SHIP DETECTION IN SATELLITE IMAGERY VIA CONVOLUTIONAL NEURAL NETWORKS

摘要

Ship detection from satellite imagery is a valuable tool for maritime traffic surveillance, detecting illegal fishing, oil discharge control, and sea pollution monitoring. Over the last decade, deep learning using convolutional neural networks (CNNs) has become the dominant paradigm for various computer vision tasks such as image classification, object detection, and segmentation, far surpassing traditional techniques. However, research and development are often focused on popular large-scale datascts, consisting of images that are vastly different from satellite imagery. We therefore trained and evaluated the Single-Shot MultiBox Detector (SSD) model for detection of ships in satellite imagery on a small-scale custom dataset. which consisted of 794 images of ships collected near harbors and coastlines. Two slightly different approaches were trialed for comparison: (1) Only the base architecture layers from the VGG-16 network were initialised with pre-trained weights produced from the ImageNet dataset and then the subsequent layers trained in a typical transfer learning setting. (2; All layers of the network were initialised with pre-trained weights produced from the Microsoft COCO dataset and then fine-tuned by retraining every layer with our dataset. The evaluation algorithms provided by the Pascal VOC competition were used to produce precision-recall (PR) curves and determine the overall mean average precision (mAP). Both approaches achieved competitive results in mAP. However the latter proved slightly more effective in both the rate of training and the final performance. The results confirm the effectiveness of the SSD model for object detection tasks, particularly in the case of satellite imagery. Our results also indicate that training with the SSD model can be greatly improved by fine-tuning pre-trained weights rather than training the entire network from scratch or transfer learning on the later layers of the network.