CNN Automatic Target Recognition in Resource-Constrained Satellites
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural networks are computationally intensive and require significant memory, making them difficult to implement in resource-constrained environments such as satellites, where size, mass, and complexity are minimized, limiting their effective use for image identification and automatic target recognition.
Innovation Solution
Implementing a convolutional neural network (CNN) based automatic target recognition system in a hardware-constrained environment by breaking image data into smaller 'chips' for parallel processing, using quantized weight values, and streaming image data with minimal buffering to reduce memory and computational requirements, suitable for implementation on a Field Programmable Gate Array (FPGA).
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a neural network is implemented for image identification, then recognition accuracy is improved, but memory usage and computational complexity increase
Solution Approach 1:
The patent divides the image data into smaller chips and processes them in parallel using multiple processing nodes. This segmentation approach reduces the computational burden on each individual node while maintaining overall recognition accuracy through collective processing of all chips.
Solution Approach 2:
The patent applies quantization to reduce the precision of weight values from 32-bit floating point to 8-bit integers. This parameter change significantly reduces memory usage and computational complexity while preserving sufficient accuracy for satellite imaging applications.
2Measurement precision
If a neural network is implemented for image identification, then recognition accuracy is improved, but memory usage increases
Solution Approach 1:
The patent quantizes the weight values from high-precision 32-bit floating point numbers to low-precision 8-bit integers. This parameter change reduces the memory required to store weights by a factor of 4, making the system feasible for satellite deployment with limited memory resources.
Solution Approach 2:
By dividing the image into multiple chips and processing them in parallel, the patent reduces the amount of data that needs to be held in memory at any given time, thereby reducing peak memory usage while maintaining recognition accuracy.
3Quantity of substance
If image data is processed with minimal buffering, then memory usage is reduced, but processing complexity increases
Solution Approach 1:
The patent segments the image into fixed-size chips that can be processed independently and in parallel. This segmentation enables streaming processing with minimal buffering, as each chip can be handled as a discrete unit without requiring large memory buffers to hold entire images.
Solution Approach 2:
The patent performs preliminary processing by dividing the image into chips and organizing them for parallel processing before the actual neural network computation. This preliminary organization reduces the need for complex buffering and random access operations during the main processing phase.
Data Source
AI summary
Methods and structures are presented for implementing an automatic target recognition system as a convolutional neural network (CNN) in a satellite or other environment with constrained resources, such as limited memory capacity and limited processing capability. For example, this allows for the automatic target recognition to be implemented on a field programmable gate array (FPGA). Image data is split into subsets of contiguous pixels, with the subsets processed in parallel in a CNN of a corresponding processing node using quantized weight values that are determined in a training process that accounts for the constraints of the automatic target recognition system. The results of the automatic target recognition process is based on the combined output of the processing nodes.