Although the Neural Radiance Fields (NeRF) has been shown to achieve high-quality novel view synthesis, existing models still perform poorly in some scenarios, particularly unbounded scenes. These models either require excessively long training times or produce suboptimal synthesis results. Consequently, we propose SD-NeRF, which consists of a compact neural radiance field model and self-supervised depth regularization. Experimental results demonstrate that SDNeRF can shorten training time by over 20 times compared to Mip-NeRF360 without compromising reconstruction accuracy.
The current convolution-based semantic segmentation network lacks long-range dependencies for infrared small target detection, which may lead to unsatisfactory detection results in the real scenario. To address the problem, this paper proposed a semantic segmentation network based on hetero-range feature fusion (HRFFnet). Compared with the common semantic segmentation networks, this network includes two feature extraction branches. One is a short-range extraction branch consisting of convolution operations, and the other is a long-range feature extraction branch consisting of transformer. The HRFFnet complements the local features extracted by the convolutional neural network by adding the transformer structure to the segmentation network to introduce the long-range information of image. And this paper also designed a hetero-range fusion module. This module is based on atrous spatial pyramid pooling and adds shortcut connection to fuse different ranges of features extracted from images, which can maintain resolution of image and improve ability of feature representation. The hetero-range fusion module fuses long-range dependencies and short-range information extracted by transformer and convolution to capture multi-scale contextual information about the scene in infrared images and facilitate the interchange of information between different features. To evaluate the HRFFnet, we compare the performance of our network against other high-performance convolution-based methods and transformer-based networks on the open SIRST dataset with different evaluation metrics. The proposed method achieves a better combined results in terms of intersection of Dice coefficient, pixel accuracy, intersection over union and receiver operating characteristic curve. The experiments and results show that the network has some superiority: one is that it can break through the limitation of range of extracting features when only using convolutional network or transformer-based network; the other one is that this network can perform better with good robustness against real and complex scenarios. So, the proposed algorithm has broad application prospects in border patrol and urban security fields.
With the development of science and technology, unmanned aerial vehicle (UAV) is more and more widely used to bring a lot of convenience to the society, but also led to serious threats to public security, personal privacy, military security and other aspects. Therefore, it is increasingly important to find unknown drones quickly and accurately. In UAV detection, the technologies based on acoustic, radio and radar detection are common, but these technologies usually require expensive equipment and strict configuration. However, the method based on machine vision has the advantages of low cost and simple configuration. In addition, detection and recognition methods based on deep learning have been fully developed, but most of them are for a single visible image, and the detection and recognition effect is limited. In this paper, a fast detection and identification method based is proposed based on the backbone of YOLOv3 (You Only Look Once version3). And dual-channel detectors were used as data sources. In this method, infrared and visible images are simultaneously input into the network for feature extraction, and the extracted depth features are concatenated. Then the multi-scale prediction network is used to regression the target location to obtain the final detection and recognition results. Finally, by collecting real UAV data sets, the network is trained and tested for comparative experiments. Experimental results show that the mAP of method in this paper is worthy of improvement, and the detection speed remains at 27images/s.
In recent years, with the applications of object detection increasingly extensive, the approaches based on Deep Learning have achieved state-of-the-art performance on challenging datasets. Some researchers have made demands on real-time performance while paying attention to the accuracy of the model. In addition, with the rapid development of the object detection model, the detection of small targets has attracted extensive attention. Although several evaluations of the models have been conducted, we have conducted a more detailed evaluation of the small targets real-time detection. In this work, we carried out an in-depth evaluation of the latest real-time object detection model. We evaluate three state-of-the- art models including Single Shot MultiBox Detector (SSD), You Only Look Once version 2 (YOLO v2), and You Only Look Once version 3 (YOLO v3) with related trade-off factors i.e. accuracy, execution time and resource usage. Experiments were conducted on benchmark datasets and a newly generated dataset for small object detection. All analyses and findings are then presented.
In order to make the system design meet the requirements of practical ghost imaging, the impact of mechanical vibration on the ghost imaging is analyzed. In ghost imaging system, the light field modulated by a digital micromirror device (DMD) is used to illuminate the target and the transmission or scattering light is detected by a single pixel detector. The target is reconstructed by combining the results of the detector and the intensity distribution of light field, so the modulation matrices of light field play a vital role in ghost imaging. By considering the form of imaging system to vibration and taking the modulation transfer function as an evaluation function, this paper quantitatively analyzes the impacts of various forms of mechanical vibration on the intensity distribution of light field. Combining engineering practice, several solutions are proposed to reduce the impact of vibration on the imaging quality. The results of simulation and experiment indicated that the analysis is correct and usable.
Ghost imaging is an indirect system that allows the imaging of an object without directly seeing the object. The speckle pattern that contains the information about light and objects has increasingly become a popular topic in pseudothermal light ghost imaging. However, existing research still has encountered problems of poor imaging quality and slow sampling speeds. We propose a ghost imaging method based on N-order speckle patterns to recover the object (NSGI). The N-order speckle patterns combine N independent laser speckles individually produced by passing an expanded and collimated He–Ne laser through a digital micromirror device (DMD). The sampling frequency can be improved by controlling the trigger signals of different DMDs. The results of the simulation and experiment have verified that our method can increase sampling speed and reconstruction accuracy. In addition, NSGI can be applied to more applications by designing multiple independent speckles with different properties.
For visual tracking with UAV, the non-rigid body change of target usually results in the accumulation of errors and decline of tracking precision. In view of this problem, a target regression tracking algorithm based on convolutional neural network is proposed. Firstly, we use the Siamese convolutional neural network to extract features which used as the input of tracker based on self-adapted scale kernel correlation filters. Then, in order to cope with the cumulative errors caused by the change of target form, a target regression network is designed to refine the location. Using the refined location to extract sample and update the filter parameters of tracker can prevent tracker from being polluted. The experimental results show that the algorithm has high tracking precision as well as fast speed compared to the state-of-the-art tracking algorithms, especially with the ability to deal with the non-rigid body change of target.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.