Visual Attention-Based Ship Detection And Recognition In Remote Sensing Images

Recently, ship detection and recognition in remote sensing imagery is becoming an intriguing subject for more and more researches because of its potential applications in areas such as fishery management, vessel traffic services, and naval warfare. The traditional detectors need to check all regions of the image carefully, but in fact the targets interest us only take up a small part of the whole image. This thorough processing demands high computational cost and increases the difficulty of target detection. However, the human vision can rapidly focus on conspicuous objects in clustered environments and selects salient visual information for further processing because of the existence of visual attention mechanism. This paper is primarily concerned with how to develop a visual attention-based system for ship detection and recognition in remote sensing imagery. The proposed attention-based system can be applied to ship detection in multispectral imagery, and also ship detection and recognition in single band optical images with high resolution. Moreover, we extend the existing frequency-based PQFT model so that it can be applied to saliency detection in multispectral imagery. The main innovations of this paper can be described into three aspects as follows:1. We introduce the selective visual attention mechanism into target detection in multispectral imagery. Since the existing computational models of visual attention are not suitable to process multi-dimensional data with its dimension more than four, we propose an approach for visual attention based on biquaternion. The proposed approach describes high-dimensional data in the form of biquaternion and utilizes its phase spectrum of biquaternion Fourier transform (PBFT) to generate the required saliency map that can be used for salient target detection. In our method, the multi-dimensional data can be processed as a whole, and features both in spatial and frequency domain can be extracted effectively. Compared with the traditional multispectral target detection method, our method has very low computational complexity and does not rely on parameter settings.2. The PFT model, which is based on phase spectrum of Fourier transform, is applied for saliency detection in single band optical images with high resolution. Using this attention model, we can detect ship candidates accurately under complex background. Combing the hierarchical discriminant regression tree (HDR Tree), a supervised classification approach based on shape and texture features is presented to distinguish between ships and nonships to remove most false alarms. In this way, we construct an salient object detection scheme that combines top-down with bottom-up processing.3. After distinguishing between ships and nonships, the two-dimensional principal component analysis (2D-PCA) is introduced to enhance the representation ability of the feature set in feature extraction. Combing the HDR Tree, a supervised classification approach based on finer texture features is presented to distinguish between civil cargo ships and military vessels. Experimental results on real remote sensing data show that the texture feature can be extracted effectively by using 2D-PCA and our method performs well in ship classification. Furthermore, the processing speed of our method is quite satisfying.Index Terms:Selective Visual Attention, Remote Sensing Imagery, Ship Detection and Recognition, Biquaternion, Fourier Transform, Shape Features, Texture Features, Hierarchical Discriminant Regression Tree (HDR Tree), Two-dimensional Principal Component Analysis (2D-PCA)...