Neural Network Obstruction Classification Using 1D Convolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous devices like robots face challenges in accurately determining whether their optical sensors are obstructed, which can lead to inaccurate environmental sensing and subsequent inappropriate actions.
Innovation Solution
A neural network is designed to classify whether an optical sensor is obstructed by using a configuration with two parallel one-dimensional convolutional layers that process images from the sensor, allowing for accurate obstruction detection even on embedded hardware with limited computational power.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a neural network is designed with high accuracy for obstruction detection, then the classification accuracy is improved, but the computational complexity and resource requirements increase
Solution Approach 1:
The neural network is segmented into two separate 1D-convolutional layers, each processing one dimension (height and width) independently. This segmentation allows the network to achieve high classification accuracy through specialized dimension-specific processing while keeping each individual layer computationally simple and efficient for embedded hardware deployment.
Solution Approach 2:
The patent transitions from traditional 2D-convolutional layers to 1D-convolutional layers, changing the dimensionality of the convolution operation. This dimensional reduction simplifies the computational complexity by processing only one dimension at a time, while maintaining high accuracy through the use of two parallel layers that collectively process both dimensions of the input image.
2Use of energy by moving object
If a neural network is optimized for embedded hardware with limited computational power, then the resource efficiency is improved, but the classification accuracy may deteriorate
Solution Approach 1:
By segmenting the convolutional processing into two separate 1D-convolutional layers, each layer can be optimized for embedded hardware with lower computational requirements. The segmentation enables resource-efficient processing on mobile robots and battery-driven devices while maintaining high classification accuracy through the combined output of both layers.
Solution Approach 2:
The dimensional reduction from 2D to 1D convolution significantly decreases the computational burden and resource consumption, making the neural network suitable for embedded hardware. The use of two parallel 1D layers compensates for the reduced complexity by collectively capturing features from both height and width dimensions, thereby preserving classification accuracy.
Data Source
AI summary
A neural network configured for classifying whether an image from an optical sensor characterizes an obstruction of the optical sensor or not. The classification is characterized by an output of the neural network for an input of the neural network and wherein the input is based on the image. The neural network comprises a first convolutional layer that characterizes a 1D-convolution along a vertical axis of a convolution output of a preceding convolutional layer and a second convolutional layer that characterizes a 1D-convolution along a horizontal axis of the convolution output. The output of the neural network is based on a first convolution output of the first convolutional layer and based on a second convolution output of the second convolutional layer.