“Gongzhu” is a card game popular in Chinese circles at home and abroad, which belongs to incomplete information game. The game process is highly reversible and has complex state space and action space. This paper proposes an algorithm that combines the Monte-Carlo (MC) method with deep neural networks, called the Deep Monte-Carlo (DMC) algorithm. Different from the traditional MC algorithm, this algorithm uses a Deep Q-Network (DQN) instead of the Q-table to update the Q-value and uses a distributed parallel training framework to build the model, which can effectively solve the problems of computational complexity and limited resources. After 24 h of training on a server with 1 GPU, the “Gongzhu” agent performed 10,000 games against the agent that uses a Convolutional Neural Network (CNN) to fit the strategies of human players. “Gongzhu” agent was able to achieve a 72.6% winning rate, and the average points per game was 63. The experimental results show that the model has better performance.
In view of the difficulty to realize the meteorological and hydrological observation on the unmanned islands or reefs, the complexity of using wireless sensor network cannot be too high and the requirement of reliability is higher, so a master-slave dual-cluster head network trust model based on adaptive weighted D-S evidence theory fusion calculation is proposed. At the same time, the reliability is adaptively adjusted according to the time sliding statistics of the node measurement data. The base station determines whether to receive data and determines the master-slave cluster head in the next cycle according to the fusion result difference and their historical data evaluation, which reduces the complexity of trust calculation and develops the remote unattended Islands and reefs meteorological and hydrological wireless sensor monitoring technology. The test results show that the technology is effective and feasible, with the advantages of convenient layout and reliable data transmission.
Personal record keeping, behaviour tracking, accurate feeding, disease prevention and control, and food traceability require the identification of dairy cows. This work proposes a unique identification that combines Mask R-CNN and ResNet101 to identify individual cows in milking parlours accurately. Using 265 Holstein cows in various positions, a facial image dataset of their faces was created using the milking hall’s webcam. The feature pyramid network-based Mask R-CNN instance segmentation model was trained to separate the cows’ faces from their backgrounds. The ResNet101 individual classification network was trained using the segmentation above data as input and cow individual numbers as output. Combining the two techniques led to the creation of the cow individual recognition model. According to experimental findings, the Mask R-CNN model has an average accuracy of 96.37% on the picture test set. The accuracy of the Resnet101-based individual classification network was 99.61% on the training set and 98.75% on the validation set, surpassing that of VGG16, GoogLeNet, ResNet34, and other networks. The study’s suggested individual recognition model outperformed the combined effect of the YOLO series model and ResNet101 in terms of test accuracy (97.58%). Furthermore, it outperforms the pairing of Mask R-CNN with VGG16, GoogleNet, and ResNet34. This research offers precision dairy farming an excellent technical basis for individual recognition.
As an important technique of driverless vehicles, 3D Point Cloud Object Detection algorithm can provide semantic information and geometrical information of various types of objects. For the problem of feature information loss in 3D Point Cloud Object Detection, this paper proposes an improved Pointpillars algorithm where a multi-scale columnar feature extraction network based on the attention mechanism. In this algorithm we extract features to obtain pseudo-images of different scales, and splicing point cloud pseudo-images of multiple scales to obtain fusion feature maps. In addition, by introducing a Convolutional Block Attention Module (CBAM), our MSCS-Piontpillars can effectively suppress the noise in the pseudo-image of the point cloud and amplify the important feature information for target classification. Comparision experimental was carried on both the KITTI 3D Object dataset and a real experiment. The results show that compared with Pointpillars, MSCS-Pointpillars (Multi-Scale Channel Spatial Attention Pointpillars) have improved the detection accuracy of cars, cyclists, pedestrians and other targets.
Due to the global COVID-19 pandemic, there is a strong demand for pharyngeal swab sampling and nucleic acid testing. Research has shown that the positive rate of nasopharyngeal swabs is higher than that of oropharyngeal swabs. However, because of the high complexity and visual obscuring of the interior nasal cavity, it is impossible to obtain the sampling path information directly from the conventional imaging principle. Through the combination of anatomical geometry and spatial visual features, in this paper, we present a new approach to generate nasopharyngeal swabs sampling path. Firstly, this paper adopts an RGB-D camera to identify and locate the subject’s facial landmarks. Secondly, the mid-sagittal plane of the subject’s head is fitted according to these landmarks. At last, the path of the nasopharyngeal swab movement in the nasal cavity is determined by anatomical geometry features of the nose. In order to verify the validity of the method, the location accuracy of the facial landmarks and the fitting accuracy of mid-sagittal plane of the head are verified. Experiments demonstrate that this method provides a feasible solution with high efficiency, safety and accuracy. Besides, it can solve the problem that the nasopharyngeal robot cannot generate path based on traditional imaging principles. It also provides a key method for automatic and intelligent sampling of nasopharyngeal swabs, and it is of great clinical value to reduce the risk of cross-infection.
