Image Decoder Conv-LSTM for Hierarchical Feature Retention

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional CNN-based image processing techniques face challenges in retaining higher-level information during the decoding process, leading to memory inefficiencies and difficulty in making accurate predictions due to the gradual forgetting of extracted features.

Innovation Solution

The use of a Conv-LSTM module in the decoder to sequentially process hierarchical feature maps, allowing for the retention and propagation of important information across multiple semantic levels without requiring excessive memory, using a single Conv-LSTM unit to update and transform hidden states.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If all hierarchical feature maps from the encoder are retained and used in the decoder, then prediction accuracy is improved, but memory requirements and system complexity increase prohibitively

Engineering Contradiction:
Improveprediction accuracyVSAvoidmemory requirements
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent extracts only the essential higher-level information from the encoder's hierarchical feature maps rather than retaining all feature maps. The Conv-LSTM module selectively processes and retains key contextual information across multiple semantic levels, discarding redundant data that would consume excessive memory while preserving the information necessary for accurate predictions.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent transforms the representation of hierarchical feature maps by processing them through a Conv-LSTM module, which changes the parameters of information retention. Instead of storing all feature maps with their original dimensions, the module compresses and transforms them into a more efficient representation that maintains predictive accuracy while reducing memory footprint.

Inventive Principle:
Principle #35Parameter changes

2Loss of information

If all hierarchical feature maps are processed simultaneously in the decoder, then information completeness is improved, but computational complexity and difficulty of processing increase

Engineering Contradiction:
Improveinformation completenessVSAvoidcomputational complexity
Core Design Contradiction:
Loss of informationVSDevice complexity

Solution Approach 1:

The patent implements continuous processing of hierarchical feature maps through the Conv-LSTM module, which sequentially integrates information across multiple semantic levels. This continuous action allows the module to maintain information completeness by progressively incorporating features from different hierarchical levels without requiring simultaneous processing of all maps, thereby reducing computational complexity.

Inventive Principle:
Principle #20Continuity of useful action

Solution Approach 2:

The patent introduces a temporal dimension to the processing of hierarchical feature maps through the recurrent nature of the Conv-LSTM module. By processing feature maps across time steps rather than simultaneously, the system transforms the computational problem from a high-dimensional simultaneous processing task into a sequential process that is more manageable while preserving information completeness.

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

3Ease of operation

If conventional CNN-based techniques are used to process hierarchical feature maps, then implementation simplicity is maintained, but ability to retain higher-level information deteriorates

Engineering Contradiction:
Improveimplementation simplicityVSAvoidretention of higher-level information
Core Design Contradiction:
Ease of operationVSLoss of information

Solution Approach 1:

The patent introduces the Conv-LSTM module as an intermediary between the encoder and decoder in the U-Net architecture. This intermediary component specifically addresses the limitation of conventional CNNs by providing a mechanism to retain and propagate higher-level information through its recurrent structure, while still integrating seamlessly into the existing encoder-decoder framework, thus maintaining a balance between implementation simplicity and information retention.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3617947B1Apparatus and method for processing image data
Publication Date: 2026.04.22 NOKIA TECHNOLOGIES OY
  • EP3617947B1 patent drawingFigure 1~2
  • EP3617947B1 patent drawingFigure 3
  • EP3617947B1 patent drawingFigure 4~5

AI summary

Apparatus for processing image data associated with at least one input image, comprising a convolutional neural network, CNN, -based encoder configured to provide a plurality of hierarchical feature maps based on said image data, a decoder configured to provide output data based on said plurality of feature maps, wherein said decoder comprises a convolutional long short-term memory, Conv-LSTM, module configured to sequentially process at least some of said plurality of hierarchical feature maps.