Safety helmet wearing is a powerful guarantee of workers’ personal safety. Aiming at the collected safety helmet wearing pictures have characteristics of high density, small pixels and difficulty to detect, a small object detection algorithm of YOLOv5 (You Only Look Once version 5) for safety helmet was proposed. Firstly, based on YOLOv5 algorithm, the bounding box regression loss function and confidence prediction loss function were optimized to improve the learning effect of the algorithm on the features of dense small objects in training. Secondly, slicing aided fine-tuning and Slicing Aided Hyper Inference (SAHI) were introduced to make the small object produce a larger pixel area by slicing the pictures input into the network, and the effect of network inference and fine-tuning was improved. In the experiments, a dataset containing dense small objects of safety helmets in the industrial scenes was used for training. The experimental results show that compared with the original YOLOv5 algorithm, the improved algorithm can increase the precision by 0.26 percentage points, the recall by 0.38 percentage points. And the mean Average Precision (mAP) of the proposed algorithm reaches 95.77%, which is improved by 0.46 to 13.27 percentage points compared to several algorithms such as the original YOLOv5 algorithm. The results verify that the introduction of slicing aided fine-tuning and SAHI improves the precision and confidence of small object detection and recognition in the dense scenes, reduces the false detection and missed detection cases, and can satisfy the requirements of safety helmet wearing detection effectively.
The target detection model of anesthesia resuscitation is often used to help medical staff to perform resuscitation detection on anesthetized patients. The targets of facial actions during patient resuscitation are small and are not obvious, and the existing Single Shot multibox Detector (SSD) is difficult to accurately detect the facial micro-action features of patients in real time. Aiming at the problem that the original model has low detection speed and is easy to have missed detection, an anesthesia resuscitation object detection method based on improved SSD was proposed. Firstly, the backbone network VGG (Visual Geometry Group)16 of the original SSD was replaced by the lightweight backbone network MobileNetV2, and the standard convolutions were replaced by the depthwise separable convolutions. At the same time, the calculation method of first increasing and then reducing the dimension of the extracted features from patient photos was used to reduce computational cost, thereby improving detection speed of the model. Secondly, the Coordinate Attention (CA) mechanism was integrated into the feature layers with different scales extracted by the SSD, and the ability of the feature map to extract key information was improved by weighting the channel and location information, so that the network positioning and classification performance was optimized. Finally, comparative experiments were carried out on three datasets: CEW(Closed Eyes in the Wild), LFW(Labeled Faces in the Wild), and HAPF(Hospital Anesthesia Patient Facial). Experimental results show that the mean Average Precision (AP) of the proposed model reaches 95.23%, and the detection rate of photos is 24 frames per second, which are 1.39 percentage points higher and 140% higher than those of the original SSD model respectively. Therefore, the improved model has the effect of real-time accurate detection in anesthesia resuscitation detection, and can assist medical staff in resuscitation detection.
Node label is widely existed supervision information in complex networks, and it plays an important role in network representation learning. Based on this fact, a Semi-supervised Representation Learning method combining Graph Auto-Encoder and Clustering (GAECSRL) was proposed. Firstly, the Graph Convolutional Network (GCN) and inner product function were used as the encoder and the decoder respectively, and the graph auto-encoder was constructed to form an information dissemination framework. Then, the k-means clustering module was added to the low-dimensional representation generated by the encoder, so that the training process of the graph auto-encoder and the category classification of the nodes were used to form a self-supervised mechanism. Finally, the category classification of the low-dimensional representation of the network was guided by using the discriminant information of the node labels. The network representation generation, category classification, and the training of the graph auto-encoder were built into a unified optimization model, and an effective network representation result that integrates node label information was obtained. In the simulation experiment, the GAECSRL method was used for node classification and link prediction tasks. Experimental results show that compared with DeepWalk, node2vec, learning Graph Representations with global structural information (GraRep), Structural Deep Network Embedding (SDNE) and Planetoid (Predicting labels and neighbors with embeddings transductively or inductively from data), GAECSRL has the Micro?F1 index increased by 0.9 to 24.46 percentage points, and the Macro?F1 index increased by 0.76 to 24.20 percentage points in the node classification task; in the link prediction task, GAECSRL has the AUC (Area under Curve) index increased by 0.33 to 9.06 percentage points, indicating that the network representation results obtained by GAECSRL effectively improve the performance of node classification and link prediction tasks.
Federated Learning (FL) is a novel privacy?preserving learning paradigm that can keep users' data locally. With the progress of the research on FL, the shortcomings of FL, such as single point of failure and lack of credibility, are gradually gaining attention. In recent years, the blockchain technology originated from Bitcoin has achieved rapid development, which pioneers the construction of decentralized trust and provides a new possibility for the development of FL. The existing research works on blockchain?based FL were reviewed, the frameworks for blockchain?based FL were compared and analyzed. Then, key points of FL solved by the combination of blockchain and FL were discussed. Finally, the application prospects of blockchain?based FL were presented in various fields, such as Internet of Things (IoT), Industrial Internet of Things (IIoT), Internet of Vehicles (IoV) and medical services.
In machine learning, data quality has a far-reaching impact on the accuracy of system prediction. Due to the difficulty of obtaining information and the subjective and limited cognition of human, experts cannot accurately mark all samples. And some probability sampling methods proposed in resent years fail to avoid the problem of unreasonable and subjective sample division by human. To solve this problem, a label noise filtering method based on Dynamic Probability Sampling (DPS) was proposed, which fully considered the differences between samples of each dataset. By counting the frequency of built-in confidence distribution in each interval and analyzing the trend of information entropy of built-in confidence distribution in each interval, the reasonable threshold was determined. Fourteen datasets were selected from UCI classic datasets, and the proposed algorithm was compared with Random Forest (RF), High Agreement Random Forest Filter (HARF), Majority Vote Filter (MVF) and Local Probability Sampling (LPS) methods. Experimental results show that the proposed method shows high ability on both label noise recognition and classification generalization.
Since some computation in reachability Query Preserving Graph Compression (QPGC) algorithm are redundant, a high-performance compression strategy was proposed. In the stage of solving the vertex sets of ancestors and descendants, an algorithm named TSB (Topological Sorting Based algorithm for solving ancestor and descendant sets) was proposed for common graph data. Firstly, the vertices of the graph data were topological sorted. Then, the vertex sets were solved in the order or backward order of the topological sequence, avoiding the redundant computation caused by the ambiguous solution order. And an algorithm based on graph aggregation operation was proposed for graph data with short longest path, namely AGGB (AGGregation Based algorithm for solving ancestor and descendant sets), so the vertex sets were able to be solved in a certain number of aggregation operations. In the stage of solving reachability equivalence class, a Piecewise Statistical Pruning (PSP) algorithm was proposed. Firstly, piecewise statistics of ancestors and descendants sets were obtained and then the statistics were compared to achieve the coarse matching, and some unnecessary fine matches were pruned off. Experimental results show that compared with QPGC algorithm: in the stage of solving the vertex sets of ancestors and descendants, TSB and AGGB algorithm have the performance averagely increased by 94.22% and 90.00% respectively on different datasets; and in the stage of solving the reachability equivalence class, PSP algorithm has the performance increased by more than 70% on most datasets. With the increasing of the dataset, using TSB and AGGB cooperated with PSP has the performance improved by nearly 28 times. Theoretical analysis and simulation results show that the proposed strategy has less redundant computation and faster compression speed compared to QPGC.
The existing similarity search algorithms do not consider the time factor. To address this problem, a meta path-based dynamic similarity search algorithm named PDSim was proposed for the heterogeneous information network. Firstly, PDSim calculated the link matrix of object under the given meta-path, thus obtained the instances ratio of meta-path between different objects. Meanwhile, the differences of establishing time were calculated. Finally, the dynamic similarity was measured under the given meta-path. In multiple instances of the similarity search, PDSim kept up with the interest variation of object which dynamically changed with time. Compared with the PathSim (Meta Path-Based Similarity) and PCRW (Path-Constrained Random Walks) methods, the clustering accuracy of Normalized Mutual Information (NMI) could be increased by 0.17% to 9.24% when applied to clustering. The experimental results show that, compared to the traditional similarity search algorithm based on link, the efficiency of dynamic similarity search and the satisfaction of user of PDSim are significantly improved, and it is a dynamic similarity search algorithm for object changes with time.
Concerning the huge calculation of sparse decomposition, a fast sparse decomposition algorithm with low computation complexity was proposed for first-order Polynomial Phase Signals (PPS). In this algorithm, firstly,two concatenate dictionaries including Df and Dp were constructed, and the atoms in the Df were constructed by the frequency, and the atoms in the Dp were constructed by the phase.Secondly, for the dictionary Df, the group testing was used to search the atoms that matched the signal, and the correlation values of the atoms and the signal were tested twice to achieve the reliability. Finally, according to the matching frequency atoms tested by group testing, the dictionary Dp was constructed, and the matching phase atoms were searched by Matching Pursuit (MP) algorithm. Therefore, the sparse decomposition of real first-order PPS was finished. The simulation results show that the computational efficiency of the proposed algorithm is about 604 times as high as that of matching pursuit and about 139 times as high as that of genetic algorithm, hence the presented algorithm has less computation complexity, and can finish sparse decomposition fast. The complexity of the algorithm is only O(N).
MapReduce is one of the popular distributed computing frameworks based on an open source cloud platform named Hadoop. However, the First-In First-Out (FIFO) scheduling algorithm of MapReduce is inefficient in resources utilization. A new tasks scheduling model based on resources matching rules was proposed and implemented. After obtaining the tasks resources requirement and remainder resources on computing nodes, the model assigned tasks to computing nodes based on resources matching degree to improve the usage efficiency of system resources. First of all, the model for MapReduce scheduling was established, the quantitative definition of resources and matching degree were given, and the corresponding calculation formulas were put forward. Second, the specific methods of resource measurement and the implementation of the algorithm were introduced. Compared with FIFO scheduling algorithm on TeraSort, GrepCount and WordCount, the experimental results show that the proposed model reduces by 22.19% in tasks completion time in the best case, and increases by 25.39% in throughput even in the worst case.