Image inpainting represents a fundamental task in computer vision, primarily focusing on the generation of missing content within an image to restore its integrity and aesthetics. Existing GAN-based approaches often yield content with ambiguity and entail high training costs. They tend to concentrate narrowly on damaged regions, leading to distortions along edges, which consequently hampers generalization. To overcome these challenges and achieve high-fidelity image inpainting, we introduce an image editing algorithm to the image inpainting task by designing two distinct networks. The first network, Edge-e4e, uses pretrained StyleGAN2 for global image generation, mitigating edge distortions in damaged regions and reducing training costs. Simultaneously, we incorporate contour information in the damaged areas to ensure the correctness of the restoration content. The second network, the Appending network, includes two style-based encoders and a generator to refine the images restored by the Edge-e4e network. Specifically, we subtract the restored images from the input images in the channel dimension to obtain a distortion map, which serves as a prior to refine the restored images. The encoders extract features from the input images and distortion map, while the generator is employed to generate optimized images. To enhance the quality of refined images, we propose integrating plugin and modulate plugin modules into the Appending network for style extraction and fusion, leveraging the available information from input images and blending it into the generator. Experimental results demonstrate that our algorithm achieves high-fidelity restoration and excellent generalization, with optimal FID and Lpips metrics of 0.0631 and 0.875, respectively.

Public security is a crucial aspect of maintaining social order. Although crime rates in western cultures may be considered socially acceptable, it is important to continually improve security measures to prevent potential risks.  With the advancements in artificial intelligence methods, particularly in deep learning and computer vision, it has become possible to detect abnormal event patterns in groups of people. This paper presents a systematic review of deep learning techniques employed for identifying gatherings of people and detecting anomalous events to enhance public security. Some of the open research areas are identified, including the lack of works addressing multiple cases of anomalies in large concentrations of people, which leaves open an important avenue for future scientific work.

Lots of mobile devices such as digital camera, mobile phones, and personal digital assistants (PDAs) adopt single image sensors to acquire scene images for real-time processing. Convolutional neural networks (CNN) have been shown good performance in lots of area about image processing. In this paper, we develop a demosaicking method based on deep residual CNN (DRCNN). Firstly, we initialize the mosaicked images using conventional demosaicking algorithms and learn the DRCNN for three color channels. In DRCNN, we adopt binary convolution unit (BCU), efficient layer aggregation networks (ELAN), and dense residual blocks (DRB). Experimental results show that our proposed algorithms outperforms existing approaches in PSNR, computational complexity, and visual quality.

Knowledge graphs convey precise semantic information that can be effectively interpreted by neural networks, and generating descriptive text based on these graphs places significant emphasis on content consistency. However, knowledge graphs are inadequate for providing additional linguistic features such as paragraph structure and expressive modes, making it challenging to ensure content coherence in generating text that spans multiple sentences. This lack of coherence can further compromise the overall consistency of the content within a paragraph. In this work, we present the generation of scientific abstracts by leveraging knowledge graphs, with a focus on enhancing both content consistency and coherence. In particular, we construct the ACL Abstract Graph Dataset (ACL-AGD) which pairs knowledge graphs with text, incorporating sentence labels to guide text structure and diverse expressions. We then implement a Siamese network to complement and concretize the entities and relations based on paragraph structure by accomplishing two tasks: graph-to-text generation and entity alignment. Extensive experiments demonstrate that the logical paragraphs generated by our method exhibit entities with a uniform position distribution and appropriate frequency. In terms of content, our method accurately represents the information encoded in the knowledge graph, prevents the generation of irrelevant content, and achieves coherent and non-redundant adjacent sentences, even with a shared knowledge graph.

Background With the rapid development of artificial intelligence technology, highly intelligent and unmanned factories have become an important trend. In the complex environments of smart factories, the long-term tracking and inspection of specified targets such as operators and special goods, as well as comprehensive visual recognition and decision-making capabilities throughout the whole production process, are critical components of automated unmanned factories. It is inevitable that issues such as target occlusion and disappearance will occur, thus exacerbating the long-term tracking challenge. Currently, there are no long-term tracking studies specifically addressing occlusions in these environments. Methods We first construct three new benchmark datasets in the complex workshop environment of a smart factory (referred to as SF-Complex3 data), which include challenging conditions such as complete occlusion and partial occlusion of targets. Next, utilising a brain memory inspired approach, we determine uncertainty estimation parameters: confidence, peak-to-sidelobe ratio (PSR), and average peak to correlation energy (APCE), to derive a continual learning based adaptive model update method. Additionally, we design a lightweight target detection model to automatically detect and locate targets in the initial frame and during re-detection. Finally, we integrate the algorithm with ground mobile robots and UAV-based imaging and processing equipment to build a new visual detection and tracking framework, termed SFC-RT (Smart Factory Complex Tracking and Identification). Results We conducted extensive tests on the benchmark UAV20L and SF-Complex3 datasets. Compared to state-of-the-art tracking algorithms, our proposed algorithm demonstrates an average performance improvement of 6% when facing key challenging attributes. Moreover, it can smoothly run on embedded platforms, including mobile robots and UAVs, at a real-time speed of 36.4 fps. Conclusions The proposed SFC-RT framework is shown to efficiently and accurately address the challenges of target loss and occlusion for long-term tracking within complex smart factory environments. It meets the requirements of real-time performance, robustness, and lightweight design.

The deployment of large language models (LLMs) on edge devices and non-server environments presents significant challenges, primarily due to constraints in memory usage, computational power, and inference time. This paper investigates the feasibility of running LLMs across such devices by focusing on optimizing memory usage, employing quantization techniques, and reducing inference time. Specifically, we utilize LLaMA 2 for biomedical text summarization and implement Low-Rank Adaptation (LoRA) quantization to compress the model size to compress the model size and fine-tune it using limited resources. Our study systematically evaluates memory consumption during both training and inference phases, demonstrating substantial reductions through efficient LoRA quantization. Our results indicate that with careful optimization, it is feasible to deploy sophisticated LLMs like LLaMA 2 on low powered devices, thereby broadening the scope of their application in resource-constrained environments.

A document by Kunal Chakma. Click on the document to view its contents.

In the global monitoring of smart cities, the demands of global object detection systems based on cloud and fog computing in intelligent systems can be satisfied by photographs with globally recognized properties. Nevertheless, conventional techniques are constrained by the imaging depth of field and can produce artifacts or indistinct borders, which can be disastrous for accurately detecting the object. In light of this, this paper proposes an artificial intelligence-based gradient learning network that gathers and enhances domain information at different sizes in order to produce globally focused fusion results. Gradient features, which provide a lot of boundary information, can eliminate the problem of border artifacts and blur in multi-focus fusion. The multiple-receptive module (MRM) facilitates effective information sharing and enables the capture of object properties at different scales. In addition, with the assistance of the global enhancement module (GEM), the network can effectively combine the scale features and gradient data from various receptive fields and reinforce the features to provide precise decision maps. Numerous experiments have demonstrated that our approach outperforms the seven most sophisticated algorithms currently in use.

In the realm of traditional representation learning for Natural Language Processing (NLP) tasks, the prevalent approach involves leveraging large language models like BERT, GPT-2, XLNet, and others, which demonstrate their prowess when trained on extensive datasets. However, the quest for effective representation learning becomes a formidable challenge when data is scarce or costly to acquire, such as in specialised domains and low-resource languages. This challenge stems from the scarcity of suitable algorithms and the inherent difficulties in learning representations from limited data. In this study, we introduce a novel and groundbreaking technique, the Attention-sequence Bidirectional Long Short-Term Memory (ABiLSTM), to tackle this challenge head-on. This method amalgamates the strengths of attention mechanisms and bidirectional Long Short-Term Memory (LSTM) networks, bolstering the model’s ability to capture intricate patterns and dependencies within sequential data, particularly in scenarios involving small or constrained datasets. Through empirical analysis, we delve into the intricate dynamics of neural network architectures, specifically examining the impact of varying numbers of hidden layers and the occurrence of training errors. These aspects are scrutinised in the context of unsupervised pretraining methods, with a focus on their effectiveness in generating robust and informative representations for open relation extraction tasks. Furthermore, we propose the use of knowledge graph techniques to tackle the challenge of imbalanced data during information extraction processes. Our study demonstrates that the use of ABiLSTM yields superior results in open relation extraction tasks, even when dealing with small or limited datasets. This advancement represents a significant contribution to the fields of natural language understanding and information extraction.

Dragon Boat Racing, a popular aquatic folklore team sport, is traditionally held during the Dragon Boat Festival. Inspired by this event, we propose a novel human-based meta-heuristic algorithm called dragon boat optimization (DBO) in this paper. It models the unique behaviors of each crew member on the dragon boat during the race by introducing social psychology mechanisms (social loafing, social incentive). Throughout this process, the focus is on the interaction and collaboration among the crew members, as well as their decision-making in different situations. During each iteration, DBO implements different state updating strategies. By modelling the crew’s behavior and adjusting the state updating strategies, DBO is able to maintain high-performance efficiency. We have tested the DBO algorithm with 29 mathematical optimization problems and 2 structural design problems. The experimental results demonstrate that DBO is competitive with state-of-the-art meta-heuristic algorithms as well as conventional methods.

Predictions arising from deep neural networks may be very accurate but not very robust, leading to uncertainty in their outcome. This critical problem is receiving growing attention from the Machine Learning (ML) community. A practical solution that is increasingly applied is calculating confidence bounds for the ML predictions. Most confidence bounds available in the literature are theoretically sound but unfeasible from a practical viewpoint. In this paper, we contribute to the literature with probabilistic confidence bounds based on conditional probabilities, and we demonstrate their operational validity by means of a real-world application that concerns the prediction of the sleeping states of car drivers.

This study introduces a cutting-edge method for few-shot font generation, capturing the complexity and subtlety of font styles with minimal reference style images. Motivated by the time-consuming and labor-intensive process of traditional font design, particularly for languages with extensive glyph sets, such as Chinese or Korean, our approach streamlines font creation by utilizing both global and local style elements through a patch-based attention mechanism and a multi-task encoder. This innovation not only addresses the high production costs and manual effort associated with conventional font design but also overcomes the limitations of prior techniques that rely on comprehensive component definitions or multi-stage training processes. By extracting global style codes for common font family characteristics and local style codes from detailed patches, our model emphasizes critical stylistic features such as serifs, stroke shapes, and spacing. The inclusion of triplet loss and style fidelity loss further refines the model's accuracy, ensuring the generated fonts faithfully replicate the desired styles. Demonstrated through experiments on Korean and Chinese characters, our method achieves better efficiency, quality, and fidelity, outpacing current state-of-the-art solutions and highlighting the potential of an attention-based patch encoding for font generation for different languages.

In the environmental security monitoring of smart cities, the infrared and visible image fusion method deployed on intelligent systems based on cloud and fog computing plays an vital role in providing enhanced images for target detection systems. However, the fusion quality can be significantly influenced by the illumination of the monitoring scenario in visible images. Therefore, conventional methods typically suffer a severe performance drop under the condition of insufficient illumination. To tackle this issue, we propose an illumination enhancement and target-aware fusion method (IETAFusion) based on artificial intelligence, which breaks the boundaries between the task of illumination enhancement and image fusion and provide a fusion result with better visual perception in nighttime scene. Specifically, we use a light-weight contrast enhancement module (CEM) restore the brightness of the visble image. Moreover, a Swin Transformer-based backbone network (STBNet) is utilized to facilitate information exchangement between the source images and enhance the capabilities of target awareness. Finally, the fused images are reconstructed by the contrast-texture retention module (CTRM) and reconstructor. The extensive experiments indicates that the proposed approach achieves improved performance both in human perception and quantitative analysis compared with the state-of-the-art (SOTA) methods.

Voice Liveness Detection (VLD) has become one of the hot research topics in the Internet of Things era. Conventional VLD methods are centralized solutions that are trained on abundant data collected from local clients and stored on the server, however, they have the risk of causing data islands and privacy leakage. To address this problem, we propose a novel word-level VLD framework based on asynchronous federated learning (FL) with pop noise, named FL-VLD. Structurally, FL-VLD takes the preprocessed voice for local model training and constructs a global model by only transmitting the learned weights after differential privacy with FL’s central server in an asynchronous manner. In addition, the local network of the framework incorporates the residual network and the spatial grouping enhancement module to optimize the complexity and accuracy of the global model. With the advantage of FL’s distributed structure, FL-VLD solves the data island problem in VLD scenario without threatening users’ privacy. Experimental results on the popular POCO dataset show that our proposal is clearly superior to the traditional centralized methods as well as overperforming other federated schemes in terms of fairness, stability, accuracy, and lightweightness. Further, for attacks involving far-field replay, synthesis, and conversion, FL-VLD has high generalisation capabilities. Finally, the ablation study attests to its efficacy.

The expansion of machine learning applications across dif- ferent domains has given rise to a growing interest in tap- ping into the vast reserves of data that are being generated by edge devices. To preserve data privacy, federated learn- ing was developed as a collaboratively decentralized privacy- preserving technology to overcome the challenges of data silos and data sensibility. This technology faces certain lim- itations due to the limited network connectivity of mobile devices and malicious attackers. In addition, data samples across all devices are typically not independent and iden- tically distributed (non-IID), which presents additional chal- lenges to achieving convergence in fewer communication rounds. In this paper, we have simulated attacks, namely Byzantine, label flipping, and noisy data attacks, besides non-IID data. We proposed Robust federated learning against poisoning attacks (RFCaps) to increase safety and accelerate conver- gence. RFCaps incorporates a prediction-based clustering and a gradient quality evaluation method to prevent attack- ers from the aggregation phase by applying multiple filters and also accelerate convergence by using the highest quality gradients. In comparison to MKRUM, COMED, TMean, and FedAvg algorithms, RFCap has high robustness in the pres- ence of attackers and has achieved a higher accuracy of up to 80% on both MNIST and Fashion-MNIST datasets.

Cardiovascular disease (CVD) believes to be a major cause of transience and indisposition worldwide. Early diagnosis and timely intervention are critical in preventing the progression of CVD and improving patient outcomes. Machine learning (ML) algorithms have emerged as powerful tools in CVD recognition, with the potential to assist physicians in making accurate and efficient diagnoses. This research paper explores the combination of multiple ML algorithms for CVD recognition, utilizing diverse datasets such as the Cleveland, Hungarian, Switzerland, statlog, and VA Long Beach datasets. Additionally, a CVD dataset comprising 12 attributes and 70,000 records is employed, demonstrating improved results through the proposed and trained model compared to previous prediction techniques for CVD. The performance of various ML techniques, including support vector machines (SVM), Naive Bayes (NB), K-Nearest Neighbor (KNN), Random Forest (RF), and Logistic Regression (LR), is evaluated and compared. The impact of feature selection and feature scaling on the models’ performance is also examined. An ensemble bagging techniques is applied which is being embedded with other classifiers. LR classifier embedded with bagging techniques proved to be our proposed model. The findings reveal that the proposed Hybrid Linear Regression Bagging Model (HLRBM) outperforms other models. Furthermore, the study highlights the significance of data preprocessing techniques, such as data normalization and class balancing, which significantly enhance the performance of all models. To this end, standard scalar and Synthetic Minority Over-sampling Technique (SMOTE) are employed. The study emphasizes the importance of selecting an appropriate ensemble technique in conjunction with various ML algorithms and preprocessing methods for CVD prediction. Overall, the research provides valuable insights into the potential of ML in improving CVD risk assessment.

In the present era, healthcare systems grapple with substan- tial volumes of medical data. However, a significant portion of this data is marked by incompleteness, inconsistency, er- rors, and unsuitability for training Machine Learning (ML) or Deep Learning (DL) algorithms. This necessitates preprocess- ing the data to render it amenable to utilization by ML/DL al- gorithms. Medical datasets predominantly feature two types of attributes: numerical and categorical values. The conver- sion of categorical features into numerical vectors is a crucial step in preparing the data for ML/DL algorithms, known as Feature Engineering (FE) based categorical encoding. Con- ventional and straightforward encoding of categorical fea- tures, termed one-hot encoding, generates multiple columns, thereby transforming data from a lower-dimensional to a higher- dimensional space. This approach poses challenges, includ- ing increased memory requirements due to the proliferation of columns. Considering these issues, this research proposes an encoding technique named “Weight of Evidence with En- tity Embedding” (WoEEE). The WoEEE approach bolsters the predictive capabilities of ML/DL algorithms by calculating the weight of evidence and concurrently mitigates dimension- ality issues. To empirically validate the proposed method, it is tested on six diverse datasets: Breast Cancer, Hospital Readmission, Vadu, Covid-19, Stroke, and Heartstatlog. Four distinct ML/DL algorithms—Decision Tree (DT), Random For- est (RF), Logistic Regression (LR), and a simple Feed-forward Neural Network (NN)— are employed for testing. The re- sults obtained demonstrate that the WoEEE approach yields an average improvement of 11.18%, 10.37%, 5.83%, 7.58%, 7.83%, and 6.83% across all combinations of datasets, classi- fiers, and encoding methods. Furthermore, an Anova test is performed to confirm the effectiveness of WoEEE in encod- ing categorical data, especially for tasks involving binary clas- sification. This enhances the treatment of categorical data in ML and data analytics scenarios. Overall, WoEEE shows po- tential as a valuable approach for categorical data encoding, making a positive contribution to the creation of effective techniques for handling this type of data in real-world appli- cations.

Anemia occurs when the hemoglobin (Hgb) value falls below a certain reference range. It requires many blood tests, radiological images, and tests for diagnosis and treatment. By processing medical data from patients with artificial intelligence and machine learning methods, disease predictions can be made for newly ill individuals and decision-support mechanisms can be created for physicians with these predictions. Thanks to these methods, which are very important in reducing the margin of error in the diagnoses made by doctors, the evaluation of data records in health institutions is also important for patients and hospitals. In this study, three hybrid models are proposed to classify non-anemia records, Hgb-anemia, folate deficiency anemia (FDA), iron deficiency anemia (IDA), and B12 deficiency anemia by combining artificial intelligence and machine learning methods TreeBagger with Crow Search Algorithm (CSO), Chicken Swarm Optimization Algorithm (CSO) and JAYA methods. The proposed hybrid models aim to achieve high performance by better emphasizing the importance of parameters. To solve the multiclass anemia classification problem, fuzzy logic-based parameter optimization is applied to improve the class-based accuracy as well as the overall accuracy in the dataset. The classification performances of the proposed methods are evaluated using accuracy, F-score, precision, and recall criteria to build a prediction model to identify the anemia type of anemic patients. A result of the study on the dataset taken from the Kaggle database found that the three proposed hybrid methods outperformed other studies using the same dataset and similar studies in the literature.