High-Dimensional Feature Detection of Small Sea-Surface Targets via Improved 1D-AlexNet

‍ ‍Currently， detecting small targets on sea surfaces have become the focus of several researchers， as it is challenging to detect them using marine radars. Because of their low signal-to-clutter ratio （SCR） and weak maneuverability， traditional existing detectors often fail to detect small targets， resulting in severe performance loss， low detection probability， and high false alarm rate. To effectively detect small targets， marine radars often observe interesting areas with increased time observation， thereby improving the SCR of target returns. Full polarization radars with a new radar system can ensure sufficient observation time and further expand the dimension of information. Therefore， a new detection method via a modified 1D-AlexNet model is proposed in this paper， where information from received returns in full polarization radars is fully exploited to comprehensively improve the detection ability of small sea-surface targets. First， twenty-four features are extracted from the time， frequency， time-frequency， and polarization domains， reflecting the obvious differences between sea clutter and returns with target in terms of intensity， fractal， geometry， and physical scattering. Second， all features are combined to construct a high-dimensional feature space， and the traditional binary detection problem can be transformed into a binary classification problem in the high-dimensional feature space. Third， a modified 1D-AlexNet classifier is designed from the perspectives of structure and parameters. At the structural level， the traditional two-dimensional AlexNet model is reduced to one dimension， and the number of layers is finely simplified to accelerate the training speed. At the parameter level， an activation function with adaptive slope adjustment is introduced to ensure the stability of the model in training. Meanwhile， classification accuracy of the model is further improved by replacing the fixed learning rate with the exponential decay function. Finally， by using the open and recognized IPIX measured datasets， experimental results show that the proposed detector can attain optimum performance in comparison with the existing several detectors and can still guarantee robust performance in complicated and various clutter environments. As a result， the improved 1D-AlexNet classifier has a simple structure and fast training speed and can be applied in practical radar rapid detection.