[Paper Review] Small-floating Target Detection in Sea Clutter via Visual Feature Classifying in the Time-Doppler Spectra

Yi Zhou, Yin Cui+3 more|arXiv (Cornell University)|Sep 9, 2020

Underwater Acoustics Research19 references4 citations

TL;DR

This paper proposes a novel radar target detection method for small-floating objects in sea clutter by leveraging visual texture features in time-Doppler spectra (TDS). It uses Local Binary Pattern (LBP) to capture texture variations caused by targets disrupting sea surface motion, and applies one-class SVM to detect outliers in the LBP feature space, achieving superior detection rates on IPIX radar data compared to existing methods.

ABSTRACT

It is challenging to detect small-floating object in the sea clutter for a surface radar. In this paper, we have observed that the backscatters from the target brake the continuity of the underlying motion of the sea surface in the time-Doppler spectra (TDS) images. Following this visual clue, we exploit the local binary pattern (LBP) to measure the variations of texture in the TDS images. It is shown that the radar returns containing target and those only having clutter are separable in the feature space of LBP. An unsupervised one-class support vector machine (SVM) is then utilized to detect the deviation of the LBP histogram of the clutter. The outiler of the detector is classified as the target. In the real-life IPIX radar data sets, our visual feature based detector shows favorable detection rate compared to other three existing approaches.

Motivation & Objective

• Address the challenge of detecting small-floating targets in sea clutter where amplitude-based separation fails due to similar backscatter amplitudes.
• Overcome limitations of Doppler-based methods that struggle with slow-moving targets indistinguishable from wave motion.
• Identify generic visual texture features in TDS images that can reliably distinguish target-containing returns from clutter-only returns.
• Develop an unsupervised detection framework that operates effectively with limited labeled data and imbalanced samples.
• Improve detection performance over existing feature-based methods by using a robust, generic texture descriptor (LBP) and outlier detection via one-class SVM.

Proposed method

• Extract time-Doppler spectra (TDS) from radar returns to visualize the Doppler-frequency characteristics of sea clutter and target returns.
• Apply Local Binary Pattern (LBP) with P=8 neighbors and radius R=1 to compute texture features in the TDS images, using the 'uniform' format for robustness.
• Represent each TDS image as a 9-bin LBP histogram to encode local texture variations, capturing disruptions caused by target presence.
• Train a one-class Support Vector Machine (SVM) on LBP histograms from clutter-only cells to model the normal clutter distribution.
• Classify any TDS image whose LBP histogram lies outside the trained one-class SVM boundary as a potential target.
• Use a Gaussian kernel with bandwidth s=1/m (m = number of training samples) and set the ν parameter to 0.4 to handle imbalanced and noisy training data.

Experimental results

Research questions

• RQ1 Can visual texture differences in TDS images reliably distinguish small-floating targets from sea clutter?
• RQ2 Does the LBP feature effectively capture the structural disruption caused by a target in the otherwise coherent motion of sea surface in TDS?
• RQ3 Can one-class SVM effectively detect target-like outliers in the LBP feature space when only clutter samples are available for training?
• RQ4 How does the proposed LBP-based detector compare in detection rate to existing multi-feature or heuristic feature-based methods on real IPIX radar data?
• RQ5 To what extent does the LBP feature generalize across different polarization modes and environmental conditions in real-world radar data?

Key findings

• The LBP-based feature space successfully separates target-containing TDS images from clutter-only ones, as evidenced by distinct histogram distributions.
• In the 1993 Dartmouth IPIX data set, the proposed method achieved a detection rate of 0.90 for HH and VH polarizations, outperforming NTFD (0.754), Tri (0.577), and Fractal (0.303) methods.
• In the 1998 Grimsby data set, the method achieved a detection rate of 0.90 for HV and VH modes, surpassing NTFD (0.997 and 1.000 respectively) in some cases and significantly outperforming Tri and Fractal methods.
• The method maintained high detection performance (0.80 overall rate) across all polarizations in the 1993 data, with detection rates exceeding 0.70 in all modes.
• The LBP feature achieved comparable or better detection performance than the multi-feature NTFD method, demonstrating its value as a generic, low-dimensional texture descriptor.
• The one-class SVM classifier effectively detected targets as outliers even with limited training data and imbalanced samples, validating its robustness in real radar scenarios.