The application of face detection and recognition technology in security monitoring systems has made a huge contribution to public security. Face detection is an essential first step in many face analysis systems. In complex scenes, the accuracy of face detection would be limited because of the missing and false detection of small faces, due to image quality, face scale, light, and other factors. In this paper, a two-level face detection model called SR-YOLOv5 is proposed to address some problems of dense small faces in actual scenarios. The research first optimized the backbone and loss function of YOLOv5, which is aimed at achieving better performance in terms of mean average precision (mAP) and speed. Then, to improve face detection in blurred scenes or low-resolution situations, we integrated image superresolution technology on the detection head. In addition, some representative deep-learning algorithm based on face detection is discussed by grouping them into a few major categories, and the popular face detection benchmarks are enumerated in detail. Finally, the wider face dataset is used to train and test the SR-YOLOv5 model. Compared with multitask convolutional neural network (MTCNN), Contextual Multi-Scale Region-based CNN (CMS-RCNN), Finding Tiny Faces (HR), Single Shot Scale-invariant Face Detector (S3FD), and TinaFace algorithms, it is verified that the proposed model has higher detection precision, which is 0.7%, 0.6%, and 2.9% higher than the top one. SR-YOLOv5 can effectively use face information to accurately detect hard-to-detect face targets in complex scenes.

The collision avoidance in real-time is crucial for unmanned surface vehicles (USVs) in a complex environment. Traditional methods make it hard to ensure the balance of control decisions. To balance safety and practicality, a collision avoidance algorithm based on deep reinforcement learning (DRL) and a two-level incentive reward based on the principle of complementarity is proposed. To address the vital sparse reward problem of Deep Deterministic Policy Gradient (DDPG), the trajectory evaluation function of the dynamic window algorithm (DWA) is referred to construct the primary reward strategy, and a secondary incentive reward is constructed based on velocity obstacle (VO) to eliminate potential collision risks. To improve the efficiency of training, the electronic chart (EC) and Unity3D are used to build an immersive simulation platform. Based on it, simulations are made to verify the performance. In addition, field experiments are first conducted in various encounter scenarios to verify the effectiveness. The results show that it can take safe collision avoidance actions and get practical paths in various situations.

Objective: We propose a deep-learning-based underwater target detection system that can effectively solve the problem of underwater optical image target detection and recognition. Methods: In this paper, based on the depth of the underwater optical image target detection and recognition and using a learning model, we put forward corresponding solutions using the concept of style migration solutions, such as training samples. A lack of variability and poor generalization of practical applications presents a challenge for underwater object identification. The UW_YOLOv3 lightweight model was proposed to solve the problems of calculating energy consumption and storage resource limitations in underwater application scenarios. The detection and recognition module, based on deep learning, can deal with the degradation process of underwater imaging by embedding an image enhancement module into the detection and recognition module for the joint tuning and transferring of knowledge. Results: The detection accuracy of the UW_YOLOv3 model designed in this paper outperformed the lightweight algorithm YOLOV3-TINY by 7.9% at the same image scale input. Compared with other large algorithms, the detection accuracy was lower, but the detection speed was much higher. Compared with the SSD algorithm, the detection accuracy was only 4.7 lower; the speed was 40.9 FPS higher; and the rate was nearly 16 times higher than Faster R-CNN. When the input scale was 224, although part of the accuracy was lost, the detection speed doubled, reaching 156.9 FPS. Conclusion: Based on our framework, the problem of underwater optical image target detection and recognition can be effectively solved. Relevant studies have not only enriched the theory of target detection and glory, but have also provided optical glasses with a clear vision for appropriate underwater application systems.

With the development of remote-sensing technology, it is important to use semantic segmentation methods to obtain detailed information in remote-sensing images. However, the objects in the images reveal significant intra-class differences and slight inter-class differences, thus affecting the acquisition of terrain information. To tackle the problems, this paper proposes an improved U-Net for the semantic segmentation of remote-sensing images. Firstly, the local importance-based pooling is introduced to alleviate the loss of feature details in the coding part. Secondly, a combined attention module with a double-branch structure is designed, which models the local relationship and the global relationship at the same time to obtain more typical features. Finally, in order to make full use of the feature information extracted from the coding part, the combined attention module and the channel attention module are added to different positions in U-Net. In order to validate the proposed method, we conduct experiments on the WHDLD dataset and compare the experimental results with other semantic segmentation methods. On the WHDLD dataset, the MPA, MIOU, and FWIOU of the proposed method reach 76%, 64.11%, and 75.64%, respectively, revealing its priority. To demonstrate the generalization of the proposed method, generalization experiments are conducted via the LandCover.ai dataset and the Massachusetts-building dataset. The simulation results testify that the proposed method provides an excellent generalization ability.