Ship Detection in Optical Remote Sensing Images Based on Wavelet Transform and Multi-Level False Alarm Identification

Fang Xu, Jinghong Liu, Chao Dong, Xuan WangRemote 2017 · DOI

TL;DR

This paper introduces a novel two-stage approach for detecting ships in optical remote sensing images. It first uses a wavelet-based saliency model to highlight potential ship candidates and then employs a multi-level discrimination method based on entropy and pixel distribution to effectively filter out false alarms from complex backgrounds like islands and clouds.

What Problem Does This Paper Address?

Detecting ships in optical remote sensing images is crucial for various applications, including maritime surveillance, rescue operations, and fishery management. While satellite and UAV imagery offers rich detail, it also presents significant challenges. These include the small size and varied appearance of ships, complex sea backgrounds with interferences like islands, coastlines, clouds, and shadows, and the need for efficient and accurate detection. Existing methods often struggle with these complexities, leading to either missed detections or a high rate of false alarms.

Methods

The proposed method operates in two main stages:

  1. Candidate Extraction via Visual Saliency:
    • Wavelet Decomposition: The input image is first converted to the CIE Lab color space and then decomposed using a multi-scale and multi-direction wavelet transform (Daubechies db.4). This captures detailed information across different frequency bands and orientations.
  2. Feature Map Generation: High-frequency wavelet coefficients are used to construct feature maps, neglecting the approximation information.
  3. Global Saliency Map Construction: A global saliency map is built by treating the feature vectors at each pixel position as a multivariate Gaussian probability density. Pixels with low probability are considered more salient. A distance decay formula is then applied to further refine the saliency map, emphasizing regions closer to high-saliency areas.
  4. Multi-Level False Alarm Identification:
    • Adaptive Segmentation: An Otsu method is used to segment potential ship regions from the saliency map. Candidates are filtered based on pixel count to remove very small or overly large regions.
  5. Entropy Discrimination: An improved entropy method is introduced, which considers spatial geometric information by applying a Gaussian filter before calculating entropy. This helps distinguish regular ship chip textures from irregular pseudo-target textures. A threshold is determined to classify chips as ship or non-ship.
  6. Pixel Distribution Discrimination: For chips with entropy values close to the threshold, a pixel distribution analysis is performed. This method uses ten judgment conditions based on the distribution of target pixels on the edges and within the chip to further eliminate false alarms caused by islands, clouds, or edge cases.

Datasets

The experiments were conducted using remote sensing images from the publicly available Google Earth service. A dataset of 273 representative images with resolutions ranging from 2-15 m was used. These images cover various scenarios, including different sea backgrounds, ship colors, sizes, and the presence of clouds, shadows, ship wakes, coastlines, small islands, and reefs. The paper also mentions the use of ground-truth maps for evaluation.

Key Insights and Ideas

  • Wavelets for Multi-Scale, Multi-Directional Features: The paper effectively leverages wavelet transforms to capture rich, multi-scale, and multi-directional features from the image, which are crucial for distinguishing ships from complex backgrounds.
  • Global Saliency for Robustness: Constructing a global saliency map based on wavelet coefficients, rather than relying solely on local contrast, helps in suppressing sea background noise and highlighting targets more consistently, even under challenging conditions.
  • Improved Entropy for Texture Regularity: The novel "improved entropy" method is a key innovation, addressing the limitations of traditional entropy by incorporating spatial information. This allows for a more accurate assessment of texture regularity, which is often different between ships and false alarms.
  • Multi-Level Discrimination for High Accuracy: The sequential application of adaptive segmentation, entropy, and pixel distribution discrimination provides a robust pipeline to progressively eliminate false alarms, leading to higher detection accuracy.
  • Handling Diverse Ship Characteristics: The method is designed to handle ships of various sizes and colors, as well as complex sea conditions, which are common challenges in real-world remote sensing applications.

Key Results

  • Superior Saliency Performance: The proposed Weighted Global Saliency (WGS) model outperformed eleven other state-of-the-art saliency models in subjective visual evaluations and quantitatively through ROC curves and AUC values. The WGS model demonstrated better background suppression and target highlighting.
  • High Detection Accuracy: The multi-level discrimination method significantly reduced false alarms. After all stages, the false alarm ratio (Far) was reduced to 5.741%, with a detection accuracy ratio (Cr) of 91.382%.
  • Robustness to Image Size: The WGS model showed better performance on images of varying sizes compared to other methods like SR, maintaining effective detection and integrity of ship regions.
  • Efficient Computation: While not the fastest, the proposed method offers a desirable speed performance, with the discrimination stages being particularly fast. The overall computation time for the discrimination stage is efficient, especially for smaller chip sizes.

Limitations and Future Work

  • Slender Reefs and Bright Coastlines: The paper acknowledges that some slender reefs or coastlines with bright edges can still be mistaken for ships, especially after binarization, due to similar entropy values. More effective features are needed to address these specific false alarms.
  • Computational Cost of Global Saliency: The calculation of the global saliency map using probability density on multiple feature maps is a time-consuming operation, contributing to the overall processing time.
  • Future Work: The authors suggest further research into more effective features to remove challenging false alarms like slender reefs and bright coastlines.

Summary updated

Abstract

Ship detection by Unmanned Airborne Vehicles (UAVs) and satellites plays an important role in a spectrum of related military and civil applications. To improve the detection efficiency, accuracy, and speed, a novel ship detection method from coarse to fine is presented. Ship targets are viewed as uncommon regions in the sea background caused by the differences in colors, textures, shapes, or other factors. Inspired by this fact, a global saliency model is constructed based on high-frequency coefficients of the multi-scale and multi-direction wavelet decomposition, which can characterize different feature information from edge to texture of the input image. To further reduce the false alarms, a new and effective multi-level discrimination method is designed based on the improved entropy and pixel distribution, which is robust against the interferences introduced by islands, coastlines, clouds, and shadows. The experimental results on optical remote sensing images validate that the presented saliency model outperforms the comparative models in terms of the area under the receiver operating characteristic curves core and the accuracy in the images with different sizes. After the target identification, the locations and the number of the ships in various sizes and colors can be detected accurately and fast with high robustness.
BibTeX
@inproceedings{xu2017ship,
  author = {Fang Xu and Jinghong Liu and Chao Dong and Xuan Wang},
  title = {{Ship Detection in Optical Remote Sensing Images Based on Wavelet Transform and Multi-Level False Alarm Identification}},
  booktitle = {Remote},
  year = {2017},
  doi = {10.3390/rs9100985},
}
Figure