As a data-driven approach, deep learning requires a large amount of annotated data for training to obtain a sufficiently accurate and generalized model, especially in the field of computer vision. However, when compared with generic object recognition datasets, aerial image datasets are more challenging to acquire and more expensive to label. Obtaining a large amount of high-quality aerial image data for object recognition and image understanding is an urgent problem. Existing studies show that synthetic data can effectively reduce the amount of training data required. Therefore, in this paper, we propose the first synthetic aerial image dataset for ship recognition, called UnityShip. This dataset contains over 100,000 synthetic images and 194,054 ship instances, including 79 different ship models in ten categories and six different large virtual scenes with different time periods, weather environments, and altitudes. The annotations include environmental information, instance-level horizontal bounding boxes, oriented bounding boxes, and the type and ID of each ship. This provides the basis for object detection, oriented object detection, fine-grained recognition, and scene recognition. To investigate the applications of UnityShip, the synthetic data were validated for model pre-training and data augmentation using three different object detection algorithms and six existing real-world ship detection datasets. Our experimental results show that for small-sized and medium-sized real-world datasets, the synthetic data achieve an improvement in model pre-training and data augmentation, showing the value and potential of synthetic data in aerial image recognition and understanding tasks.

The method of collecting aerial images or videos by unmanned aerial vehicles (UAVs) for object search has the advantages of high flexibility and low cost, and has been widely used in various fields, such as pipeline inspection, disaster rescue, and forest fire prevention. However, in the case of object search in a wide area, the scanning efficiency and real-time performance of UAV are often difficult to satisfy at the same time, which may lead to missing the best time to perform the task. In this paper, we design a wide-area and real-time object search system of UAV based on deep learning for this problem. The system first solves the problem of area scanning efficiency by controlling the high-resolution camera in order to collect aerial images with a large field of view. For real-time requirements, we adopted three strategies to accelerate the system, as follows: design a parallel system, simplify the object detection algorithm, and use TensorRT on the edge device to optimize the object detection model. We selected the NVIDIA Jetson AGX Xavier edge device as the central processor and verified the feasibility and practicability of the system through the actual application of suspicious vehicle search in the grazing area of the prairie. Experiments have proved that the parallel design of the system can effectively meet the real-time requirements. For the most time-consuming image object detection link, with a slight loss of precision, most algorithms can reach the 400% inference speed of the benchmark in total, after algorithm simplification, and corresponding model’s deployment by TensorRT.

Object detection in computer vision requires a sufficient amount of training data to produce an accurate and general model. However, aerial images are difficult to acquire, so the collection of aerial image datasets is a priority issue. Building on the existing research on image generation, the goal of this work is to create synthetic aerial image datasets that can be used to solve the problem of insufficient data. We generated three independent datasets for ship detection using engine and generative model. These synthetic datasets are rich in virtual scenes, ship categories, weather conditions, and other features. Moreover, we implemented domain-adaptive algorithms to address the issue of domain shift from synthetic data to real data. To investigate the application of synthetic datasets, we validated the synthetic data using six different object detection algorithms and three existing real-world, ship detection datasets. The experimental results demonstrate that the methods for generating synthetic aerial image datasets can complete the insufficient data in aerial remote sensing. Additionally, domain-adaptive algorithms could further mitigate the discrepancy from synthetic data to real data, highlighting the potential and value of synthetic data in aerial image recognition and comprehension tasks in the real world.