International Journal of Pattern Recognition and Artificial Intelligence

With the development of social media, multimodal hate speech that relies on images and text has become an emerging way of spreading hate. The detection of multimodal hate speech is gradually becoming an increasingly challenging task. While many works based on neural networks and multimodal machine learning were proposed to detect multimodal hate speech, only few attempts have been made in terms of the interpretability of the task. This leads to difficulties in analyzing prediction results and model improvement. Therefore, this paper investigates the interpretable multimodal hate speech detection task and develops an intrinsically interpretable deep learning method by leveraging the multimodal architecture. Specifically, we leverage a multimodal pretrained model as the backbone of the final detection results and parallel an interpretability module via a joint training approach, which calculates the input tokens and fine-grained tags through a filter-gate attention mechanism. The interpretability module provides an interpretable basis for the final result judgment. We conduct experiments on the hate speech detection dataset and demonstrate that our proposed method not only significantly outperforms other methods but also provides interpretable insights into the decisions of our model.

In the research of Tangut character recognition, it was attempted to detect the stroke of character with standard Hough transform (SHT), but hindered by the inability of SHT to detect the curve-line. In this paper, Hough transform with guidance of endpoints (HTGE) was proposed to achieve better performance on the aspect of curve and short line detection. HTGE is implemented based on the conception of hypothesis and verification, first, to assume there is line between each pair of endpoints, and then the existence of line is checked according to the corresponding area accumulated value in the parameter domain. With setting up of test object and evaluation criterion, the experiment was carried out to determine the key parameter for the best performance, and it can be concluded from the experiment result that considerable performance can be achieved with HTGE in the application of line detection.

The production of nitrogen oxides (NO_x) in coal-fired boiler combustion has been found as a significant source of environmental pollution. Flue gas denitrification is a standard NO_x control technology for small- and medium-sized coal-fired boilers. Achieving steady-state control in flue gas denitrification can be challenging since coal-fired boiler systems have complexity and significant delay. A model based on a learning-based K-nearest neighbor (KNN) query mechanism created for NO_x output soft prediction is proposed in this study. First, a knowledge base in the proposed model is established through spatial division in accordance with the previous combustion parameters. Moreover, the clusters are established based on the output NO_x values. Next, the domain of values of combustion parameters for the respective cluster is obtained. Second, the optimal cluster is selected using the knowledge base for an input vector q with new combustion parameters ($q_1$, $q_2$, $\dots$, $q_n)$. Lastly, the K tuples in the cluster the closest to the values of the input vector q are adopted to predict the output NO_x value of q. The predicted NO_x value can serve as a feedforward signal to control the output of the reductant for accurate denitrification. As revealed by the experimental results, the proposed practical model, capable of conducting the prediction in a sub-second time, is highly competitive with existing techniques. Furthermore, a deep learning algorithm (DLA) is designed, whereas it underperforms the KNN model.

Adversarial training is by far one of the most effective methods to improve the robustness of deep neural networks against adversarial examples. However, the trade-off between robustness and accuracy is still a challenge in adversarial training. Previous methods used adversarial examples with a fixed perturbation budget or specific perturbation budgets for each example, which is inefficient in improving the trade-off and lacks the ability to control the trade-off flexibly. In this paper, we show that the largest element of logit, $z_{\mathrm{\text{max}}}$, can roughly represent the minimum distance between an example and its neighboring decision boundary. Thus, we propose group adaptive adversarial training (GAAT) that divides the training dataset into several groups based on $z_{\mathrm{\text{max}}}$ and develops a binary search algorithm to determine the group perturbation budgets for each group. Using the group perturbation budgets to perform adversarial training can fine-tune the trade-off between robustness and accuracy. Extensive experiments conducted on CIFAR-10 and ImageNet-30 show that our GAAT can achieve a more perfect trade-off than TRADES, MMA, and MART.

There has been a lot of research on edge-computing task offloading in deep reinforcement learning (DRL). Deep reinforcement learning is one of the important algorithms in the current AI field, but there is still room for improvement in the time cost and adaptive correction ability of the algorithm. This paper studies the application of DRL algorithms in edge-computing task offloading, and its key innovation is to propose an MADRLCO algorithm, which is based on the design idea of the Actor–Critic framework, uses the DNN model to act as an Actor, and can more accurately locate the initial decision through iterative training, and use the LSTM model to optimize the Critic, which can be more accurate. The optimal decision can be located in a short period of time. The main work of this paper is divided into three parts: (1) The AC algorithm of the Actor–Critic framework in DRL is proposed to be applied to edge-computing task offloading. (2) To address the weak generalization ability of the basic version of the Actor–Critic algorithm in multi-objective optimization, the sequential quantitative correction and adaptive correction parameter K method are used to optimize the Critic frame, thereby improving the generalization ability of the model in multi-objective decision-making and improving the rationality of decision-making results. (3) Aiming at the problem of large time cost in the critical framework of the model, a search algorithm for resource allocation-related parameters based on the time-series prediction method is proposed (time-series forecasting is a research branch of pattern recognition), which reduces the time overhead of the algorithm and improves the adaptive correction capability of the model. The algorithm in this paper can adapt to not only the time-varying network channel state, but also the time-varying number of device connections. Finally, it is proved by experiments that compared with the DRL calculation offloading algorithm based on DNN plus binary search, the MADRLCO algorithm reduces the model training time by 66.27%, and in the environment of the time-varying number of devices in the metasystem, the average model average. The standard calculation rate is 0.0403 higher than that of the current optimal algorithm.

Lung cancer is an aggressive disease among all cancer-based diseases, because of causing huge mortality in humans. Thus, earlier discovery is a basic task for diagnosing lung cancer and it helps increase the survival rate. Computed tomography (CT) is a powerful imaging technique used to discover lung cancer. However, it is time-consuming for examining each CT image. This paper develops an optimized deep model for classifying the lung nodules. Here, the pre-processing is done using Region of Interest (ROI) extraction and adaptive Wiener filter. The segmentation is done using the DeepJoint model wherein distance is computed with a congruence coefficient for extracting the segments. The nodule identification is done by a grid-based scheme. The features such as Global Binary Pattern (GBP), Texton features, statistical features, perimeter and area, barycenter difference, number of slices, short axis and long axis and volume are considered. The lung nodule classification is done to classify part solid, solid nodules and ground-glass opacity (GGO) using Deep Residual Network (DRN), which is trained by the proposed Shuffled Shepard Sine–Cosine Algorithm (SSSCA). The developed SSSCA is generated by the integration of the Sine–Cosine Algorithm (SCA) and Shuffled Shepard Optimization Algorithm (SSOA). The proposed SSSCA-based DRN outperformed with the highest testing accuracy of 92.5%, sensitivity of 93.2%, specificity of 83.7% and $F_1$-score of 81.5%.

Plants diseases is a major threat to agricultural production. Reduced yield due to plant diseases can lead to immeasurable economic losses. Therefore, the detection and classification of plant diseases are of great significance. Most of the existing plant disease detection methods focus on improving the identification accuracy. However, besides accuracy, real-time performance cannot be ignored. In this paper, a new module named 2-way residual dense layer is presented to effectively decrease the number of parameters in our network. In this module, depth separable convolution is introduced, which reduces the amount of parameter calculation and achieves a performance of over 98%. Our network is verified by an open dataset which includes 4503 images from four classes, including Mango, Arjun, Alstonia Scholaris, Guava, Bael, Jamun, Jatropha, Pongamia Pinnata, Basil, Pomegranate, Lemon, and Chinar. The leaf images of these plants have healthy and diseased condition. The experimental results showed that this method can be practically applied to the identification of plant leaf diseases and provide a basis for the identification of other leaf diseases.

In this paper, we propose a structure-preserving and denoising low-light enhancement method that uses the coefficient of variation. First, we use the coefficient of variation to process the original low-light image, which is used to obtain the enhanced illumination gradient reference map. Second, we use the total variation (TV) norm to regularize the reflectance gradient, which is used to maintain the smoothness of the image and eliminate the artifacts in the reflectance estimation. Finally, we combine the above two constraint terms with the Retinex theory, which contains the denoising regular term. The final enhanced and denoised low-light image is obtained by iterative solution. Experimental results show that our method can achieve superior performance in both subjective and objective assessments compared with other state-of-the-art methods (the source code is available at: https://github.com/bbxavi/SPDLEM.).

Time series is a very common but important data type. A large number of time series data are generated in various professional research fields and daily life. Although there are many models being developed to deal with time series, the cluster methods for time series are insufficient and need to improve. This paper is focused on time series clustering, which uses deep learning approach to discover the shape characteristics of time series. We establish a new neural network model of time series clustering to jointly optimize the representation learning and clustering tasks of time series. Focusing on shape features with time series, we built the Soft-DTW layer into the neural network to learn the interpretable time series representation. Maximized regularization mutual information is used to jointly optimize representation learning and clustering tasks. Experiments show that this model can help obtain an excellent representation of time series. In comparison with the benchmark model, the best clustering effect is achieved in the proposed model on multiple data sets. This model has broad applicability in time series data.

In order to solve the key problems of secondary surveillance radar signal debugging on special devices such as DSP and FPGA, a parallel processing scheme of secondary surveillance radar response signal is proposed based on CPU–GPU architecture. It can effectively reduce the difficulty of code development and improve the portability of program. The parallel optimization design of each processing model of response signal is made through the characteristics of shared memory in CPU–GPU architecture to improve the processing speed of the algorithm. The proposed scheme is tested and analyzed on different graphics cards by the secondary surveillance radar Mode-5 response signal in this paper. The experimental results showed that it takes 8390.52960 us to implement the signal processing algorithm based on NVIDIA Tesla K40c graphics card, which can save 51.98% of time than NVIDIA Quadro K4200 graphics card, and makes it possible to do the real-time processing of the secondary surveillance radar signal.

Aspect-based summarization differs from generic text-summarization in which the generated summary must be conditioned on a given topic. A fundamental challenge to the aspect-based summarization approach is the lack of labeled data for training models, which limits the usage of supervised methods. One approach to address this issue is to introduce human intervention to generate unique datasets per aspect. However, there is a large number of possible aspects to summarize which makes this option impossible to scale. This limits the use of typical modeling techniques, and requires methods which excel in few-shot, or ideally zero-shot regimes. Hence, in this research, we propose a modular, two-step approach that does not need any aspect-based supervision. This research combines recent advances in zero-shot text classification and generic summarization in a novel way. The backbone of the proposed approach is a transformer network trained for the task of textual entailment, which is used to reduce a document to the set of on topic sentences. In the experiments, our model achieves a new state of the art compared to other unsupervised models on the MA-News dataset (ROUGE-1 35.70 and ROUGE-2 15.52), and even outperforms fine-tuned models without any supervision of its own.

Although the CenterNet object detection method can be used to achieve high accuracy, it cannot be used to detect objects with overlapping or almost overlapping centers. However, a good trade-off between detection speed and accuracy cannot be achieved when utilizing this method. We improved a series of backbones using CenterNet, output the fusion of multiple feature maps in backbone, and used the feature pyramid network (FPN) mechanism for multiscale object detection. Additionally, we improved the head of the detector and considered adding intersection over union (IoU) branches. Finally, the loss function was improved to improve detection accuracy. Based on the above research, the CenterNet Plus object detection method is proposed in this paper. Through experiments on the COCO dataset, it can be seen that the use of a multiscale FPN mechanism not only solves the problem of center point overlap but also helps improve the accuracy of detection. CenterNet Plus can be used to greatly improve the detection accuracy on the premise of having a higher detection speed.

Currently, there are considerable challenges associated with the theoretical derivation and numerical calculation of Green’s function for spherical layered earths. In order to study the coupling effect between wide-area interconnected systems and the earth, this earth model is widely used. Artificial intelligence methods have been applied in order to overcome these challenges. First, Green’s function in the form of infinite Legendre series is derived using intelligent symbolic operations, and the weight function of Legendre series is recursively computed by means of a recursive algorithm. It is on this basis that the intelligent complex image method has been developed for solving spherical layered Green’s functions, which transforms the summation of infinite series of Green’s functions into the superposition of complex image potentials. Additionally, the upper error limit of the algorithm has been determined, and examples have been provided to demonstrate the accuracy of the algorithm. A solution based on multiple precision algorithms has been proposed to resolve the numerical singularity problem and the extremely slow convergence of the spherical layered Green’s function at the earth’s scale. It has been demonstrated that the intelligent complex image method has advantages in terms of computing speed and accuracy. A spherical layered Green’s function at the earth scale is presented in this paper as an effective and intelligent solution.

Object detection has been an important research branch in the field of computer vision. The single-shot-detection (SSD) is an object detection model based on deep learning, which can achieve a good balance between the detection accuracy and the detection speed, but has the problem of poor recognition accuracy for small objects. To address this limitation, this paper improves the structure of the SSD feature pyramid and up-samples the shallow feature map with small object information and fuses it with the upper feature map, thus enhancing the ability of the shallow feature map to represent detailed information. In this way, not only the overall detection accuracy of the SSD is improved, but also a relatively high detection speed is maintained. The proposed model is verified by experiments on two common datasets, the Pascal VOC and MS COCO datasets. On the Pascal VOC07+12, MS COCO14, and VOC07+12+COCO datasets, the improved model achieves the mean average precision values of 80.1% (+3.3% compared with the conventional model), 49.9% (+6.8%), and 82.1% (+3.0%), respectively. Meanwhile, the proposed model can achieve the detection speed of 42.2 frames per second.