Convolutional Neural Network (CNN) is sensitive to the local features of data due to the complex classification boundaries and too many parameters. As a result, the accuracy of CNN model will decrease significantly when it is attacked by adversarial attacks. However, the Topological Data Analysis (TDA) method pays more attention to the macro features of data, which naturally can resist noise and gradient attacks. Therefore, an image classification algorithm named MCN (Mapper-Combined neural Network) combining topological data analysis and CNN was proposed. Firstly, the Mapper algorithm was used to obtain the Mapper map that described the macro features of the dataset. Each sample point was represented by a new feature using a multi-view Mapper map, and the new feature was represented as a binary vector. Then, the hidden layer feature was enhanced by combining the new feature with the hidden layer feature extracted by the CNN. Finally, the feature-enhanced sample data was used to train the fully connected classification network to complete the image classification task. Comparing MCN with pure convolutional network and single Mapper feature classification algorithm on MNIST and FashionMNIST data sets, the initial classification accuracy of the MCN with PCA (Principal Component Analysis) dimension reduction is improved by 4.65% and 8.05%, the initial classification accuracy of the MCN with LDA (Linear Discriminant Analysis) dimensionality reduction is improved by 8.21% and 5.70%. Experimental results show that MCN has higher classification accuracy and stronger anti-attack capability.
Aiming at the forgetting and underutilization of the text information of image in image captioning methods, a Scene Graph-aware Cross-modal Network (SGC-Net) was proposed. Firstly, the scene graph was utilized as the image’s visual features, and the Graph Convolutional Network (GCN) was utilized for feature fusion, so that the visual and textual features were in the same feature space. Then, the text sequence generated by the model was stored, and the corresponding position information was added as the textual features of the image, so as to solve the problem of text feature loss brought by the single-layer Long Short-Term Memory (LSTM) Network. Finally, to address the issue of over dependence on image information and underuse of text information, the self-attention mechanism was utilized to extract significant image information and text information and fuse then. Experimental results on Flickr30K and MS-COCO (MicroSoft Common Objects in COntext) datasets demonstrate that SGC-Net outperforms Sub-GC on the indicators BLEU1 (BiLingual Evaluation Understudy with 1-gram), BLEU4 (BiLingual Evaluation Understudy with 4-grams), METEOR (Metric for Evaluation of Translation with Explicit ORdering), ROUGE (Recall-Oriented Understudy for Gisting Evaluation) and SPICE (Semantic Propositional Image Caption Evaluation) with the improvements of 1.1,0.9,0.3,0.7,0.4 and 0.3, 0.1, 0.3, 0.5, 0.6, respectively. It can be seen that the method used by SGC-Net can increase the model’s image captioning performance and the fluency of the generated description effectively.
Text sentiment analysis has gradually become an important part of Natural Language Processing(NLP) in the fields of systematic recommendation and acquisition of user sentiment information, as well as public opinion reference for the government and enterprises. The methods in the field of sentiment analysis were compared and summarized by literature research. Firstly, literature investigation was carried out on the methods of sentiment analysis from the dimensions of time and method. Then, the main methods and application scenarios of sentiment analysis were summarized and compared. Finally, the advantages and disadvantages of each method were analyzed. According to the analysis results, in the face of different task scenarios, there are mainly three sentiment analysis methods: sentiment analysis based on emotion dictionary, sentiment analysis based on machine learning and sentiment analysis based on deep learning. The method based on multi-strategy mixture has become the trend of improvement. Literature investigation shows that there is still room for improvement in the techniques and methods of text sentiment analysis, and it has a large market and development prospects in e-commerce, psychotherapy and public opinion monitoring.
The dust accumulation on photovoltaic panels will reduce the conversion efficiency of photovoltaic power generation, and easily cause damage to the photovoltaic panels at the same time. Therefore, it is of great significance to recognize the dust accumulation of photovoltaic panels intelligently. Aiming at above problems, a dust accumulation degree recognition model of photovoltaic panel based on improved deep residual network was proposed. Firstly, the NeXt Residual Network (ResNeXt)50 was improved by decomposing convolution and fine-tuning down-sampling. Then, the Coordinate Attention (CA) mechanism was fused to embed the location information into channel attention, the channel relationship and long-term dependence were encoded by using the accurate location information, and the feature map was decomposed into two one-dimensional codes by using the two-dimensional global pooling operation, thereby enhencing the representation of the objects of attention. Finally, the cross-entropy loss function was replaced by the Supervised Contrast (SupCon) learning loss function to effectively improve the recognition accuracy. Experimental results show that in the recognition of the dust accumulation of photovoltaic panel at four levels of real photovoltaic power stations, the improved ResNeXt50 model has a recognition accuracy of 90.7%, which is increased by 7.2 percentage points compared with that of the original ResNeXt50. The proposed model can meet the basic requirements of intelligent operation and maintenance of photovoltaic power stations.
Users’ social media contains their past personal experiences and potential life patterns, and the study of their patterns is of great value for predicting users’ future behaviors and performing personalized recommendations for users. By collecting Weibo data, 11 types of events were defined, and a three?stage Pipeline system was proposed to detect personal events by using BERT (Bidirectional Encoder Representations from Transformers) pre?trained models in three stages respectively, including BERT+BiLSTM+Attention, BERT+FullConnect and BERT+BiLSTM+CRF. The information of whether the text contained defined events, the event types of events contained, and the elements contained in each event were extracted from the Weibo, and the specific elements are Subject (subject of the event), Object (event element), Time (event occurrence time), Place (place where the event occurred) and Tense (tense of the event), thereby exploring the change law of user’s personal event timeline to predict personal events. Comparative experiments and analysis were conducted with classification algorithms such as logistic regression, naive Bayes, random forest and decision tree on a collected real user Weibo dataset. Experimental results show that the BERT+BiLSTM+Attention, BERT+FullConnect, BERT+BiLSTM+CRF methods used in three stages achieve the highest F1?score, verifying the effectiveness of the proposed methods. Finally, the personal event timeline was visually built according to the extracted events with time information.
Spatio-temporal prediction task is widely applied in neuroscience, transportation, meteorology and other fields. As a typical spatio-temporal prediction task, temperature prediction needs to dig out the inherent spatio-temporal characteristics of temperature data. Aiming at the problems of large prediction error and insufficient spatial feature extraction in the existing temperature prediction algorithms, a temperature prediction model based on Graph Convolutional Network and Gated Recurrent Unit (GCN-GRU) was proposed. Firstly, the methods of weight redistribution and multi-order neighbor connection were used to modify Graph Convolutional Network (GCN) in order to effectively mine the unique spatial characteristics of the meteorological data. Secondly, the matrix multiplication of each recurrent unit in the Gated Recurrent Unit (GRU) was replaced by graph convolution operation, and all the recurrent units were connected in series to form a graph convolutional gating layer. Then, the graph convolutional gating layer was used to build the main network structure to extract the spatio-temporal characteristics of the data. Finally, the temperature prediction results were output through a fully connected output layer. Compared with the single models such as Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM), GCN-GRU had the Mean Absolute Error (MAE)reduced by 0.67 and 0.83 respectively; compared with the prediction model combined with Chebyshev graph convolution and Long Short-Term Memory (Cheb-LSTM) and the prediction model combined with Graph Convolutional Network and Long Short-Term Memory (GCN-LSTM), the proposed model had the MAE reduced by 0.36 and 0.23 respectively.
NVM (Non-Volatile Memory) storage system has the potential to provide high throughput, including near-memory read and write speeds, byte addressing features, and support for multi-way forwarding. However, the existing system software stack is not designed for NVM, which makes the system software stack have many factors that affect system access performance. Through analysis, it is found that the lock mechanism of the file system has a large overhead, which makes the concurrent access of data to be a difficult problem under multi-core environment. In order to alleviate these problems, a lock-free file reading and writing mechanism as well as a byte-based read and write interface were designed. By eliminating the file-based lock mechanism, the coarse-grained access control was changed, and the self-management request was used to improve the concurrency of the process. When designing the new file access interface that can utilize byte addressing, the read-write asymmetry as well as the different characteristics of the read and write operations of the NVM storage device were considered. These designs reduce the overhead of the software stack and make the full use of NVM features to provide a high concurrent, high throughput, and durable storage system. Finally, based on the open source NVM simulator PMEM, the FPMRW prototype system was implemented, and the universal test tool Filebench was used to test and analyze the FPRRW. The results show that the FPMRW can improve the system throughput by 3%-40% compared with EXT+PMEM and XFS+PMEM.
Echo hiding is a steganographic technique with audio as carrier. Currently, the steganalysis methods for echo hiding mainly use the cepstral coefficients as handcrafted-features to realize classification. However, when the echo amplitude is low, the detection performance of these traditional methods is not high. Aiming at the low echo amplitude condition, a steganalysis method for echo hiding based on Convolutional Neural Network (CNN) was proposed. Firstly, Short-Time Fourier Transform (STFT) was used to extract the amplitude spectrum coefficient matrix as the shallow feature. Secondly, the deep feature was extracted by the designed CNN framework from the shallow feature. The network framework consisted of four convolutional blocks and three fully connected layers. Finally, the classification results were output by Softmax. The proposed method was steganographically evaluated on three classic echo hiding algorithms. Experimental results indicate that the detection rates of the proposed method under low echo amplitude are 98.62%, 98.53% and 93.20% respectively. Compared with the existing traditional handcrafted-features based methods and deep learning based methods, the proposed method has the detection performance improved by more than 10%.
In visual detection of subminiature accessory, the extracted target contour will be affected by the existence of foreign matter in the field like dust and hair crumbs. In order to avoid the impact for measurement brought by foreign matter, a method of culling foreign matter fake information based on prior knowledge was put forward. Firstly, the corners of component image with foreign matter were detected. Secondly, the corner-distribution features of standard component were obtained by statistics. Finally, the judgment condition of foreign matter fake imformation was derived from the corner-distribution features of standard component to cull the foreign matter fake information. Through successful application in an actual engineering project, the processing experiments on three typical images with foreign matter prove that the proposed algorithm ensures the accuracy of the measurement, while effectively culling the foreign matter fake information in the images.
A new image retrieval method based on enhanced micro-structure and context-sensitive similarity was proposed to overcome the shortcoming of high dimension of combined image feature and intangible combined weights. A new local pattern map was firstly used to create filter map, and then enhanced micro-structure descriptor was extracted based on color co-occurrence relationship. The descriptor combined several features with the same dimension as single color feature. Based on the extracted descriptor, normal distance between image pairs was calculated and sorted. Combined with the iterative context-sensitive similarity, the initial sorted image series were re-ranked. With setting the value of iteration times as 50 and considering the top 24 images in the retrieved image set, the comparative experiments with Multi-Texton Histogram (MTH) and Micro-Structure Descriptor (MSD) show that the retrieval precisions of the proposed algorithm respectively are increased by 13.14% and 7.09% on Corel-5000 image set and increased by 11.03% and 6.8% on Corel-10000 image set. By combining several features and using context information while keeping dimension unchanged, the new method can enhance the precision effectively.