Small object detection in UAV aerial images based on deep learning has a wide range of applications in military intelligence reconnaissance, battlefield surveillanceand assessment, military object captureand verification, intelligent traffic management, infrastructure inspectionand maintenance, disaster preventionand control, searchand rescue, crop managementand analysis, ecological protectionand monitoringand other fields, and has become a current research hotspot in recent years. This review article gives a comprehensiveand in-depth investigation on small object detection in UAV aerial images based on deep learning in the past five years. First of all, the definitionand challenges of small object detection in UAV aerial images are introduced. Secondly, small object detection methods in drone aerial images are summarized in terms of discriminative feature learning, super-resolution technology, real-time lightweight detection, and other improvement ideas. Then, small object detection datasets of UAV aerial images are systematically summarized, and the performances of different algorithms are analyzed based on the VisDrone Challenge. Finally, the specific applications of small object detection in UAV aerial images in the militaryand civilian fields are comprehensively presented, and the potential future development directions of small object detection in UAV aerial imagesand some concerns about UAV aerial photography are also discussed. It is expected that this review would inspire relevant researchers to further promote the development of small object detection in UAV aerial images based on deep learning.

Unmanned Aerial Vehicle (UAV) plays an important role in military reconnaissance, agricultural monitoring, urban planning and management due to its advantages of light weight, flexibility, wide coverage and low cost. In these applications, due to the limited field of view of UAV aerial photography, it is impossible to obtain high-definition panoramic images of the complete target area. Therefore, image mosaic technology is essential. In recent years, the development of deep learning has brought renewed attention to UAV aerial image mosaic technology. This paper reviews the research progress of UAV aerial image mosaic methods in the last 10 years. Firstly, the development process and main process of UAV aerial image mosaic method are introduced. Secondly, the traditional methods of UAV aerial image mosaic are described in detail according to the three steps of image preprocessing, image registration and image fusion. The advantages and disadvantages of the methods used in each step are also compared. This paper then describes in detail the UAV aerial image mosaic method based on deep learning from three aspects: semantic segmentation, image registration, and image mosaic framework which both based on deep learning. Common UAV aerial image mosaic datasets and image mosaic performance evaluation indicators are sorted out, and five typical application fields of UAV aerial image mosaic technology are listed. Finally, this paper highlights several technical challenges that UAV aerial image mosaic still faces and provides an outlook on the future work directions.

The you only look once (YOLO) series of deep learning-based object identification algorithms has several advantages, including great adaptability, high detection accuracy, and quick detection speed. Make it the preferred choice for object detection in drone aerial images. Due to the complex and ever-changing nature of aerial images and videos captured by drones, the YOLO series object detection algorithms from the UAV vision have become a challenging problem. Therefore, it is necessary to comprehensively summarize and analyze the recent research and application achievements of YOLO series algorithms for object detection from the UAV vision. Firstly, the YOLO structural diagram is redrawn and the development, connection, and optimization features of the 10 fundamental versions of YOLO are summarized. The roles and objectives of various YOLO optimization strategies are pointed out, indicating the direction for further theoretical research. Secondly, the optimization methods of YOLO in the application field of drone object detection are summarized. This includes preprocessing before data training, optimization of training and inference, application of attention modules, convolutional optimization, detection heads, and optimization of loss functions. At the same time, some application datasets and data expansion methods from the UAV vision are sorted out. Through studies, the effectiveness of YOLO series algorithms on the UAV vision dataset and the influence of convolutional block attention module (CBAM) on various YOLO improvements are assessed. Finally, it is pointed out that there are still some problems and challenges in the research of YOLO series algorithms for object detection from the UAV vision. This article looks forward to possible development trends and research directions in this field.

No matter in the field of military investigation or maritime law enforcement, the ship target detection and identification technology is very important. Especially with the development of optical remote sensing satellite, a large number of ship imaging data have been obtained, and how to quickly and accurately locate and identify ship targets from a large number of optical imaging data is a challenging work. Firstly, this paper summarizes the development process and technical process of ship target detection and recognition technology in optical remote sensing images. Then, the research progress in the acquisition and preprocessing of optical remote sensing images, the separation of land and sea, the detection of ship targets and the recognition of ship targets are reviewed, and the methods and progress of target detection and recognition of optical remote sensing images using traditional and deep learning methods are emphatically discussed. Then, 11 kinds of remote sensing image data sets including ship targets and performance evaluation indexes are introduced. Finally, the main problems in ship target detection and recognition technology are analyzed, and the future development direction of ship target detection and recognition technology is given.

Unmanned Aerial Vehicles (UAV) have been widely used due to their advantages of small size, light weight, and simple operation. The combination of the deep learning method and UAV system can help to quickly and accurately detect the desired target on UAV aerial images with high definition and a wide field of view. Related topics have become one of the current research hotspots. This paper reviews the research progress of deep learning methods for object detection and semantic segmentation in UAV aerial images in the past ten years. First, the characteristics and wide range of application scenarios of UAVs and their aerial images are summarized, and the development process of target detection and semantic segmentation methods for UAV aerial images is briefly described. Then, the object detection and semantic segmentation methods of UAV aerial images based on deep learning are classified according to different network models, and their improvement strategies, application scenarios, contributions, and limitations are summarized. Subsequently, the data sets of aerial images taken by UAVs in recent years are collected and sorted out, and the evaluation indicators of commonly used convolutional neural network models are summarized. Finally, the existing problems in this field are pointed out, and prospects for future research trends are presented.