Research On Weak Sea-Surface Target Detection Based On Deep Learning

Sea-surface target detection has direct and vital applications in both the military and civilian fields,such as resource exploration,ship reconnaissance,maritime rescue,naval warning and meteorological research.When detecting,there are several types of targets referred to as weak targets,i.e.,the small-scale targets,stealthy targets,high-maneuvering targets,and distant targets,due to the fact that their reflected signals are weak,such as the sea skimmers,periscopes,and stealth speedboats in the field of military,and the victims,floating wreckage and damaged fishing boats in the field of civil.However,the weak targets floating on the sea surface are often of low observability,and thus how to accurately identify them from sea clutter has become an important scientific problem to be solved.In this paper,we consider constructing a discriminative 2-D image,as by both the temporal and spatial(i.e.,range dimension)information can be taken into account,and then adopt the convolutional neural network(CNN)to design a constant false alarm rate(CFAR)detector to address the hign-precision object detection problem.Specifically,the main contents and contributions are summarized as follows.1.We propose a scheme for detecting weak targets on the seasurface from sea clutter based on 2-D images containing both the temporal and spatial feature.By analyzing the temporal and spatial correlation characteristics of the reflected signals,it is found that both the temporal information under a certain observation time and the spatial information hidden between different range cells are beneficial to target detection.Therefore,we choose to combine the two kinds of information and convert the returned signals into 2-D images to fully describe the differences between target signals and clutter signals.We extensively investigate three different preporities of the returned signals,i.e.,the amplitude,Dopplerspectrum,and short-time Fourier transform,from the time,frequency,and time-frequency domains,and rhen convert them into images.Interestingly,during the research,it is found that compared with the original features,the constucted images are less sensitive to the sea spike effect and could clearly highlight the position of the target signal under low sea states.2.We evaluate and select the most distinguishable 2-D images which characterize the difference between target and clutter signals best.The detection performance of the CNN is highly dependent on the discriminability between the input target image and the clutter image itself.However,the ability of different image construction method to depict the discriminability is different.Therefore,we define the concept of image distinctiveness,and adopt histogram matching method to calculate the image distinctiveness,so as to select the most excellent image constrction method.By calculating the Bhattacharyya distance between the target image and the clutter image histogram,the performance of the proposed amplitude-based,Doppler-spectrum-based,and short-time-Fourier-transform-based images to describe the difference between the target and the clutter is compared.It is found that the amplitude-based image exhibits the best performance.3.We design a constant false alarm rate detector based on the CNN for identifying the target images.The detection of weak targets on the sea surface can be considered as a binary classification problem based on the constructed 2-D images converted from the returned signals.As a classical image classification method in deep learning,the CNN exhibit excellent performance on many issues,and thus it is a potential solution to apply it in the problem of sea-surface weak targets detection.However,research from this aspect has so far been very rare,partly limited by the requirement of the CFAR detection.To address this issue,we in this paper modify the CNN to accept the FAR as an input and then adjust the classification threshold in the softmax layer,and thus our proposed CNN-based detector can work in the given FAR to meet the need of practical applications and evaluations.This paper designs a CNN-based detector to identify the target image in the constructed amplitude-based images converted from the radar echo signals.The performance of the CNN-based detector is evaluated from the three perspectives of the average signal-to-clutter ratio,false alarm rate(FAR),and observation time,which are the most concerned by the actual demand,and compared with several classical detectors.Experimental results demonstrate that the detector designed in this paper significantly improves the probability of the weak targets detection problem.