Transposed Convolution for Spatial Dimension Preservation

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Advanced driver assistance systems (ADAS) and autonomous vehicle (AV) systems face challenges in preserving the spatial dimension of images processed by convolutional neural networks (CNNs), which can affect object segmentation and localization.

Innovation Solution

The use of transposed convolutional layers, which are trained jointly with convolutional layers, allows for improved spatial dimension preservation or reconstruction. This approach enables the use of a lower-dimensional representation of input images while maintaining the spatial integrity necessary for ADAS and AV applications.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If convolutional neural networks process images, then object detection capability is improved, but spatial dimension is lost

Engineering Contradiction:

Improveobject detection capabilityVSAvoidspatial dimension

Core Design Contradiction:

Measurement precisionVSShape

Solution Approach 1:

The patent applies transposed convolutional layers that perform the opposite operation of standard convolutional layers. While standard convolutions reduce spatial dimensions, transposed convolutions reconstruct and expand spatial dimensions by applying learned transformation weights in reverse, thereby recovering spatial information while maintaining object detection capabilities

Inventive Principle:

Principle #13The other way round (Inversion)

Solution Approach 2:

The patent changes the operational parameters of convolutional operations by using transposed convolutions with specific kernel configurations and padding strategies. This allows the system to transform downsampled feature maps back to higher resolution spatial representations, preserving spatial dimensions needed for precise object localization while maintaining the feature extraction benefits of convolutional processing

Inventive Principle:

Principle #35Parameter changes

2Shape

If image interpolation is used, then spatial dimension is preserved, but reconstruction accuracy deteriorates

Engineering Contradiction:

Improvespatial dimensionVSAvoidreconstruction accuracy

Core Design Contradiction:

ShapeVSManufacturing precision

Solution Approach 1:

The patent replaces traditional mechanical image interpolation methods (such as nearest-neighbor, bilinear, or bicubic interpolation) with a learned transposed convolutional system. This neural network-based approach dynamically learns optimal reconstruction transformations from data, achieving superior reconstruction accuracy compared to fixed interpolation algorithms while preserving spatial dimensions

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The transposed convolutional layers are trained jointly with the convolutional network, allowing the reconstruction process to adapt automatically to the specific features and characteristics of the input data. The system serves its own reconstruction needs by learning task-specific transformation parameters rather than relying on generic interpolation methods

Inventive Principle:

Principle #25Self-service

Data Source

PatentUS12299840B2Transposed convolution on downsampled data

Publication Date: 2025.05.13 MOBILEYE VISION TECH LTD

AI summary

The present subject matter provides technical solutions facing technical problems associated with preserving spatial dimension of images used in CNNs. Transposed convolutional layers may be used to provide improved spatial dimension preservation or reconstruction. In contrast with image interpolation, transposed convolutional layers may use a set of weights to reconstruct input images. When using CNNs for ADAS and AV applications, the transposed convolutional layers may be trained jointly with convolutional layers during the CNN training process. This may provide the ability to use a lower-dimensional representation of input images, while preserving the spatial dimension of images for use in ADAS and AV systems.