Controlled Distortion Imager for Neural Network Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks are limited in processing high-resolution images, especially for applications requiring global image analysis, due to their restricted input image resolution, which leads to poorer analysis outcomes and necessitates new training methods for handling images with on-purpose controlled distortion.
Innovation Solution
A method to train and utilize convolutional neural networks with images having targeted distortion, where an imager creates digital images with controlled distortion, increasing resolution in specific zones, and a distorted image dataset generator processes existing datasets to add distortion, allowing the network to learn and process these images effectively.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If convolutional neural networks process images at their maximum supported resolution (512x512), then processing speed and computational efficiency are maintained, but the analysis precision and quality of global image analysis deteriorate when full-resolution images (20-50 MPx) are required
Solution Approach 1:
The patent divides the image processing task into two segments: a first neural network processes a first image (potentially at lower resolution or with different preprocessing), and a second neural network processes a second image (at full resolution or with enhanced features). This segmentation allows each network to be optimized for its specific resolution and task, enabling full-resolution analysis without requiring a single overly complex network that would be computationally infeasible.
Solution Approach 2:
The patent implements a nested architecture where one neural network is embedded within another. The first neural network processes input and generates intermediate results that are fed into the second neural network, which performs further processing at higher resolution. This nested structure allows progressive refinement of image analysis from lower to higher resolutions, achieving high-precision global analysis while managing computational complexity through hierarchical processing.
2Loss of information
If the input image resolution is increased to 512x512 or higher, then more global scene information is available for analysis, but the processing power requirements and computational resources increase significantly
Solution Approach 1:
By segmenting the processing across two neural networks operating at different resolutions and with different computational demands, the patent reduces the peak processing power required compared to processing the entire high-resolution image in a single network. The first network handles lower-resolution or preprocessed data with lower computational cost, while the second network focuses on specific regions or features at higher resolution, optimizing the energy-information tradeoff.
Solution Approach 2:
The patent applies partial action by having the first neural network process the entire image at a baseline resolution to extract global features, then the second neural network applies excessive action by processing specific regions or enhanced versions of the image at higher resolution. This selective high-resolution processing provides sufficient global scene information for the second network's specific tasks without requiring the first network to consume excessive processing power.
3Area of stationary object
If existing neural networks are used with distorted images from wide-angle imagers, then the field of view is increased, but the analysis accuracy deteriorates because the networks are trained on non-distorted images
Solution Approach 1:
The patent changes the training parameters of the neural networks by providing distorted images (from wide-angle imagers) as input to the first neural network and generating correspondingly distorted target outputs for training the second neural network. This parameter change in the training data distribution allows the networks to learn the distortion characteristics and compensate for them, maintaining analysis accuracy while utilizing the increased field of view provided by wide-angle imaging.
Solution Approach 2:
The patent creates a distorted copy of the input image through the first neural network's processing and uses this distorted representation as the basis for training the second neural network. By copying and transforming the input data through the distortion process, the system enables the second network to learn how to accurately analyze distorted images while preserving the benefits of the wide-angle field of view.
Data Source
AI summary
A method for training and using a convolutional neural network with images having on-purpose distortion is disclosed. By distorting on-purpose the image to create zone of interest with higher number of pixels, the resulting output from the neural network is improved. The distorted images are created using an imager device either comprising an optical system specifically designed to output distorted images or comprising software or hardware image distortion manipulation algorithm to create the distorted images from normal images. Also described is a method for training the neural network using a distorted image generator from various existing datasets.