Encoding-Decoding Network Loss Layer Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

During the backpropagation process in CNNs for image segmentation, the loss values become smaller and smaller, making it difficult to adjust the parameters of each filter effectively, leading to suboptimal segmentation results.

Innovation Solution

The implementation of an encoding-decoding network with multiple encoding filters, corresponding decoding filters, and intermediate dilation convolution filters, along with multiple loss layers that interact with the decoding filters, allows for the computation and propagation of individual losses to each filter, ensuring accurate parameter optimization.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If a single loss layer is used in the decoding network, then the device complexity is reduced, but the loss values become too small during backpropagation to effectively adjust filter parameters

Engineering Contradiction:
Improvenumber of loss layersVSAvoidparameter adjustment accuracy
Core Design Contradiction:
Device complexityVSManufacturing precision

Solution Approach 1:

The patent divides the single loss computation into multiple separate loss layers, each computing loss independently at different stages of the decoding network. This segmentation allows each loss to be propagated separately through backpropagation, preventing the cumulative diminishment that occurs with a single loss layer and enabling effective parameter adjustment at each stage.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces an additional dimension to the loss computation by adding multiple loss layers along the network depth dimension. Instead of computing a single loss at the final output, losses are computed at multiple intermediate stages, creating a multi-layered loss structure that preserves gradient magnitudes throughout the network.

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

2Manufacturing precision

If multiple loss layers are added to compute individual losses for each filter, then the parameter optimization accuracy is improved, but the device complexity increases

Engineering Contradiction:
Improveparameter optimization accuracyVSAvoidnumber of loss layers
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

Each loss layer in the patent serves multiple functions: it computes the loss for its corresponding decoding filter, provides a gradient signal for backpropagation, and enables independent parameter optimization for that filter stage. This multi-functionality justifies the added complexity by delivering comprehensive parameter optimization benefits throughout the network.

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Productivity

If convolution operations are applied multiple times to reduce feature map size, then the computation efficiency is improved, but the loss values become too small for effective learning

Engineering Contradiction:
Improvecomputation efficiencyVSAvoidloss value magnitude
Core Design Contradiction:
ProductivityVSManufacturing precision

Solution Approach 1:

The patent performs preliminary loss computation at intermediate stages before the final decoding output. By computing losses earlier in the decoding process when feature maps are still relatively large and gradient magnitudes are stronger, the system captures useful gradient signals before they diminish through multiple subsequent convolution operations.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3467713B1Learning method and learning device for image segmentation and testing method and testing device using the same
Publication Date: 2022.08.17 STRADVISION
  • EP3467713B1 patent drawingFigure 1
  • EP3467713B1 patent drawingFigure 2A
  • EP3467713B1 patent drawingFigure 2B

AI summary

A method for improving image segmentation by using a learning device is disclosed. The method includes steps of: (a) if a training image is obtained, acquiring (2- K)th to (2-1)th feature maps through an encoding layer and a decoding layer, and acquiring 1st to Hth losses from the 1st to the Hth loss layers respectively corresponding to H feature maps, obtained from the H filters, among the (2-K)th to the (2-1)th feature maps; and (b) upon performing a backpropagation process, performing processes of allowing the (2-M)th filter to apply a convolution operation to (M-1)2-th adjusted feature map relayed from the (2-(M-1))th filter to obtain M1-th temporary feature map; relaying, to the (2-(M+1))th filter, M2-th adjusted feature map obtained by computing the Mth loss with the M1-th temporary feature map; and adjusting at least part of parameters of the (1-1)th to the (1-K)th filters and the (2-K)th to the (2-1)th filters.