Alzheimer’s disease (AD) poses a significant global health challenge, making early and efficient detection essential for timely intervention. While three-dimensional (3D) T1-weighted MRI is widely used for AD detection, its clinical applicability is often hindered by high acquisition costs, prolonged scan durations, complex preprocessing pipelines, and the heavy computational requirements of volumetric deep learning models. Additionally, 3D models typically demand large annotated datasets and extended training time—constraints that limit their deployment in real-world, resource-limited settings. In contrast, two-dimensional (2D) T1-weighted MRI slices provide a cost-effective and computationally efficient alternative. These slices retain localized pathological features vital for diagnosis and can be leveraged effectively through modern representation learning. In practical settings, 3D scans may be missing or degraded, while 2D slices remain accessible and informative. This study introduces a hybrid deep learning-based framework for AD classification using 2D T1-weighted MRI slices from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). We introduce a series of methodological innovations, including advanced preprocessing techniques, weighted majority voting with confidence scores, the integration of statistical and pixel-based features, and a hybrid architecture that fuses pre-trained deep learning models with classical machine learning classifiers. Our best-performing model, the Weighted Augmented Hybrid Network (WAHNet), achieves an accuracy of 80.1
In the modern context, as the data generated is exponential, finding meaningful patterns from large datasets is an urgent need. A ‘Topic Evolution Model’ could generate the evolutions related to a topic of user interest and assist in the exploration of patterns. In a generic setting, the proposed ‘topic evolution model’ assists researchers and domain experts for the relevant information extraction on scientific field progress and innovations in technological field or domain from large archives. The evolution patterns uncover the emerging, decay/fading, peculiar, and long-lasting research topics, and subtopics. The performance evaluation on coherence metrics asserts that the proposed model significantly minimizes the domain expert user efforts in topic analysis, as evolving patterns easily reveal underlying statistical and machine learning details. The perplexity metrics highlights the capability of the topic model towards the cognitive view of the user, i.e., change of ideas and knowledge through a period of time reducing the citation bias.
In Magnetic Resonance Imaging (MRI), reduction in scanning time and high reconstruction quality are of paramount value. The recent sparsity based optimization techniques offer a way out to this end. The weighted norm minimization in Compressed Sensing (CS) has become popular due to its capability in providing adaptive sparse solutions. The ℓ _1-p , p>1 , minimization on the other hand has recently gained attention of researchers due to its superiority in dealing with highly coherent matrices. Intending to exploit the best of both, we propose a weighed ℓ _1-p minimization problem for p>1 . While presenting its solver, we demonstrate that the proposed optimization problem reconstructs an MR image faithfully from a small set of its Fourier samples. We compare and contract our method with its existing counterparts that use CS based ideas.
Facial expressions are utilized often in the day to day communication and are considered significant as they can mirror the internal emotional states of a person. Automatic Facial Expression Recognition (FER) systems aim at classifying the facial images into various expressions. To do this task accurately, better feature descriptors are to be developed to effectively capture the facial information. In this regard, novel local texture based feature extraction technique, Petersen Graph based Binary Pattern (PGBP), inspired by the Generalized Petersen Graph has been proposed. PGBP extracts three feature values in a 5 × 5 overlapping neighborhood. The experiments have been performed on MUG, TFEID and KDEF datasets with respect to six expressions in person independent setup. The experimental results demonstrated that the proposed method outperformed the existing methods in terms of recognition accuracy.
Deep Neural Networks (DNN) for resource-constrained mobile devices tend to reduce diagnosis latency and alleviate privacy concerns of the patients in the healthcare domain. Since mobile devices are limited by computational power and battery capacity, traditional DNN requires optimization. This paper analyzes the effect of compression techniques to detect Atrial Fibrillation (AFib) from Electrocardiogram signals. Specifically, we design a framework that compresses Residual Neural Network (ResNet) and Long Short Term Memory (LSTM) model using Weight Pruning, Weight sharing, Quantization, Huffman encoding, and Tensorflow Lite for AFib detection in android devices. The compressed models are evaluated on PhysioNet Computing in Cardiology Challenge 2017 database, and models are evaluated using size, inference time, precision, accuracy, sensitivity, and specificity. The models are tested on desktop computer and android smartphone.
The successful application of deep neural networks for solving complex tasks like image classification, object detection and segmentation depends critically on the availability of large number of labelled training samples. To achieve good generalization for a reasonably complex model with about 60 million parameters, as in AlexNet, one needs about one million labelled training samples. In almost all practical applications, like natural image classification and segmentation, plenty of unlabelled samples are available but labelling these samples is a tedious manual task. We introduce a novel mechanism to automatically label all the samples in an unlabelled dataset. Starting with completely unlabelled dataset, an iterative algorithm incrementally assigns labels along with a confidence to all training samples. During each iteration, 10–30 new representative samples are generated in a latent space learned using a variational autoencoder and labels for these samples are obtained from a human expert. The proposed idea is demonstrated on MNIST dataset without using the labels provided in the dataset. At regular intervals of training, the low dimensional latent vectors are clustered and only cluster centers are annotated. The manual labels of cluster centers are propagated to other samples in the cluster based on the distance and a confidence function. The loss function in successive training is modified to incorporate the manual information provided. We run multiple experiments with different choices of clustering algorithm, confidence function and distance metric and compare the results. With GMM clustering, best classification accuracy of 93.9
Data imbalance handling is important to have unbiased learning during model training of classification tasks. Synthetic minority over-sampling technique (SMOTE) is primarily used for data imbalance handling. Conventional SMOTE algorithm and its variants mainly deal with upsampling the class data that is based solely on the amplitude values of features and their neighbors. However, the spatio-temporal data corresponding to satellite remote sensing images, comes with the additional location information, i.e. longitude and latitude. This has to be incorporated in the data upsampling case in order to have semantically and physically useful data. Hence, we propose a new pipeline named, 'Spatial-SMOTE' to upsample the data by retaining the significance of spatial distribution aspect in the overall upsampling process. The effectiveness of this approach is shown on land-use land-cover classification task using the time series data for a particular study area. We identified the relation between different classes based on their semantic distances and formulated two cases- one with high semantic distance and other with low. It is to be observed that the proposed method of Spatial-SMOTEing the minority class works well for both the cases. We also tested the effectiveness of the proposed approach over synthetically induced class imbalances for both low and high semantically differing classes.
This paper proposes a local oriented structural information based scheme for developing weighted total variation mechanism to despeckle B-mode ultrasound images. This weighing mechanism is devised with a goal to preserve the structures present in the image while yielding a piecewise smoothness in the restored version. To achieve this goal, the weights are computed based on the direction of the geometry of the organs present in the image. The use of oriented structural information helps to control the smoothing around the edges depending on the direction instead of inhibiting it near the edges. Using two radiometric measurements, the developed algorithm's performance is compared to those of state-of-the-art algorithms. It has been observed from the results that the proposed algorithm attains better piecewise smoothness and high contrast in the restored images. Along with it, the structures are conserved in de-speckled images. The performance of the proposed algorithms can be corroborated from the qualitative and quantitative analysis.
Nowadays, the attention mechanism is widely used in image segmentation. Recently many attention gate (AG) based models have gained interest for use in the delineation processes. However, adding only AG in the models can sometimes produce fallacious results due to irregular lesion boundaries and illumination variation conditions. To address these issues, we propose two novel methods, Attention FocusNet and SE-Attention U-Net. Attention FocusNet focuses on the use of AG to get much improved hierarchical attention maps and lowers the details of irrelevant information in the model to create a segmented image mask. SE-Attention U-Net, on the other hand, re-calibrates the weights of the feature vector and pass it through AG to focus on relevant information and suppressing any superficial information. We experimented on three medical image datasets, which are ISIC 2017, ISIC 2018 and HAM10000. Validation results show that the proposed methods outperform the state-of-the-art methodologies and obtained better segmentation accuracy.
Electroencephalography (EEG) based emotion recognition has become a subtle research area because of its promising applications. An effective emotion recognition relies on significant and stable features. In this paper, we propose an EEG based emotion recognition methodology based on a hybrid feature extraction combined with Genetic Algorithm (GA) based feature selection. The features are extracted from three domains: time, frequency and discrete wavelet The proposal is evaluated on DEAP dataset where the emotional states are classified using a GA optimized Multi-Layer Perceptron. The proposed model identifies a. two classes of emotions viz. Low/High Valence with an average accuracy of 95.96% and Low/High Arousal with an average accuracy of 95.39%, b. four classes of emotions viz. High Valence-Low Arousal, High Valence-High Arousal, Low Valence-Low Arousal and Low Valence-High Arousal with 91.88% accuracy, which are better compared to the existing results reported in the literature.
Feature selection (FS) is an important pre-processing step in building machine learning models. minimum Redundancy and Maximum Relevance (mRMR) approach has emerged as one of the successful algorithms in obtaining irredundant feature subset involving only bivariate computations. In the current digital age, owing to the prevalence of very large scale datasets, an imminent need has arisen for scalable solutions using distributed/parallel algorithms. MapReduce solutions are proven to be one of the best approaches to design fault-tolerant and scalable solutions. This work analyses the existing Horizontal MapReduce approaches for mRMR feature selection and identifies the limitations thereof. It is observed that existing approaches involve redundant and repetitive computations and lacks a metadata framework to diminish them. This motivated us to propose Horizontal partitioning based MapReduce solutions namely HMR_mRMR, is an Iterative MapReduce algorithms and is designed under Apache Spark. Appropriate usage of metadata framework and solution formulation optimizes the computations in the proposed approaches. The comparative experimental study is conducted with existing approaches to establish the importance of HMR_mRMR.
Few-shot learning techniques enable deep learning models to identify novel classes from one or only a few examples of it. This is achieved by using the subconcepts learnt from pre-training on base classes that share some of the characteristics of the novel classes. Most of the existing approaches use episodic training or episodic fine tuning (N-way K-shot) to learn these subconcepts. However, most studies in literature require different training models that are tied to the different N-way K-shot test settings; N represents the number of classes and K represents the number of samples in each class. Second, these models are trained for a large number of episodes and the memory requirements are also high for increasing values of N and K. In contrast, we propose a method that combines a β -variational autoencoder with cosine similarity classifier trained only once in an end-to-end (non-episodic) fashion on the complete training set. The end-to-end training enables the model to learn disentangled content and style representation. After training, the weights of novel classes are generated by normalizing the encoder’s content representation. Now the model can perform any N-way K-shot classification. Unlike the majority of results reported in literature, our work includes the results for the case where the test set includes both base and novel class images. We tested our model on the Omniglot dataset. Our model performed on par with the state-of-the-art methods. In a few settings it also outperformed the existing methods.
With the ongoing development of deep learning techniques in recent years, the convolutional neural networks (CNNs) have shown remarkable performance breakthrough in remote sensing image scene classification. However, the performance of these deep models largely depends on the number of available training samples or labeled images. Although the knowledge transferring and pre-training techniques can handle such situation, these may become ineffective due to domain difference. On the other side, the existing data augmentation approaches often produce training samples with too low diversity to help in performance improvement. In order to address these issues, in this work, we propose PReLim as a novel modeling paradigm for remote sensing scene classification under limited labeled samples scenario. PReLim is based on the notion of local and global filtering of scene fragment mixture, which overcomes both the sample diversity and the domain difference issue. Experimental analyses with the benchmark UCMerced and SIRI-WHU datasets demonstrate the effectiveness of PReLim in achieving the state-of-the-art accuracy using limited number of training samples.
Restoring face occluded images is a challenging task as missing regions need to be filled based on existing surrounding information. Information extraction from a single image without exploiting context results in degraded quality of images. In this paper, we propose a Deep Convolution Generative Adversarial Network (DCGAN) based inpainting for restoring face occluded images using perceptual and contextual loss functions. The encoder of DCGAN is trained using a combination of context and prior losses. The generator of DCGAN is trained with random latent vector and occluded image encoding from the encoder to generate missing image region. The masking applied during training makes the method robust as the masks are applied at the center of the image, thereby occluding the most informative parts of the image. Experiments performed on CelebA dataset show that DCGAN successfully predicts masked missing regions, outperforming the state-of-the-art methods.
In fingerprint-based authentication system, cancelable fingerprint templates are generated to defend the fingerprint information. In this paper, we proposed a novel cancelable fingerprint template using Visual Secret Sharing (VSS). Using VSS, each fingerprint image is encrypted into different shares. Finally, these shares are preserved in distinct databases and treated as fingerprint template. Traditional VSS schemes are suffering from pixel expansion and contrast reduction. We have used grid-based VSS and data embedding mechanisms to succeed these limitations. The proposed fingerprint templates satisfy ideal properties of cancelable templates such as non-invertibility, diversity, and revocability without altering the performance of the authentication system. To enhance the speed of the template generation and reconstruction, we have used General Purpose Graphical Processing Unit (GPGPU) to fulfill the operations. The experimental evaluation validates that the reconstructed fingerprints have equivalent performance as the initial fingerprints with upgraded security.
Consumer review helpfulness has a significant role in purchase decision making in an online shopping environment. Deep learning modules with pre-trained word embeddings are predominantly used to asses review helpfulness. Pre-trained word embeddings are trained on generic corpora and lack in incorporating domain knowledge and sentiment information of a word. Moreover, pre-trained embeddings fail to capture the subtle change of semantics of same word with different parts of speech. In this work, we propose HORIE (Heplfulness of Online Reviews with Improved Embedding) which improve pre-trained embedding with domain, sentiment and parts of speech information and analyse helpfulness as classification problem. In HORIE, domain knowledge is acquired from domain specific corpora. The average of pre-trained and domain specific embedding is combined with vectorized sentiment information, extracted from lexical dictionaries, along with POS tag information. Later, we apply a dual CNN based model for classification of reviews. HORIE is tested with five different domain and compare our performance with existing embeddings. We also compare our approach with handcrafted feature sets and existing helpfulness classification technique. AUROC is used as a metric. Our approach shows improvement over existing approaches.
Graphic recognition in scanned floor plan images is crucial to understanding indoor scene layouts and their properties. Digitization of floor plans by scanning often introduces visible and abrupt disturbances. Recognizing symbols under such a scenario with higher accuracy has always been a subject of importance in the document image analysis community. Classical machine learning methods with hand-crafted features and deep neural networks were able to achieve the task with reasonable efficiency. This paper proposes FloorCaps, a hybrid network of the recently proposed Capsule network and VGG19. The proposed network takes advantage of the VGG19 network to learn the feature encoding from region images in a floor plan and uses the Capsule network for the classification or decoding task. FloorCaps can recognize and classify the region elements present and perform room labeling in a scanned floor plan image in an end-to-end manner with much higher accuracy than the state-of-the-art methods.
In this work, a lip movements related hybrid visual feature is proposed to analyze the Lombard effect in automatic speech recognition task. First, a robust lip contour method is employed to automatically detect the lip contours. Three vertical heights and width of the lip contour are considered as geometric features. By using Lucas-Kanade algorithm, the vertical velocity and horizontal velocity are estimated from eight points (end point of the vertical heights and width of the lip contour) between every successive frames. These motion parameters are concatenated to geometric lip features, results a hybrid visual feature. As compared to geometric features, the proposed hybrid visual feature improves the word recognition accuracy by 7.5
