Depression is a psychological illness characterized by the recurrent occurrence of a negative emotional state. The common cause of the rise in suicide cases that appear globally is major depressive disorder (MDD). Therefore, proper diagnosis and therapy are necessary to lessen the impact of depression. MDD is a serious and common ailment that causes functional frailty. However, its exact symptoms remain unknown. Therefore, the task of manually detecting MDD is challenging and subjective. Although electroencephalogram (EEG) signals have demonstrated potential in diagnosis, we still need to enhance their accuracy, clinical value, and efficiency. This study employs a unique method known as Blended Multilevel Features with Constraint Fusion Network (BMFCNet) to identify MDD. The residual-inception (RI) module of the BMFCNet extracts the most pertinent and discriminative characteristics from the high-level (HL) and low-level (LL) features. To combine improved LL and HL traits, a constraint fusion technique is introduced to weigh and fuse the LL and HL features adaptively. Furthermore, we demonstrate the effectiveness of our method by analogizing it with sixteen state-of-the-art (SOTA) methodologies utilizing two benchmark datasets.

Inverse kinematics, crucial in robotics, involves computing joint configurations to achieve specific end-effector positions and orientations. This task is particularly complex for six-degree-of-freedom (six-DoF) anthropomorphic robots due to complicated mathematical equations, nonlinear behaviours, multiple valid solutions, physical constraints, non-generalizability and computational demands. The primary contribution of this work is to address the complex inverse kinematics problem for six-DoF anthropomorphic robots through the systematic exploration of AI models. This study involves rigorous evaluation and Bayesian optimization for hyperparameter tuning to identify the optimal regressor, balancing both accuracy and computational efficiency. Utilizing five-fold cross-validation on a publicly available dataset, the selected model demonstrates exceptional performance in predicting six joint angles for end effector configuration, yielding an average mean square error of 1.934 × 10−3 to 3.522 × 10−3. Its computational efficiency, with a prediction time of approximately 1.25 ms per sample, makes it a practical choice. Additionally, the study employs Explainable AI, using SHAP (SHapley Additive exPlanations) analysis to gain an understanding of feature importance. This analysis not only enhances model interpretability but also reaffirms the efficacy in this challenging multi-input multi-output predictive task. This research advances state-of-the-art models and neural networks by prioritizing computational efficiency alongside accuracy—a critical yet often overlooked factor. Pioneering a significant advancement in anthropomorphic robot kinematics, it balances accuracy and efficiency, offering practical robotic automation solutions.

Schizophrenia (SCZ) is a severe neurological and physiological syndrome that perverts a patient’s perception of reality. SCZ exhibits several symptoms, including hallucinations, delusions, aberrant behavior, and thinking. It affects their professional, academic, personal, and social lives. Neurologists use a variety of verbal and visual tests to determine SCZ. However, these methods are laborious, time-consuming, superficial, and vulnerable to mistakes. Therefore, it is necessary to create an automated model for SCZ detection. Convolutional neural networks have swiftly established themselves in the field of mental health care due to the growth of deep learning in recent decades. Electroencephalogram (EEG) data records the variations in the neural dynamics of human memory. Using EEG data, this study proposes an automatic SCZ detection method using separable convolution attention network (SCZ-SCAN). The proposed network employs depth-wise separable convolution and attention networks on high-level and low-level to aggregate characteristics of 2-D scalogram images acquired from the continuous wavelet transform. The depth-wise separable convolutions help to create a lightweight framework, while attention techniques concentrate on significant features and reduce futile computations by removing the transmission of irrelevant features. The proposed approach has an average classification accuracy of 99% and 95% on the IBIB-PAN and EEG data from the basic sensory task in SZ dataset. Moreover, statistical hypothesis testing is performed using Wilcoxon’s Rank-Sum test to signify the model performance and it proves that SCZ-SCAN is statistically efficient to nine cutting-edge methods.

Experimental results show that the PSFAN statistically defeats 11 contemporary methods, proving its effectiveness for medical industrial applications.

Environmental microbes are certainly present in our surroundings since they are essential to the growth and survival of human advancement. The detailed analysis of environmental microorganisms (EMs) is very important to recognize, understand and make use of microbes as well and prevent damage. Extracting the discriminatory features from a limited-size dataset is very challenging for a deep learning model and a pure transformer-based network cannot achieve good classification results on a limited-size dataset due to the lack of muti-scale features. In this study, a novel vision transformer-based deep neural network is proposed by integrating the transformer with CNN for the classification of EM using microscopic images. The proposed network EMViT-Net has three main modules: a transformer module, a CNN module and a multilayer perceptron module. The transformer model extracted multiscale features to generate more discriminatory information from the images. A new separable convolutional parameter-sharing attention (SCPSA) block is integrated with the CNN module in the core of EMViT-Net, which makes the model robust to capture the local and global features, and simultaneously reduces the computational complexity of the model. The data augmentation is performed to introduce the variability in the dataset and counter the problem of overfitting and data imbalance. After extensive experiments and detailed analysis, it has been determined that the proposed model EMViT-Net outperforms the other existing methods and achieves state-of-the-art results with an accuracy of 71.17% which proves the effectiveness of the model for the classification of environmental microbes.

Facial expression analysis plays a crucial role in various fields, such as affective computing, marketing, and clinical evaluation. Despite numerous advances, research on facial expression recognition (FER) has recently been proceeding from confined lab circumstances to in-the-wild environments. FER is still an arduous and demanding problem due to occlusion and pose changes, intra-class and intensity variations caused by illumination, and insufficient training data. Most state-of-the-art approaches use entire face for FER. However, the past studies on psychology and physiology reveals that mouth and eyes reflect the variations of various facial expressions, which are closely related to the manifestation of emotion. A novel method is proposed in this study to address some of the issues mentioned above. Firstly, modified homomorphic filtering is employed to normalize the illumination, then the normalized face image is cropped into five local regions to emphasize expression-specific characteristics. Finally, a unique blended feature attention network (BFAN) is designed for FER. BFAN consists of both residual dilated multi-scale feature extraction module and spatial and channel-wise attention modules. These modules help to extract the most relevant and discriminative features from the high-level and low-level features. Then, both feature maps are integrated and passed on to the dense layers followed by a softmax layer to compute probability scores. Finally, the Choquet fuzzy integral is applied to the computed probability scores to get the final outcome. The superiority of the proposed method is exemplified by comparing it with eighteen existing approaches on seven benchmark datasets.

Content-based medical image retrieval (CBMIR) systems can assist and help doctors and radiologists in reliable diagnosis by retrieving relevant medical cases from past cases that share similarities with the current case. Therefore, there is a need for improved retrieval accuracy in CBMIR systems. Retrieving similar Brain Tumor magnetic resonance imaging (MRI) slices from the same class as the query is inherently challenging. For different types of tumors, there is no class-specific structure, size or shape. Further, MRIs being present in multiple views increase discrepancy in interview retrieval. This leads to high inter-class similarity but at the same time high intra-class variations. Skewed data quantities for Brain Tumor types like Meningioma further add to another challenge. It is necessary to formulate rich and generic MRI representations to address these issues. Toward the same, it is essential to model spatial contexts on a multi-scale across the tumor-affected regions and enhances the feature representational learning by extracting generic features. Hence, we propose a Weight Initialization Framework with Densely Connected Networks to improve generalization for Brain Tumor MRI retrieval. The proposed framework uplifts DenseNet-based models for feature extraction as they incorporate feature reuse and feature learning in a multi-scale manner. Further, a weight Initialization Framework (WIF) is used for improvising the representational learning. Specifically, WIF initializes weights of the DenseNet model by transfer learning-based adaptation, which then involves freezing the initial few layers. The freezing step ensures rich low-level features, even for long-tailed classes like Meningioma, while the remaining trainable layers are fine-tuned to incorporate domain-inherent feature learning. The proposed approach outperforms state-of-the-art by a margin of 1.70% and 1.69% on standard mAP and p@10, respectively. Concretely, when DenseNet and WIF are jointly employed, a stark increment in performance for the Meningioma class is observed, suggesting the generalizability of the framework.

Weld quality inspection is a critical step in industrial manufacturing, where accurate and efficient defect detection directly impacts product reliability and operational safety.

Although many state-of-the-art (SOTA) deep learning methods have been designed and tested for weld defect classification, most remain computationally demanding, limiting their suitability for real-time or embedded inspection systems, and/or their predictive performance, when evaluated under strict protocols, is often insufficient to ensure robust operation in diverse industrial environments.

This paper presents the Efficient feature Modulation and Aggregation Network (EMAN), a novel lightweight convolutional architecture integrating efficient attention mechanisms and structural optimizations, alongside EMAN-S, an ultra-compact variant with extreme parameter efficiency. Both models were initially evaluated on four datasets of varying complexity: RIAWELC, NEU-CLS, StitchingNet, and Ball Screw Drive Defect. On the original RIAWELC dataset, the proposed architecture was compared against nine state-of-the-art (SOTA) models. EMAN achieved 99.71% accuracy and EMAN-S attained 99.80%, surpassing the baseline models across most metrics. However, our analysis revealed severe data leakage within the original RIAWELC distribution, indicating that these exceptionally high baseline results were heavily biased.

To mitigate this issue, we evaluated a selection of the best-performing models on a rigorously filtered version of the RIAWELC dataset devoid of data leakage. Under these strict conditions, EMAN and EMAN-S demonstrated robust generalization, achieving accuracies of 82.56% and 79.99%, recall rates of 78.27% and 74.50%, specificities of 94.53% and 93.68%, precisions of 78.48% and 74.33%, and F1 scores of 77.97% and 74.29%, respectively. Furthermore, after re-implementing the best-performing literature models, we rigorously benchmarked them against our proposed architectures on this filtered dataset as well as the three other aforementioned datasets.

Across the remaining benchmarks, EMAN achieved 99.61%, 98.87%, and 99.73% accuracy on the NEU-CLS, StitchingNet, and Ball Screw Drive Defect datasets, respectively, surpassing all compared architectures. The slim variant, EMAN-S, followed closely with accuracies of 97.89%, 93.41%, and 99.18%, respectively, still outperforming or matching the top competitors despite utilizing a mere 145 k parameters. Overall, the results indicate the superior performance and strong generalization capabilities of both proposed architectures across all utilized datasets.

Regarding inference latency, both proposed models operate at exceptional speeds: following TensorRT optimization on an NVIDIA RTX 4090 graphics card, EMAN-S achieves over 800 FPS, while EMAN reaches 500–550 FPS. Even under CPU-only execution (using an Intel Core i7-14700K), EMAN-S sustains a throughput of over 360 FPS, confirming its suitability for deployment in embedded edge devices and on-site industrial inspection systems.

These results confirm the proposed architectures as effective, scalable solutions for real-time industrial defect classification, combining state-of-the-art accuracy with exceptional computational efficiency.

Visually impaired people (VIPs) encounter various challenges in their daily lives, and there is a need for portable, user-friendly device for real-time assistance to give them guidance regarding their surroundings. This article presents an artificial-intelligence-based innovative wearable assistive device—artificial intelligence (AI)-SenseVision—to analyze visual and sensory information about the objects and obstacles present in the scene to perceive the surrounding environment. The device is a complete amalgamation of sensor and computer-vision-based technologies that generate auditory information with the name of identified objects or audio warnings for detected obstacles. The performance of the trained deep-learning model is rigorously tested in complex and real-life scenarios using various statistical parameters for experimental validation. Moreover, the trained deep-learning models have been integrated into a low-cost single-board processor to make a standalone cost-effective device. All data processing is done within an optimized single hardware setup, and the user can easily access different modes, such as indoor and outdoor mode, while also enabling object counting in observed scenes. The proposed system has low-cost sensors, multiple operational modes, easy integration, and small volume, making this assistive device helpful for VIPs for independent navigation and collision prevention.

Image captioning, amalgamation of computer vision and natural language processing describe the contents of image in the form of natural language. Earlier methods focus on visual features and ignoring the relationship between different objects and regions for generation of corresponding captions. To address this concern this paper presents a novel transformer-based model, Higher-order Object Relation Network Transformer (HORN-T) to generate semantically and syntactically correct descriptions of images. The proposed model, for extraction of higher-order relationship between different objects and the scenes exploits appearance feature and spatial configuration. To further enhance the extracted features, Parallel Co-Attention (PCo-AM)-based transformer encoder is incorporated. Furthermore, for generation of coherent and diverse captions, the proposed model utilizes a triple attention (Convolutional Attention Module (CAM), Modified Multi-head Attention (MHA), and Position-aware attention (PAA)) based transformer decoder. To quantitatively and qualitatively evaluate the effectiveness of the proposed framework extensive experiments are conducted on three benchmarks for image captioning Flickr8K, Flickr30K, and MSCOCO. The proposed model provided state-of-the-art results on all the benchmarks by generating more informative descriptions with significant n-grams captured from a complex word space. An ablation is also carried out to study the influence different attention modules and parameters.

Accurate detection of simultaneous grip and elbow flexion is critical for advanced rehabilitation monitoring, intuitive prosthetic control and human-machine interaction. Existing approaches, particularly those relying solely on surface electromyography (sEMG), often struggle to capture overlapping muscle activations and the biomechanical context of dynamic upper-limb movements. This study proposes a multimodal, multi-label deep learning framework integrating sEMG, pressure distribution and inertial sensing for concurrent detection of grip and elbow flexion. Signals were acquired from three sEMG channels, four pressure sensors embedded in a Wii Balance Board and a nine-axis inertial measurement unit. All modalities were synchronized, standardized and segmented into fixed-length windows of 100 samples. A customized convolutional neural network and bidirectional long-short-term memory architecture with temporal attention was employed to model spatial-temporal dependencies across sensor streams. Multi-label sigmoid outputs enabled simultaneous prediction of both movements. Evaluation was conducted using subject-level cross-validation with file-level train-test splitting. The proposed framework achieved a window-level macro F1-score of 0.9573 with a Hamming loss of 0.0674 and file-level aggregation yielded a macro F1-score of 0.9783. Feature analysis indicated dominant contributions from inertial and pressure signals with sEMG providing complementary activation information. The proposed multimodal sensor fusion enables robust multi-label detection with attention-guided temporal weighting of complex upper-limb movements. This framework helps address key limitations of single-modality EMG systems and supports scalable real-time rehabilitation and assistive applications.

Introducing adaptability in image dehazing solution increases the scope to accompany a broader category of images, leading to performance improvement. Another important aspect is to correct color imbalance. We propose a dynamically adaptive prior named the ratio-prior, capable of screening an image for colorcast. Through simulations we demonstrate that in a hazy image, the assumption of inherent colorcast in a night-time scene and none in day-time, reported in recent image dehazing methods, is not always true. Secondly, ratio-prior serves as a parameter for enhancing image visibility and saturation. The proposed image enhancement model relies on an iterative method to generate controlled exposure images which other fusion-based solutions have overlooked. We also incorporate a multi-scale fusion framework which fuses the generated multi-exposure images by exploiting image features via weight maps. The bright weight map for instance reduces glare due to light sources in an image. The proposed approach has been benchmarked with other state-of-the art methods both qualitatively and quantitatively affirming its effectiveness. The quantitative evaluations demonstrate significant gains over recent baselines. On the O-HAZE dataset, the proposed method achieves a + 14.86 dB improvement in peak signal-to-noise ratio and a + 0.371 increase in structural similarity index measurement over multi-scale fusion, along with reduced perceptual error (LPIPS: 0.313 versus 0.517). Similarly, on the HazeRD dataset, it outperforms multi-scale fusion with a + 10.36 dB peak signal-to-noise ratio gain and a + 0.439 structural similarity index measurement increase.

In this proposal, a physics-driven methodology for secure key generation and image encryption is proposed by employing the underlying stochastic nature of optical wave propagation through atmospheric turbulence. An optical wave propagating through this medium suffers random phase and amplitude distortions, leading to intermittent intensity fluctuations at the receiver end. This phenomenon serves as a high-entropy source of randomness that has been exploited in the proposed method.

Building on this principle, a realistic optical propagation pipeline is developed by integrating turbulence-induced fluctuations, multiplicative distortion emanating from coherent interference, and additive noise. This results in a stochastic optical observation, and is further transformed into cryptographic binary sequences using a nonlinear deep quantization module. The robustness and statistical independence of the key generation mechanism is further enhanced by processing the quantized outputs using dithered thresholding, repetition-based error reconciliation, and cryptographic privacy amplification, resulting in a 256-bit secret key.

The generated bit stream is employed in a hybrid image encryption framework that utilizes Arnold transform-based permutation and a zero-propagation 2D stream-diffusion engine. This combination ensures good confusion and diffusion properties while localizing transmission errors during decryption. The generated key stream has been statistically validated using the NIST SP 800-22 test suite, the results confirms the presence of randomness. The encryption scheme achieves good entropy, negligible pixel correlation, and resistance to differential attacks, as demonstrated by NPCR and UACI values.

Liver cancer, a leading global cause of death, requires early and accurate diagnosis for better treatment outcomes and reduced mortality rates. This retrospective cohort study presents a machine learning-based predictive framework utilizing a combined panel of minimally invasive circulating biomarkers for early diagnosis and multi-grade prediction of liver cancer. This curated dataset is composed of urine/blood biomarkers like β-hCG, PD-L1 and Alpha-Fetoprotein. A hybrid and heterogeneous ensemble learning-based prediction technique has been proposed for the thorough analysis of collected data about the subject and associated biomarkers from highly diverse and broad age groups. The model's heterogeneity involved applying multiple distinct learning algorithms to the same training dataset, with the outcomes of each classifier used to expand the feature space and guide decision-making in subsequent stages. The proposed ensemble model is characterized by an iterative methodology incorporating both bagging and boosting techniques, ensuring adaptability and sustained enhancement, thereby surmounting limitations inherent in conventional static ensembles. The best hyperparameters using evolutionary algorithm achieved an accuracy of 93.33% and an F1-score of 94.29% on the test samples. The developed model accurately predicts liver cancer risk and grade, outperforming conventional methods. This AI model minimizes invasive procedures and healthcare costs, enhancing overall public health. The proposed minimally invasive biomarker-based framework may be particularly useful in resource-limited clinical settings where advanced imaging infrastructure and invasive diagnostic procedures are not readily accessible.

Schizophrenia (SZ) is a significant mental health illness that impacts people’s thoughts, feelings, and behaviour. It may cause a variety of hallucinations, delusions, and disorganised thought and behaviour, making daily life difficult. SZ patients require lifetime care. Early intervention improves the long-term prognosis by controlling symptoms before major complications occur. It has been determined that electroencephalography (EEG) is an effective biomarker for SZ detection. However, it is challenging to extract useful information from EEG signals because of their low intensity and instability. Thus, the vision transformer (ViT) that operates in the time-frequency domain is used in this work to automatically detect SZ. However, pure transformer designs frequently need a substantial quantity of training data or additional supervision. And, there is always a scarcity of datasets in the medical field. Therefore, we present a deep neural network (MGViT) incorporated ViT that combines the benefits of transformers in creating long-range dependencies and CNNs in extracting low-level features. Initially, the multichannel EEG signals are converted to recurrence plots (RP) to capture both spatial and temporal information. Later, RP images are processed through MGViT for early identification of SZ. The proposed approach is trained on two publicly accessible EEG datasets and obtained 98.92% and 88.17% accuracies, respectively. Further, the proposed work is contrasted with seven competitive techniques and achieved a better performance with an improvement of 1–26% in dataset-1 and 2–24% in dataset-2, proving its effectiveness in medical applications.