A Deep Neural Network-Based Method for Detecting Aircraft in High-Resolution Radar Images with Limited Training Data

This research work comes up with a deep neural network-based method for detecting aircraft and ships in radar imagery, even with limited training data. The proposed method comprises four major steps: development of a reference network by using the High Resolution SAR Images Dataset (HRSID) for detecting vessels, changing the parameters and hyperparameters of the network, adding the SAR Aircraft Detection Dataset (SADD) with very high-resolution radar images to the SADD, and performing the aircraft detection network through transfer learning. It is worth mentioning that the method has delivered an average precision of around 80% in the ship as well as the plane detection. In order to overcome the challenge of less available training data, the function that performs the layer-by-layer freezing is applied. It helps to find the ideal depth for transference of features between the detection tasks. Additionally, the step-wise feature transitive transfer approach was applied, where multiple source tasks and datasets were brought into cooperation to gain further improvement in model performance on the target task. The result of the research points out that transferring features from different source datasets improves detection accuracy by a large margin. The research carried out provided evidence for the effectiveness of radar data, especially that of SAR imagery from missions such as ICEYE and Capella, in automatic reconnaissance. The trust of SAR systems, which work effectively during the night and in difficult meteorological conditions, in turn, boosts the method's deployability for robust and continuous object detection.