Grid Cell Weighted Convolution Filters for Image Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image segmentation methods using deep learning technologies apply a single convolution filter to all areas of an image, leading to inefficient processing and suboptimal results when the image contains similar arrangements of objects and roads, such as in advanced driver assistance systems, as they do not utilize the similarities effectively.

Innovation Solution

The method employs a convolutional neural network (CNN) to generate encoded and decoded feature maps, which are then divided into grid cells, allowing each grid cell to have its own weighted convolution filters. These filters are learned through backpropagation using segmentation loss and ground truth, with the largest area serving as a reference for linear transformation of filter values for other cells.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If a single convolution filter with the same weight is applied to all areas of the image, then the processing is simpler and more uniform, but the segmentation accuracy deteriorates for images with similar arrangements of objects and roads

Engineering Contradiction:
Improvefilter application complexityVSAvoidsegmentation accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The image is divided into multiple grid cells, and each grid cell is processed by a dedicated convolution filter with unique weights. This segmentation approach allows different regions with similar arrangements (e.g., roads, sidewalks, buildings) to be processed with specialized filters that capture their specific characteristics, thereby improving segmentation accuracy while maintaining manageable complexity through systematic organization.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Different convolution filters are assigned to different grid cells based on their spatial locations and characteristics. Each filter is locally optimized for its specific region, allowing the system to adapt to local patterns such as road arrangements, building layouts, and sidewalk configurations. This local specialization improves segmentation precision without requiring a complete redesign of the entire filtering system.

Inventive Principle:
Principle #3Local quality

2Speed

If a single convolution filter is applied to the whole area, then the computational process is faster, but the processing efficiency deteriorates due to unnecessary operations on similar areas

Engineering Contradiction:
Improveprocessing speedVSAvoidprocessing efficiency
Core Design Contradiction:
SpeedVSProductivity

Solution Approach 1:

By dividing the image into grid cells and applying specialized filters to each cell, the system eliminates redundant computations that would occur with a single global filter. Each filter is trained specifically for its region, so operations are performed only where needed with appropriate specificity, improving overall processing efficiency despite the increased number of filters.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system changes the parameters (weights) of convolution filters based on the specific characteristics of each grid cell. This parameter adaptation allows the processing to be more efficient by focusing computational resources on distinguishing features specific to each region, rather than applying uniform processing that wastes cycles on already-similar areas.

Inventive Principle:
Principle #35Parameter changes

3Measurement precision

If multiple weighted convolution filters are applied to each grid cell, then the segmentation accuracy is improved, but the device complexity increases

Engineering Contradiction:
Improvesegmentation accuracyVSAvoidfilter configuration complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The image is systematically divided into grid cells, providing a structured framework for organizing multiple convolution filters. This segmentation approach manages complexity by creating a regular, predictable pattern of filter deployment, where each cell receives a dedicated filter. The systematic organization allows the system to handle multiple filters efficiently without becoming unmanageably complex.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The convolutional neural network architecture provides a universal framework that can handle multiple filters across multiple grid cells using the same fundamental operations. This multi-functionality allows the system to process many regions with different characteristics using a consistent methodological approach, reducing the effective complexity by applying universal principles across diverse situations.

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

4Manufacturing precision

If multiple weighted convolution filters are learned through backpropagation, then the segmentation precision is improved, but the training time and computational resources increase

Engineering Contradiction:
Improvesegmentation precisionVSAvoidtraining time
Core Design Contradiction:
Manufacturing precisionVSLoss of time

Solution Approach 1:

By segmenting the image into grid cells and training dedicated filters for each cell, the overall training problem is divided into smaller, more manageable sub-problems. This segmentation of the learning task allows for more efficient optimization, as each filter can be trained on a smaller subset of data with specific local characteristics, reducing the effective training time compared to training a single global filter on the entire image.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system performs preliminary organization by dividing the image into grid cells before the actual filter training begins. This preliminary action structures the training data in a way that facilitates more efficient learning, allowing each filter to focus on specific regional patterns from the outset. This pre-organization reduces the computational burden during the actual backpropagation process by eliminating the need for filters to learn all patterns from scratch.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3686792A1Method and device for segmenting image to be used for surveillance using weighted convolution filters for respective grid cells by converting modes according to classes of areas to satisfy level 4 of autonomous vehicle, and testing method and testing device using the same
Publication Date: 2020.07.29 STRADVISION
  • EP3686792A1 patent drawingFigure 1
  • EP3686792A1 patent drawingFigure 2
  • EP3686792A1 patent drawingFigure 3

AI summary

A method for segmenting an image by using each of a plurality of weighted convolution filters for each of grid cells to be used for converting modes according to classes of areas is provided to satisfy level 4 of an autonomous vehicle. The method includes steps of: a learning device (a) instructing (i) an encoding layer to generate an encoded feature map and (ii) a decoding layer to generate a decoded feature map; (b) if a specific decoded feature map is divided into the grid cells, instructing a weight convolution layer to set weighted convolution filters therein to correspond to the grid cells, and to apply a weight convolution operation to the specific decoded feature map; and (c) backpropagating a loss. The method is applicable to CCTV for surveillance as the neural network may have respective optimum parameters to be applied to respective regions with respective distances.