Yu Xia
ENHANCING THE CLARITY OF BLURRY DRONE IMAGES OF VESSELS AT SEA *
Аннотация:
to address the trade-off between real-time performance and image clarity in UAV-based maritime ship target acquisition, this paper proposes an image compression and blur reconstruction method. The approach leverages an improved YOLOv8 detection model and the Real-ESRGAN network, incorporating dataset construction, network training and tuning, and deployment to achieve real-time reconstruction of high-quality ship target images on the ground under limited bandwidth and computational resources. First, the improved YOLOv8 model is employed for precise detection and localization of ship targets in the images. Then, the Real-ESRGAN network is utilized to reconstruct compressed and blurry images, restoring high-resolution details. Experimental results demonstrate that the proposed method significantly enhances image clarity and detection accuracy while substantially reducing bandwidth consumption. It meets the high real-time requirements of UAV-based ship recognition and performs exceptionally well in resource-constrained environments. This work provides an effective solution for UAV-based maritime ship monitoring, improving UAV surveillance and recognition capabilities and laying a foundation for the broader application of UAVs in ocean monitoring.
Ключевые слова:
drone imaging, naval ships, two-way feature integration model, Real-ESRGAN system, optimized YOLOv8 recognition model, oceanic vessel surveillance
DOI 10.24412/2712-8849-2025-384-437-448
1. Introduction. Unmanned aerial vehicles (UAVs) play a crucial role in both civilian and military applications. To effectively support the task of maritime vessel detection and recognition, UAVs are required to capture high-resolution images of ship targets from as great a distance as possible [1]. However, due to communication bandwidth limitations, the ground station can only receive compressed and degraded images [2], making it impossible to obtain high-quality, lossless imagery. This significantly impacts the accuracy of image analysis and interpretation at the ground station. As maritime target detection and recognition are key applications of UAVs, there is an urgent need for effective reconstruction of compressed and blurred UAV-captured images of ships at sea.In recent years, deep learning techniques have made remarkable progress in the field of computer vision, particularly in the challenging task of image super-resolution reconstruction. Super-resolution reconstruction aims to recover high-resolution images from low-resolution inputs, a process that is crucial for improving image quality, enhancing visual experience, and meeting the growing demands for image processing. Deep learning algorithms have played a significant role in this domain due to their powerful feature learning capabilities. The SRCNN algorithm, proposed in references [4-5], is one of the early representations of applying deep learning to super-resolution reconstruction. SRCNN effectively learns the mapping from low-resolution to high-resolution images using deep convolutional networks and spatial pyramid pooling techniques. Further advancing this research, the VDSR algorithm, introduced in reference [7], utilizes a very deep network structure that can restore more details and texture information in image reconstruction by learning from a large number of samples. The EDSR algorithm, proposed in reference [9], represents further innovation in network structure. EDSR employs an efficient residual network design, maintaining gradient effectiveness even in deeper networks, enabling higher-quality image reconstruction. The GAN algorithm introduced in reference [11] has revolutionized the image generation field and has also been applied to super-resolution reconstruction. GAN generates highly realistic high-resolution images by training an adversarial network between a generator and a discriminator.Since paired high and low-resolution images are rare in the real world, existing methods primarily generate low-resolution images by degrading high-resolution images, forming datasets by using original and degraded images for training. The low-resolution images in the datasets are degraded from high-resolution images using methods such as blurring, downsampling, noise, and JPEG compression. However, real-world low-resolution image degradation is more complex and diverse, and simple degradation combinations fail to simulate real data, resulting in poor generalization of the trained model. UAVs can capture real-time low-resolution compressed images during ocean exploration flights and can also retrieve high-resolution lossless images afterward. Therefore, both low-resolution compressed images and high-resolution lossless images can be combined to form a dataset, improving the models generalization performance. Due to the differences in dataset construction methods, existing methods are difficult to directly apply to UAV image super-resolution reconstruction and require optimization and improvement.This paper proposes a super-resolution reconstruction technique suitable for UAV-based maritime target detection. First, the improved YOLOv8 object detection network is used to quickly screen the original images containing target objects, accurately selecting all image frames with targets. After manual verification, a maritime vessel target dataset is formed. Subsequently, the super-resolution Real-ESRGAN network is trained using the maritime vessel target dataset, converting low-resolution images into high-definition super-resolution images. Finally, the super-resolution reconstruction network, consisting of YOLOv8 and Real-ESRGAN, is deployed to the UAVs ground control station, enabling real-time conversion of compressed images to high-resolution images.2. Construction of the Maritime Vessel Target Dataset.The UAV performs flight missions and captures real-time images of ground targets. Multiple mission datasets are collected, including low-resolution compressed images X transmitted in real-time by the UAV and high-resolution lossless images Z offloaded at the end of the flight. The UAV video data is processed for target detection using the improved YOLOv8 detection model to construct the maritime vessel target dataset. The entire network consists of three components: feature extraction (Backbone), feature fusion (Neck), and detection head (Head). On the basis of YOLOv8, bidirectional feature fusion and attention mechanisms are employed for feature integration.2.1. Feature Extraction Network.The feature extraction network (Backbone) utilizes the Darknet-53 model [13], a network structure designed in [14] for YOLO-based object detection systems. Darknet-53 emphasizes balancing speed and performance. Compared to VGG16, it has fewer layers, but each convolutional layer generally uses larger kernels (e.g., 7×7 or 5×5), and residual connections (Residual Connection) are added between some layers to facilitate better information flow across layers. Detailed network structure is shown on the right side of Figure 1, comprising Conv convolution modules and Residual Blocks, serially stacked four times. Research has shown that Darknet-53, due to its fewer parameters and higher computational efficiency, is suitable for real-time processing of large image datasets in object detection tasks.Fig. 1. Overall Framework of the Improved YOLOv8 Detection Model.2.2. Feature Fusion Network.The feature fusion network (Neck) adopts a bidirectional feature fusion model. CNNs employ a hierarchical method in the feature extraction process. Generally, as the network depth increases, the receptive field of each feature point also increases, allowing the extraction of higher-level abstract features and richer semantic information. Conversely, shallow features focus more on capturing simple details such as contours and textures. To overcome the semantic limitations of low-level features, a bidirectional feature fusion module is employed, effectively supplementing information from high-level features with that from low-level features through a specialized aggregation and reuse mechanism.The fusion process is illustrated in Figure 2, where the semantic information from high-level features is effectively integrated with the detailed information from low-level features to improve the accuracy and efficiency of target detection.Fig. 2. Flowchart of the Bidirectional Feature Fusion Module.The attention mechanism optimizes feature representation by weighting the importance of features. This paper uses two attention mechanisms: channel attention mechanism and spatial attention mechanism. The integrated processing flow for these mechanisms is shown in Figure 3.Fig. 3. Flowchart of the Bidirectional Feature Fusion Module.3. Image Reconstruction Based on Real-ESRGAN.3.1. Generative Network.The generative network of the Real-ESRGAN model is built upon the ESRGAN framework. It processes low-resolution, compressed images and generates high-resolution images with a fourfold increase in resolution. For 2× and 1× upscaling, pixel-unshuffle is employed to downscale the image while increasing the number of channels, redistributing original pixels into a lower-resolution but channel-enriched version. This processed image is then fed into the generative network, which outputs an upscaled image with 2× or 1× resolution (as illustrated in Figure 4).Fig. 4. Generative Network.3.2. Discriminative Network.The discriminative network in Real-ESRGAN adopts a U-Net architecture, composed of an encoder-downsampling section and a decoder-upsampling section. It is designed to differentiate between generated and real images. The input consists of either the generated image
Номер журнала Вестник науки №3 (84) том 2
Ссылка для цитирования:
Yu Xia ENHANCING THE CLARITY OF BLURRY DRONE IMAGES OF VESSELS AT SEA // Вестник науки №3 (84) том 2. С. 437 - 448. 2025 г. ISSN 2712-8849 // Электронный ресурс: https://www.вестник-науки.рф/article/21808 (дата обращения: 13.12.2025 г.)
Вестник науки © 2025. 16+