Adaptive Loss Weight Map for Attention-Driven Image Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image segmentation methods using CNNs fail to properly learn important regions with small pixel counts, as they assign uniform weights to all areas, leading to inadequate learning and segmentation of critical regions.

Innovation Solution

The method employs an adaptive loss weight map to generate larger weights for important regions, using a CNN with encoding, decoding, softmax, and loss weight layers to generate feature maps, segmentation scores, softmax scores, predicted error values, and adjusted softmax loss values, allowing for increased weight differences and improved learning of critical regions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If uniform weights are assigned to all areas in image segmentation, then the learning process is simple and computationally efficient, but important regions with small pixel counts cannot be properly learned

Engineering Contradiction:
Improvesegmentation accuracy of important regionsVSAvoidcomplexity of loss function
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent applies local quality by introducing a loss weight map that assigns different weights to different spatial regions of the image. Specifically, important regions (such as objects of interest) are assigned higher weights while background regions receive lower weights. This is achieved through the loss weight layer that generates the weight map based on the input image and ground truth, and then multiplies these weights with the softmax loss values for corresponding pixels during backpropagation.

Inventive Principle:
Principle #3Local quality

2Measurement precision

If adaptive loss weights are applied to emphasize important regions, then segmentation accuracy of critical areas improves, but the computational complexity and number of operations increase

Engineering Contradiction:
Improvesegmentation precision of small regionsVSAvoidlearning speed
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent applies preliminary action by pre-processing the input image and ground truth through the loss weight layer to generate the loss weight map before computing the final loss. This allows the network to focus on important regions from the beginning of the learning process, rather than requiring multiple iterations or post-processing steps to achieve the same effect.

Inventive Principle:
Principle #10Preliminary action

3Reliability

If standard softmax loss is used for all pixels, then the training process is straightforward and converges quickly, but errors in important regions are not sufficiently reflected in the loss

Engineering Contradiction:
Improveerror reflection in loss functionVSAvoidstructure of loss calculation
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent applies feedback by using the generated loss weight map to modulate the softmax loss calculation. The loss weight layer computes weights based on the input image and ground truth, then feeds these weights back into the loss calculation process. This creates a feedback loop where the network adapts its learning focus based on the importance of different regions, ensuring that errors in critical areas have a proportionally larger impact on the gradient updates.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP3686780B1Learning method and learning device for attention-driven image segmentation by using at least one adaptive loss weight map to be used for updating HD maps required to satisfy level 4 of autonomous vehicles and testing method and testing device using the same
Publication Date: 2023.03.01 STRADVISION
  • EP3686780B1 patent drawingFigure 1
  • EP3686780B1 patent drawingFigure 2
  • EP3686780B1 patent drawingFigure 3

AI summary

A method for an attention-driven image segmentation by using at least one adaptive loss weight map is provided to be used for updating HD maps required to satisfy level 4 of autonomous vehicles. By this method, vague objects such as lanes and road markers at distance may be detected more accurately. Also, this method can be usefully performed in military, where identification of friend or foe is important, by distinguishing aircraft marks or military uniforms at distance. The method includes steps of: a learning device instructing a softmax layer to generate softmax scores; instructing a loss weight layer to generate loss weight values by applying loss weight operations to predicted error values generated therefrom; and instructing a softmax loss layer to generate adjusted softmax loss values by referring to initial softmax loss values, generated by referring to the softmax scores and their corresponding GTs, and the loss weight values.