The impact of convolution neural networks (CNNs) in the supervised settings provided tremendous increment in performance. The representations learned from CNN's operated on hyperspherical manifold led to insightful outcomes in face recognition, face identification and other supervised tasks. A broad range of activation functions is developed with hypersphere intuition which performs superior to softmax in euclidean space. The main motive of this research is to provide insights. First, the stereographic projection is implied to transform data from Euclidean space (R^n) to hyperspherical manifold (S^n) to analyze the performance of angular margin losses. Secondly, proving both theoretically and practically that decision boundaries constructed on hypersphere using stereographic projection obliges the learning of neural networks. Experiments have proved that applying stereographic projection on existing state-of-the-art angular margin objective functions led to improve performance for standard image classification data sets (CIFAR-10,100).

Plant diseases have become a major threat in farming and provision of food. Various plant diseases have affected the natural growth of the plants and the infected plants are the leading factors for loss of crop production. The manual detection and identification of the plant diseases require a careful and observative examination through expertise. To overcome manual testing procedures an automated identification and detection can be implied which provides faster, scalable and precisive solutions. In this research, the contributions of our work are threefold. Firstly, a bi-linear convolution neural network (Bi-CNNs) for plant leaf disease identification and classification is proposed. Secondly, we fine-tune VGG and pruned ResNets and utilize them as feature extractors and connect them to fully connected dense networks. The hyperparameters are tuned to reach faster convergence and obtain better generalization during stochastic optimization of Bi-CNN (s). Finally, the proposed model is designed to leverage scalability by implying the Bi-CNN model into a real-world application and release it as an open-source. The model is designed on variant testing criteria ranging from 10% to 50%. These models are evaluated on gold-standard classification measures. To study the performance, testing samples were expanded by 5x (ie, from 10% to 50%) and it is found that the deviation in the accuracy was quite low (0.27%) which resembles the consistent generalization ability. Finally, the larger model obtained an accuracy score of 94.98% for 38 distinct classes.

Cardiovascular disease are one of the leading causes of an increase in the mortality rate due to irregular heart beats. Photoplethysmogram (PPG) technique is one of the noninvasive evaluation of blood pressure (BP) offers a reliable, feasible, and cost-efficient solution than other conventional techniques. PPG technique is highly induced by motion artifacts and its characteristics depend on the physiological condition of the person. While the collection of data, the PGG must be calibrated. In this research, a novel approach using a dual-tree complex wavelet transform (DT-CWT) based feature extraction technique with GRU network for the classification of hypertension is proposed. DT-CWT gives shift invariance compared to Continuous wavelet transforms and dual tree structure helps to extract the real and imaginary coefficients of the features. DT-CWT helps to integrate the signal patterns even disintegrating them during testing procedure. Further, these extracted features are fed into a variant neural architecture consisting of sequential GRU layers stacked over fully connected dense layers. It is observed, utilizing GRU layers led to extract precise features for sequential signal data by out-performing existing models. The proposed model attained an state-of-the-art accuracy score of 98.82% on BIDMC-PGG dataset by overhauling existing loops in research. Finally, the larger model obtained an accuracy score of 94.98% for 38 distinct classes.

The epidemic of malaria is a death-dealing infectious disease caused by mosquito spreading across the world. In this technological era, automated diagnosis is worthwhile for accurate and faster solutions. In this work, a novel application-oriented diagnostic model was deployed to detect and classify thin-blood smear images infected with malaria. The Features from thin-blood smears are extracted using a series of Convolution-Neural-Networks and classified with novel Capsule-Networks by understanding the spatial relationship of thin-blood films. The experimentation is compared to VGG-16, ResNet-50, DenseNet-121 architectures with immense depth varying from 16 to 121 layers. The proposed Capsule-Network is 8 layer deep and tends to outperform attaining classification accuracy of 96.9% and specificity, sensitivity scores were 94.95%, 98.18%. This novel model was deployed as a web-application to act as an aid for such a havoc problem and to transfer applicability to every user in need by classifying images in less than 3 seconds.

Coronary heart disease (CHD) is one of the leading cause of an increase in the mortality rate. Various factors are influencing to estimate the presence of CHD in an individual biologically. We made inferences with these influencing factors and their peculiarity by analyzing with a various set of algorithms which help in obtaining a precisive decision support system for the presence of CHD in an individual. We considered a set of linear models (Logistic Regression, Linear Discriminant Analysis), Na¨ıve Bayes, Classification and Regression Trees (CART), Support Vector Machines (SVM). K-Nearest Neighbor (k-NN) for k=1 to 21. Next, we have considered a set of ensemble models. Furtherly, we computed a Multi-Layer Perceptron (MLP) and a Deep Neural Network to evaluate the performance through a deep learning approach. So, with this analysis, we found a linear model (Logistic Regression) tend to give on par results both in the case of Cleveland as well as Framingham dataset. This analysis led to design an intuitive approach for CHD classification and would give insights on how to use them in the medical research field.