Convolutional Neural Network Super Resolution for Infrared Imagery
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current IR uncooled technology produces low-resolution images, limiting their usefulness in applications such as agriculture for phenotype identification, irrigation monitoring, and early disease detection in plants.
Innovation Solution
A deep neural network approach using a convolutional neural network (CNN) is employed to generate high-resolution images from low-resolution images, specifically designed for low-power devices. This method utilizes depthwise-separable convolution and bottleneck layers to reduce computational complexity, enabling efficient image processing on low-power hardware.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If conventional CNN-based SR algorithms are used, then image resolution is improved, but computational complexity and power consumption increase
Solution Approach 1:
The patent segments the convolutional operations into depthwise convolution and pointwise convolution stages. The depthwise convolution processes each channel independently with 3x3 kernels, and the pointwise convolution uses 1x1 kernels to combine channels. This segmentation reduces computational complexity from O(k^3) to O(k^2) while maintaining image resolution enhancement capability.
Solution Approach 2:
The patent applies different convolution operations to different parts of the feature extraction process. Depthwise convolution is used for spatial feature extraction within each channel, while pointwise convolution is used for channel-wise feature integration. This localized application of different operations optimizes computational efficiency at each stage of the processing pipeline.
2Manufacturing precision
If deep neural networks are used for super-resolution, then image quality is improved, but power consumption increases
Solution Approach 1:
The network is segmented into multiple layers with bottleneck structures, where each layer performs localized feature extraction. This segmentation allows the model to process information in smaller, more energy-efficient steps rather than requiring a single large computational pass, reducing overall power consumption while maintaining image quality.
Solution Approach 2:
The patent changes the parameters of the convolutional operations by using depthwise separable convolutions instead of standard convolutions. This parameter change reduces the number of multiplications required per pixel, directly lowering computational power requirements while preserving the ability to generate high-quality super-resolution images.
3Measurement precision
If high-resolution image generation is performed, then detection capability is improved, but computational cost increases
Solution Approach 1:
The computational process is segmented into multiple processing stages with bottleneck layers that reduce feature dimensionality between stages. This segmentation allows the system to achieve high detection capability through progressive feature refinement while controlling computational cost by processing fewer features at each intermediate stage.
Solution Approach 2:
The patent transitions between different dimensional representations of the data by using bottleneck layers that compress feature dimensions and then expand them in subsequent layers. This dimensional transformation allows the network to capture complex patterns necessary for high detection capability while managing computational requirements through efficient dimensionality reduction and expansion.
Data Source
AI summary
An artificial neural network for processing low-resolution images to generate super-resolution images includes feed-forward connections between layers. The network includes an input layer, one or more convolution layers, wherein the input layer is connected to a first convolution layer of the convolution layers, and a output layer connected to a last convolution layer of the convolution layers. An input image is input to the input layer and to at least one of the convolution layers, an initial output of the input layer is input to at least one of the convolution layers, and a layer output of at least one of the convolution layers is input to at least one subsequent convolution layer.