used to remove unwanted features, then processing simplicity is maintained, but video quality improvement is insufficient and processing requirements increase

Engineering Contradiction:
Improvevideo qualityVSAvoidprocessing requirements
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent segments video data into multiple layered representations (foreground layer, background layer, reflection layer, shadow layer) using neural network decomposition. This segmentation allows targeted processing of unwanted features like reflections and shadows in specific layers while preserving desired content in other layers, thereby improving video quality without requiring complex global processing of the entire video data.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces latent variable representations as intermediaries between the input video data and the final output. The neural networks encode video data into compressed latent representations, process these intermediaries to separate unwanted features, and then decode back to video data. This intermediary approach simplifies the processing by working in a lower-dimensional space rather than directly manipulating full-resolution video data.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Manufacturing precision

If complex neural network systems are used to decompose video data into layered representations, then video quality and feature separation are improved, but computational complexity and processing time increase

Engineering Contradiction:
Improvefeature separation qualityVSAvoidprocessing time
Core Design Contradiction:
Manufacturing precisionVSLoss of time

Solution Approach 1:

The patent applies preliminary action by pre-training the neural network models on large datasets to learn effective video decomposition patterns. Once trained, the models can rapidly decompose new video data without requiring complex real-time computations. The preliminary learning phase captures the essential relationships between different video layers, enabling fast inference later.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent utilizes parameter changes in the form of learned transformation parameters within the neural networks. The models learn optimal parameter transformations that map input video data to decomposed layered representations. By optimizing these parameters during training, the system achieves high-quality feature separation with efficient forward-pass computations during actual video processing.

Inventive Principle:
Principle #35Parameter changes

3Reliability

If video data is decomposed into multiple layered representations, then unwanted features like reflections and shadows are removed, but data processing complexity and memory requirements increase

Engineering Contradiction:
Improveunwanted feature removalVSAvoidmemory requirements
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The patent extracts unwanted features like reflections and shadows into separate dedicated layers (reflection layer, shadow layer) during the decomposition process. This extraction allows the system to isolate and remove these unwanted elements by zeroing out or discarding their respective layers, while preserving the cleaned video data by combining the remaining foreground and background layers. This approach removes unwanted features effectively without requiring complex post-processing memory operations.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent transforms the video processing problem from the spatial-temporal domain into an additional layer dimension. By decomposing video data into multiple layered representations (adding a layer dimension), the system can manipulate unwanted features by simply adjusting values in specific layers rather than applying complex spatial-temporal filtering operations, thereby reducing memory requirements for processing.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

4Manufacturing precision

If traditional methods are used to handle video with motion and varying lighting conditions, then processing simplicity is maintained, but video quality and artifact reduction are insufficient

Engineering Contradiction:
Improvevideo quality and artifact reductionVSAvoidhandling of complex scenarios
Core Design Contradiction:
Manufacturing precisionVSAdaptability or versatility

Solution Approach 1:

The patent applies dynamics by using recurrent neural network components and temporal convolution layers that adapt to varying lighting conditions and motion patterns in video data. The model dynamically adjusts its processing based on the specific characteristics of each video sequence, learning to handle different scenarios (reflections, shadows, occlusions, lighting changes) through trained temporal patterns rather than relying on fixed static processing rules.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP3857457B1Neural network systems for decomposing video data into layered representations
Publication Date: 2024.09.11 DEEPMIND TECH LTD
  • EP3857457B1 patent drawingFigure 1
  • EP3857457B1 patent drawingFigure 2
  • EP3857457B1 patent drawingFigure 3

AI summary

A computer-implemented neural network system for decomposing input video data. A video data input receives a sequence of video image frames. The sequence is encoded, using a 3D spatio-temporal encoder neural network, into a set of latent variables representing a compressed version of the sequence. A 3D spatio-temporal decoder neural network processes the set of latent variables to generate two or more sets of decomposed video data; these may be stored, communicated, and/or made available to a user interface. Input video including undesired features such as reflections, shadows, and occlusions may thus be decomposed into two or more video sequences, one in which the undesired features are suppressed, and another containing the undesired features.