Named Entity Recognition (NER), a pivotal task in information extraction, is aimed at identifying named entities of various types within text. Traditional NER methods, however, often fall short in providing sufficient semantic representation of text and preserving word order information. Addressing these challenges, a novel approach is proposed, leveraging dual Graph Neural Networks (GNNs) based on multi-feature fusion. This approach constructs a co-occurrence graph and a dependency syntax graph from text sequences, capturing textual features from a dual-graph perspective to overcome the oversight of word interdependencies. Furthermore, Bidirectional Long Short-Term Memory Networks (BiLSTMs) are utilized to encode text, addressing the issues of neglecting word order features and the difficulty in capturing contextual semantic information. Additionally, to enable the model to learn features across different subspaces and the varying degrees of information significance, a multi-head self-attention mechanism is introduced for calculating internal dependency weights within feature vectors. The proposed model achieves F1-scores of 84.85% and 96.34% on the CCKS-2019 and Resume datasets, respectively, marking improvements of 1.13 and 0.67 percentage points over baseline models. The results affirm the effectiveness of the presented method in enhancing performance on the NER task.

The diagnosis and treatment of breast cancer (BC) are significantly subject to medical imaging techniques, with segmentation being crucial in delineating pathological regions for precise diagnosis and treatment planning. This comprehensive analysis explores a variety of segmentation methodologies, encompassing classical, machine learning, deep learning (DL), and manual segmentation, as applied in the medical imaging field for BC detection. Classical segmentation techniques, which include edge-driven and threshold-driven segmentation, are highlighted for their utilization of filters and region-based methods to achieve precise delineation. Emphasis is placed on the establishment of clear guidelines for the selection and comparison of these classical approaches. Segmentation through machine learning is discussed, encompassing both unsupervised and supervised techniques that leverage annotated images and pathology reports for model training, with a focus on their efficacy in BC segmentation tasks. DL methods, especially models such as U-Net and convolutional neural networks (CNNs), are underscored for their remarkable efficiency in segmenting BC images, with U-Net models noted for their minimal requirement for annotated images and achieving accuracy levels up to 99.7%. Manual segmentation, though reliable, is identified as time-consuming and susceptible to errors. Various metrics, such as Dice, F-score, Intersection over Union (IOU), and Area Under the Curve (AUC), are used for assessing and comparing the segmentation techniques. The analysis acknowledges the challenges posed by limited dataset availability, data range inadequacy, and confidentiality concerns, which hinder the broader integration of segmentation methods into clinical practice. Solutions to overcome these challenges are proposed, including the promotion of partnerships to develop and distribute extensive datasets for BC segmentation. This approach would necessitate the pooling of resources from multiple organizations and the adoption of anonymization techniques to safeguard data privacy. Through this lens, the analysis aims to provide a thorough analysis of the practical implications of segmentation methods in BC diagnosis and management, paving the way for future advancements in the field.

Skin cancer, a significant health concern globally, necessitates innovative strategies for its early detection and classification. In this context, a novel methodology employing the state-of-the-art EfficientNetB0 deep learning architecture has been developed, aiming to augment the accuracy and efficiency of skin cancer diagnoses. This approach focuses on automating the classification of skin lesions, addressing the challenges posed by their complex structures and the subjective nature of conventional diagnostic methods. Through the adoption of advanced training techniques, including adaptive learning rates and Rectified Adam (RAdam) optimization, a robust model for skin cancer classification has been constructed. The findings underscore the model's capability to achieve convergence during training, illustrating its potential to transform dermatological diagnostics significantly. This research contributes to the broader fields of medical imaging and artificial intelligence (AI), underscoring the efficacy of deep learning in enhancing diagnostic processes. Future endeavors will explore the realms of explainable AI (XAI), collaboration with medical professionals, and adaptation of the model for telemedicine, ensuring its continued relevance and applicability in the dynamic landscape of skin cancer diagnosis.

In the realm of digital image security, this study presents an innovative encryption methodology for color images, significantly advancing the traditional Vigenere cipher through the integration of two extensive pseudorandom substitution matrices. These matrices are derived from chaotic maps widely recognized for their cryptographic utility, specifically the logistic map and the skew tent map, chosen for their straightforward implementation capabilities in encryption systems and their high sensitivity to initial conditions. The process commences with the vectorization of the original image and the computation of initial values to alter the starting pixel's value, thereby initiating the encryption sequence. A novel aspect of this method is the introduction of a Vigenere mechanism that employs dynamic pseudorandom affine functions at the pixel level, enhancing the cipher's robustness. Subsequently, a comprehensive permutation strategy is applied to bolster the vector's integrity and elevate the temporal complexity against potential cryptographic attacks. Through simulations conducted on a varied collection of images, encompassing different sizes and formats, the proposed encryption technique demonstrates formidable resilience against both brute-force and differential statistical attacks, thereby affirming its efficacy and security in safeguarding digital imagery.

In the realm of facial expression recognition (FER), the identification and classification of seven universal emotional states, surprise, disgust, fear, happiness, neutrality, anger, and contempt, are of paramount importance. This research focuses on the application of convolutional neural networks (CNNs) for the extraction and categorization of these expressions. Over the past decade, CNNs have emerged as a significant area of research in human-computer interaction, surpassing previous methodologies with their superior feature learning capabilities. While current models demonstrate exceptional accuracy in recognizing facial expressions within controlled laboratory datasets, their performance significantly diminishes when applied to real-time, uncontrolled datasets. Challenges such as degraded image quality, occlusions, variable lighting, and alterations in head pose are commonly encountered in images sourced from unstructured environments like the internet. This study aims to enhance the recognition accuracy of FER by employing deep learning techniques to process images captured in real-time, particularly those of lower resolution. The objective is to augment the accuracy of FER in real-world datasets, which are inherently more complex and collected under less controlled conditions, compared to laboratory-collected data. The effectiveness of a deep learning-based approach to emotion detection in photographs is rigorously evaluated in this work. The proposed method is exhaustively compared with manual techniques and other existing approaches to assess its efficacy. This comparison forms the foundation for a subjective evaluation methodology, focusing on validation and end-user satisfaction. The findings conclusively demonstrate the method's proficiency in accurately recognizing emotions in both laboratory and real-world scenarios, thereby underscoring the potential of deep learning in the domain of facial emotion identification.

The development of an adaptive Lane Keeping Assistance System (LKAS) is presented, focusing on enhancing vehicular lateral stability and alleviating driver workload. Traditional LKAS with static parameters struggle to accommodate varying driver behaviors. Addressing this challenge, the proposed system integrates adaptive driver characteristics, aligning with individual driving habits and intentions. A novel lane departure decision model is introduced, employing time-space domain fusion to effectively discern driver's lane change intentions, thus informing system decisions. Further innovation is realized through the application of reinforcement learning theory, culminating in the creation of a master controller for lane departure intervention. This controller dynamically adjusts to driver behavior, optimizing lane keeping accuracy. Extensive simulations, coupled with hardware-in-the-loop experiments using a driving simulator, substantiate the system's efficacy, demonstrating marked improvements in lane keeping precision. These advancements position the system as a significant contribution to the field of driver assistance technologies.

In this study, an integrated pest and disease recognition system for agricultural drones has been developed, leveraging deep learning technologies to significantly improve the accuracy and efficiency of pest and disease detection in agricultural settings. By employing convolutional neural networks (CNN) in conjunction with high-definition image acquisition and wireless data transmission, the system demonstrates proficiency in the effective identification and classification of various agricultural pests and diseases. Methodologically, a deep learning framework has been innovatively applied, incorporating critical modules such as image acquisition, data transmission, and pest and disease identification. This comprehensive approach facilitates rapid and precise classification of agricultural pests and diseases, while catering to the needs of remote operation and real-time data processing, thus ensuring both system efficiency and data security. Comparative analyses reveal that this system offers a notable enhancement in both accuracy and response time for pest and disease recognition, surpassing traditional detection methods and optimizing the management of agricultural pests and diseases. The significant contribution of this research is the successful integration of deep learning into the domain of agricultural pest and disease detection, marking a new era in smart agriculture technology. The findings of this study bear substantial theoretical and practical implications, advancing precision agriculture practices and contributing to the sustainability and efficiency of agricultural production.

In the domain of intellectual property protection, the embedding of digital watermarks has emerged as a pivotal technique for the assertion of copyright, the conveyance of confidential messages, and the endorsement of authenticity within digital media. This research delineates the implementation of a non-blind watermarking algorithm, utilizing alpha blending facilitated by discrete wavelet transform (DWT) to embed watermarks into genuine images. Thereafter, an extraction process, constituting the inverse of embedding, retrieves these watermarks. The robustness of the embedded watermark against prevalent manipulative attacks, specifically median filter, salt and pepper (SAP) noise, Gaussian noise, speckle noise, and rotation, is rigorously evaluated. The performance of the DWT-based watermarking is quantified using the peak signal-to-noise ratio (PSNR), an objective metric reflecting fidelity. It is ascertained that the watermark remains tenaciously intact under such adversarial conditions, underscoring the proposed method's suitability for applications in digital image security and copyright verification.

This investigation delineates an optimised predictive model for employee attrition within a substantial workforce, identifying pertinent models tailored to the specific context of employee and organisational variables. The selection and refinement of the appropriate predictive model serve as cornerstones for enhancements and updates, which are integral to honing the model's precision in prognosticating potential departures. Through meticulous optimisation, the model demonstrates proficiency in pinpointing the pivotal factors contributing to employee turnover and elucidating the interdependencies among salient variables. A suite of 27 general and eight critical variables were scrutinised. Pertinent correlations were unearthed, notably between monthly income and job satisfaction, home-to-work distance and job satisfaction, as well as age with both job satisfaction and performance metrics. Drawing from prior studies in analogous domains, a three-stage analytical methodology encompassing data exploration, model selection, and implementation was employed. The rigorous training of the optimised model encompassed both attrition factors and variable correlations, culminating in predictive outcomes with a precision of 90% and an accuracy of 87%. Implementing the refined model projected that 113 out of 709 employees, equating to 15.93%, were at a heightened risk of exiting the organisation. This quantitative foresight equips stakeholders with a strategic tool for preemptive interventions to mitigate turnover and sustain organisational vitality.

In Sub-Saharan Africa, particularly in Nigeria, Lassa fever poses a significant infectious disease threat. This investigation employed count regression and machine learning techniques to model mortality rates associated with confirmed Lassa fever cases. Utilizing weekly data from January 7, 2018, to April 2, 2023, provided by the Nigeria Centre for Disease Control (NCDC), an analytical comparison between these methods was conducted. Overdispersion was indicated (p<0.01), prompting the exclusive use of negative binomial and generalized negative binomial regression models. Machine learning algorithms, specifically medium Gaussian support vector machine (MGSVM), ensemble boosted trees, ensemble bagged trees, and exponential Gaussian Process Regression (GPR), were applied, with 80% of the data allocated for training and the remaining 20% for testing. The efficacy of these methods was evaluated using the coefficients of determination (R²) and the root mean square error (RMSE). Descriptive statistics revealed a total of 30,461 confirmed cases, 4,745 suspected cases, and 772 confirmed fatalities attributable to Lassa fever during the study period. The negative binomial regression model demonstrated superior performance (R²=0.1864, RMSE=4.33) relative to the generalized negative binomial model (R²=0.1915, RMSE=18.2425). However, machine learning algorithms surpassed the count regression models in predictive capability, with ensemble boosted trees emerging as the most effective (R²=0.85, RMSE=1.5994). Analysis also identified the number of confirmed cases as having a significant positive correlation with mortality rates (r=0.885, p<0.01). The findings underscore the importance of promoting community hygiene practices, such as preventing rodent intrusion and securing food storage, to mitigate the transmission and consequent fatalities of Lassa fever.

In the realm of agriculture, crop yields of fundamental cereals such as rice, wheat, maize, soybeans, and sugarcane are adversely impacted by insect pest invasions, leading to significant reductions in agricultural output. Traditional manual identification of these pests is labor-intensive and time-consuming, underscoring the necessity for an automated early detection and classification system. Recent advancements in machine learning, particularly deep learning, have provided robust methodologies for the classification and detection of a diverse array of insect infestations in crop fields. However, inaccuracies in pest classification could inadvertently precipitate the use of inappropriate pesticides, further endangering both agricultural yields and the surrounding ecosystems. In light of this, the efficacy of nine distinct pre-trained deep learning algorithms was evaluated to discern their capability in the accurate detection and classification of insect pests. This assessment utilized two prevalent datasets, comprising ten pest classes of varied sizes. Among the transfer learning techniques scrutinized, adaptations of ResNet-50 and ResNet-101 were deployed. It was observed that ResNet-50, when employed in a transfer learning paradigm, achieved an exemplary classification accuracy of 99.40% in the detection of agricultural pests. Such a high level of precision represents a significant advancement in the field of precision agriculture.

Diabetic retinopathy, a severe ocular disease correlated with elevated blood glucose levels in diabetic patients, carries a significant risk of visual impairment. The essentiality of its timely and precise severity classification is underscored for effective therapeutic intervention. Deep learning methodologies have been shown to yield encouraging results in the detection and categorisation of severity levels of diabetic retinopathy. This study proposes a dual-level approach, wherein the MobileNetV2 model is modified for a regression task, predicting retinopathy severity levels and subsequently fine-tuned on fundus images. The refined MobileNetV2 model is then utilised for learning feature embeddings, and a Support Vector Machine (SVM) classifier is trained for grading retinopathy severity. Upon implementation, this dual-level approach demonstrated remarkable performance, achieving an accuracy rate of 87% and a kappa value of 93.76% when evaluated on the APTOS19 benchmark dataset. Additionally, the efficacy of data augmentation and the handling of class imbalance issues were explored. These findings suggest that the novel dual-level approach provides an efficient and highly effective solution for the detection and classification of diabetic retinopathy severity levels.

In addressing the challenge of obstacle scattering inversion amidst intricate noise conditions, a model predicated on convolutional neural networks (CNN) has been proposed, demonstrating high precision. Five distinct noise scenarios, encompassing Gaussian white noise, uniform distribution noise, Poisson distribution noise, Laplace noise, and impulse noise, were evaluated. Far-field data paired with the Fourier coefficients of obstacle boundary curves were employed as network input and output, respectively. Through the convolutional processes inherent to the CNN, salient features within the far-field data related to obstacles were adeptly identified. Concurrently, the statistical characteristics of the noise were assimilated, and its perturbing effects were diminished, thus facilitating the inversion of obstacle shape parameters. The intrinsic capacity of CNNs to intuitively learn and differentiate salient features from data eradicates the necessity for external intervention or manually designed feature extractors. This adaptability confers upon CNNs a significant edge in tackling obstacle scattering inversion challenges, particularly in light of fluctuating data distributions and feature variability. Numerical experiments have substantiated that the aforementioned CNN model excels in addressing scattering inversion complications within multifaceted noise conditions, consistently delivering solutions with remarkable precision.

The task of interpreting multi-variable time series data, while also forecasting outcomes accurately, is an ongoing challenge within the machine learning domain. This study presents an advanced method of utilizing Long Short-Term Memory (LSTM) recurrent neural networks in the analysis of such data, with specific attention to both target and exogenous variables. The novel approach aims to extract hidden states that are unique to individual variables, thereby capturing the distinctive dynamics inherent in multi-variable time series and allowing the elucidation of each variable's contribution to predictive outcomes. A pioneering mixture attention mechanism is introduced, which, by leveraging the aforementioned variable-specific hidden states, characterizes the generative process of the target variable. The study further enhances this methodology by formulating associated training techniques that permit concurrent learning of network parameters, variable interactions, and temporal significance with respect to the target prediction. The effectiveness of this approach is empirically validated through rigorous experimentation on three real-world datasets, including the 2022 closing prices of three major stocks - Apple (AAPL), Amazon (AMZN), and Microsoft (MSFT). The results demonstrated superior predictive performance, attributable to the successful encapsulation of the diverse dynamics of different variables. Furthermore, the study provides a comprehensive evaluation of the interpretability outcomes, both qualitatively and quantitatively. The presented framework thus holds substantial promise as a comprehensive solution that not only enhances prediction accuracy but also aids in the extraction of valuable insights from complex multi-variable datasets.