NCC Layer Filter Design for Infrared Small Target Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for small target detection on infrared imagery are inefficient in distinguishing targets from clutter and false targets, and existing solutions lack a widely-accepted, off-the-shelf approach for real-time computation, especially in systems with limited data and computational constraints.
Innovation Solution
A machine learning approach using a normalized-cross-correlational (NCC) layer in neural networks for designing a filter bank, which is optimized for FPGA systems through the MAD-NCC layer, avoiding square root operations for real-time computation and minimizing resource use, to enhance target detection and reduce false alarms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If conventional filters (mean, median, top-hat, Gaussian) are used to reduce background clutter, then clutter reduction is achieved, but the ability to differentiate targets from false targets is insufficient
Solution Approach 1:
The patent transforms the filter design from fixed heuristic parameters to learned parameters through supervised training. The NCC layer learns optimal filter shapes and characteristics from training data, adapting parameters dynamically to distinguish targets from false targets while reducing clutter.
Solution Approach 2:
The patent implements supervised training with feedback loops where the neural network learns from labeled training data containing both targets and false targets. The training process provides feedback to adjust filter parameters, enabling the system to differentiate targets from false targets effectively.
2Measurement precision
If deep CNNs are used for target detection, then detection performance is improved, but the risk of overfitting increases when training data is limited
Solution Approach 1:
The patent uses a normalized cross-correlation layer instead of full deep CNNs, applying partial action by implementing only the essential correlation operation with normalization. This reduced architecture achieves effective detection while requiring fewer training parameters, thereby reducing overfitting risk when training data is limited.
Solution Approach 2:
The normalization operation changes the parameter space by standardizing inputs, which regularizes the learning process. This parameter transformation helps prevent overfitting by constraining the effective capacity of the model while maintaining detection performance.
3Measurement precision
If NCC-based matching is used for target detection, then template matching effectiveness is improved, but filter selection and design become ambiguous
Solution Approach 1:
The patent implements self-service by using supervised training to automatically learn optimal filters from data. The system serves itself by generating appropriate filter selections through training, eliminating the need for manual filter design or ambiguous filter selection processes.
4Measurement precision
If standard NCC layer is implemented for real-time computation, then detection accuracy is improved, but computational complexity increases due to square root operations
Solution Approach 1:
The patent replaces the computationally expensive square root operation with simpler arithmetic operations (subtraction and division) in the MAD-NCC layer. This substitution uses cheaper computational operations that can be executed faster, enabling real-time processing while maintaining detection accuracy.
5Measurement precision
If standard NCC layer is used, then normalization is achieved, but resource consumption increases for FPGA implementations
Solution Approach 1:
The patent replaces resource-intensive square root operations with simpler arithmetic operations in the MAD-NCC layer. This substitution uses computationally cheaper operations that consume fewer FPGA resources (LUTs, DSP blocks, memory) while achieving the same normalization objective.
Data Source
AI summary
A filter design method for a small target detection on infrared imagery using a normalized-cross-correlation layer in neural networks, including the steps of: Normalizing inputs and filters of a convolutional neural network, wherein normalizing inputs and filters of the convolutional neural network provides faster convergence in a limited database. Defining a forward function of a normalization layer in the convolutional neural network, wherein the forward function of the normalization layer in the convolutional neural network is used for training a neural network. Defining a derivative function of the normalization layer for a back propagation in a neural network training phase. Training created neural networks with datasets, wherein the datasets consist of target and background views and using trained neural networks in the small target detection.