Neural Network Encoder Adaptation for Low-Visibility Semantic Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing semantic segmentation models struggle to maintain accuracy in varying weather conditions, particularly in low visibility due to fog or snow, as they are typically trained on high visibility images and lack sufficient data for adaptation to low visibility scenarios.

Innovation Solution

A method and system for adapting a neural network model by dividing it into an encoder and a second part, where the encoder part's weights are adjusted using random images from both source (high visibility) and target (low visibility) domains to minimize distances in feature and probability distributions, allowing the model to perform semantic segmentation across varying conditions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If a semantic segmentation model is trained on labeled images from multiple weather conditions, then the model's performance in varying weather conditions improves, but the time and resources required for training increase significantly

Engineering Contradiction:
Improvemodel performance in varying weather conditionsVSAvoidtraining time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The model is pre-trained on source domain images (good weather conditions) before being adapted to the target domain (bad weather conditions). This preliminary training establishes a baseline performance that can then be fine-tuned with domain adaptation techniques, reducing the overall training time compared to training from scratch on all weather conditions.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The training process is segmented into two distinct phases: source domain training and target domain adaptation. The model architecture is also segmented into an encoder part (adapted to target domain) and a second part (fixed from source domain). This segmentation allows efficient resource utilization by only adapting specific parts of the model to the target domain.

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If labeled images from target domain (low visibility conditions) are collected and used for training, then the model's accuracy in low visibility conditions improves, but the difficulty and time of obtaining such data increase

Engineering Contradiction:
Improvemodel accuracy in low visibility conditionsVSAvoidease of obtaining labeled training data
Core Design Contradiction:
Measurement precisionVSEase of manufacture

Solution Approach 1:

The method uses unlabeled images from the target domain (low visibility conditions) instead of requiring labeled images. The encoder part of the model is adapted to the target domain by minimizing distribution distances between source and target domains, effectively copying the learned features from the source domain and adapting them to the target domain without requiring manual annotations.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The domain adaptation process is self-supervised, using the unlabeled target domain images themselves to guide the adaptation. The model automatically learns to adapt to the target domain by minimizing distribution distances, without requiring external labeling or manual intervention, thus serving itself with the available unlabeled data.

Inventive Principle:
Principle #25Self-service

3Reliability

If the entire model is adapted to the target domain, then the model's performance in target domain improves, but the complexity of the adaptation process increases

Engineering Contradiction:
Improvemodel performance in target domainVSAvoidadaptation process complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The model is divided into two parts: an encoder part that is adapted to the target domain and a second part that remains fixed from the source domain training. This segmentation reduces the number of parameters that need to be adapted, simplifying the adaptation process while still improving target domain performance through encoder adaptation.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The adaptation process extracts only the necessary components (encoder part parameters) for modification while leaving the rest of the model (second part) unchanged. This extraction approach reduces the complexity of adaptation by focusing only on the parts of the model that need to be domain-specific.

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP3872695B1A method and system of adapting an initial model of a neural network
Publication Date: 2025.09.24 TOYOTA JIDOSHA KK
  • EP3872695B1 patent drawingFigure 1
  • EP3872695B1 patent drawingFigure 2~3
  • EP3872695B1 patent drawingFigure 4~5

AI summary

This method adapts an initial model (Mγ̂) trained with labeled images (xs,ys) of a source domain into an adapted model (Mγ). It comprises: - copying the initial model into the adapted model; - dividing the adapted model into an encoder part (E) and a second part (F), wherein the second part (F) is configured to process features (f̂s) output from said encoder part (E); - adapting said adapted model to a target domain (xt) using images (xs) of the source and target domains while fixing the parameters (WFi) of said second part (F) and minimizing (E236) a function (f) of the following two distances: - a distance (D1) between features (f̂s, fs) of the source domain (xs) output of the encoders (E) of the initial model and of the adapted model; and - a distance (D2) measuring a distribution distance between probabilities (Pr(f̂s)∼p) of features obtained for images of the source domain and of the target domain.