Aspect-Aided Dynamic Non-Negative Sparse Representation-Based Microwave Image Classification

Xinzheng Zhang, Qiuyue Yang, Miaomiao Liu, Yunjian Jia, Shujun Liu, Guojun LiSensors 2016 · DOI

TL;DR

This paper introduces a novel "Aspect-Aided Dynamic Non-Negative Sparse Representation" (ADNNLS) algorithm for classifying microwave radar images. By dynamically selecting relevant training data based on the target's viewing angle (aspect angle), ADNNLS significantly improves classification accuracy compared to methods that consider all available training data, especially in challenging scenarios.

What Problem Does This Paper Address?

Classifying targets in microwave radar images is crucial for applications like security and surveillance. However, these images are highly sensitive to the viewing angle (aspect angle) of the target. Even small changes in aspect angle can drastically alter the image, making traditional classification methods struggle. Existing approaches, like template matching or standard sparse representation, often fail to account for this aspect-dependent variability, leading to reduced accuracy. The challenge is to develop a classification method that can effectively handle these significant image variations caused by changing aspect angles.

Methods

The core innovation of this paper is the Aspect-Aided Dynamic Non-Negative Sparse Representation (ADNNLS) algorithm. Here's a breakdown of its key steps:

  1. Aspect Angle Estimation: For each incoming test image, the algorithm first estimates its aspect angle. This is done by correlating the test image with a library of training images with known aspect angles and selecting the aspect angle of the most correlated training image.
  2. Aspect Sector Determination: Based on the estimated aspect angle, a "sector" of relevant viewing angles is defined. Only training samples within this sector are considered potential candidates for building the classification dictionary.
  3. Smooth Self-Representation Clustering: Within this aspect sector, a "smooth self-representation learning" technique is employed to further refine the selection of training samples. This method divides the training samples into two groups: those closely correlated with the test sample (active atoms) and those less correlated (inactive atoms).
  4. Dynamic Dictionary Construction: For each test sample, a dynamic dictionary is created using only the identified "active atoms." This means the dictionary is tailored to the specific viewing angle of the test image, effectively discarding irrelevant training data.
  5. Non-Negative Sparse Representation: The test image is then represented using a non-negative sparse representation model with respect to this dynamic dictionary. This process aims to find a sparse set of coefficients that best reconstructs the test image from the selected active atoms.
  6. Classification: The final class label is determined by minimizing the reconstruction error across different target classes.

The "dynamic" aspect refers to the dictionary being updated for each test sample, and "aspect-aided" highlights the crucial role of aspect angle information in this selection process. The "non-negative" constraint is also important, as it aligns better with human visual perception and can lead to more effective representations.

Datasets

The experiments were conducted using the Moving and Stationary Target Acquisition and Recognition (MSTAR) database. This database contains synthetic aperture radar (SAR) images of various targets, including BMP2, BTR70, and T72, captured at different aspect and depression angles. The paper specifically utilized images of BMP2, BTR70, and T72 for training and testing, with variations in aspect angles from 0° to 360° and specific depression angles (17° for training, 15° for testing). Additional experiments also used a dataset with four targets (BRDM2, 2S1, ZSU23/4, SLICY) to evaluate performance under varying depression angles.

Key Insights and Ideas

  • Aspect Angle is Paramount: The paper strongly emphasizes that aspect angle is the most critical factor influencing microwave radar image characteristics. Ignoring this leads to significant performance degradation.
  • Dynamic Dictionaries are Key: Instead of using a static, all-encompassing dictionary, ADNNLS creates a tailored dictionary for each test sample based on its viewing angle. This "dynamic" approach is crucial for handling aspect variations.
  • "Active Atoms" for Efficiency: The concept of "active atoms" (relevant training samples) and "inactive atoms" (irrelevant training samples) within an aspect sector is a clever way to reduce computational complexity and noise interference.
  • Smooth Self-Representation for Refinement: The use of smooth self-representation clustering within the aspect sector provides a more robust way to select the most correlated training samples, further enhancing the dictionary's relevance.
  • Non-Negativity Matters: The non-negative constraint in the sparse representation is shown to be beneficial, aligning better with visual perception and potentially improving sparsity.

Key Results

  • Improved Accuracy: ADNNLS significantly outperformed traditional Sparse Representation Classifier (SRC) and Non-Negative Least Squares (NNLS) methods. For example, in one experiment, ADNNLS achieved a recognition rate of 99.49% at 800 dimensions, while SRC, NNLS, and an aspect-aided SRC (ADSRC) achieved 92.23%, 96.8%, and 96.97%, respectively.
  • Robustness to Aspect Angle: The algorithm demonstrated superior performance across various aspect sectors, with the highest recognition rate (99.63%) achieved with a 30° aspect sector. Even with a full 360° sector, it maintained a high rate of 97.39%.
  • Robustness to Noise: ADNNLS showed better robustness to noise compared to other algorithms, maintaining over 90% recognition rate across a broad range of Signal-to-Noise Ratios (SNRs) from 5 dB to 50 dB.
  • Robustness to Depression Angle: Even with large variations in depression angle (from 17° training to 45° testing), ADNNLS maintained a high accuracy (79.21%), outperforming other methods.
  • Reduced Interference: The aspect-aided approach effectively reduced interference from training samples with poor aspect correlation, leading to more accurate inference.

Limitations and Future Work

  • Aspect Angle Estimation Error: While the aspect angle estimation method is effective, it is not perfect and can have errors (around 2° to 5° in the experiments). The paper acknowledges that the sector size needs to be large enough to accommodate these errors.
  • Computational Cost: While dynamic dictionary construction reduces overall computation compared to brute-force methods, the process of aspect angle estimation and clustering adds computational overhead.
  • Future Work: The paper suggests further research into the influence of PCA feature dimensions on performance and comparisons with other state-of-the-art algorithms like IJSRC. The authors also mention that the proposed method showed promising anti-noise behavior, implying further investigation into its noise resilience could be beneficial.

Summary updated

Abstract

Classification of target microwave images is an important application in much areas such as security, surveillance, etc. With respect to the task of microwave image classification, a recognition algorithm based on aspect-aided dynamic non-negative least square (ADNNLS) sparse representation is proposed. Firstly, an aspect sector is determined, the center of which is the estimated aspect angle of the testing sample. The training samples in the aspect sector are divided into active atoms and inactive atoms by smooth self-representative learning. Secondly, for each testing sample, the corresponding active atoms are selected dynamically, thereby establishing dynamic dictionary. Thirdly, the testing sample is represented with ℓ 1 -regularized non-negative sparse representation under the corresponding dynamic dictionary. Finally, the class label of the testing sample is identified by use of the minimum reconstruction error. Verification of the proposed algorithm was conducted using the Moving and Stationary Target Acquisition and Recognition (MSTAR) database which was acquired by synthetic aperture radar. Experiment results validated that the proposed approach was able to capture the local aspect characteristics of microwave images effectively, thereby improving the classification performance.
BibTeX
@inproceedings{zhang2016aspect,
  author = {Xinzheng Zhang and Qiuyue Yang and Miaomiao Liu and Yunjian Jia and Shujun Liu and Guojun Li},
  title = {{Aspect-Aided Dynamic Non-Negative Sparse Representation-Based Microwave Image Classification}},
  booktitle = {Sensors},
  year = {2016},
  doi = {10.3390/s16091413},
}
Figure