Facial expression recognition (FER) is the task of determining a person’s current emotion. It plays an important role in healthcare, marketing, and counselling. With the advancement in deep learning algorithms like Convolutional Neural Network (CNN), the system’s accuracy is improving. A hybrid CNN and k-Nearest Neighbour (KNN) model can improve FER’s accuracy. This paper presents a hybrid CNN-KNN model for FER on the Raspberry Pi 4, where we use CNN for feature extraction. Subsequently, the KNN performs expression recognition. We use the transfer learning technique to build our system with an EfficientNet-Lite model. The hybrid model we propose replaces the Softmax layer in the EfficientNet with the KNN. We train our model using the FER-2013 dataset and compare its performance with different architectures trained on the same dataset. We perform optimization on the Fully Connected layer, loss function, loss optimizer, optimizer learning rate, class weights, and KNN distance function with the k-value. Despite running on the Raspberry Pi hardware with very limited processing power, low memory capacity, and small storage capacity, our proposed model achieves a similar accuracy of 75.26% (with a slight improvement of 0.06%) to the state-of-the-art’s Ensemble of 8 CNN model.

A skin lesion is a portion of skin that observes abnormal growth compared to other areas of the skin. The ISIC 2018 lesion dataset has seven classes. A miniature dataset version of it is also available with only two classes: malignant and benign. Malignant tumors are tumors that are cancerous, and benign tumors are non-cancerous. Malignant tumors have the ability to multiply and spread throughout the body at a much faster rate. The early detection of the cancerous skin lesion is crucial for the survival of the patient. Deep learning models and machine learning models play an essential role in the detection of skin lesions. Still, due to image occlusions and imbalanced datasets, the accuracies have been compromised so far. In this paper, we introduce an interpretable method for the non-invasive diagnosis of melanoma skin cancer using deep learning and ensemble stacking of machine learning models. The dataset used to train the classifier models contains balanced images of benign and malignant skin moles. Hand-crafted features are used to train the base models (logistic regression, SVM, random forest, KNN, and gradient boosting machine) of machine learning. The prediction of these base models was used to train level one model stacking using cross-validation on the training set. Deep learning models (MobileNet, Xception, ResNet50, ResNet50V2, and DenseNet121) were used for transfer learning, and were already pre-trained on ImageNet data. The classifier was evaluated for each model. The deep learning models were then ensembled with different combinations of models and assessed. Furthermore, shapely adaptive explanations are used to construct an interpretability approach that generates heatmaps to identify the parts of an image that are most suggestive of the illness. This allows dermatologists to understand the results of our model in a way that makes sense to them. For evaluation, we calculated the accuracy, F1-score, Cohen’s kappa, confusion matrix, and ROC curves and identified the best model for classifying skin lesions.

There has been much recent work on fraud and Anti Money Laundering (AML) detection using machine learning techniques. However, most algorithms are based on supervised techniques. Studies show that supervised techniques often have the limitation of not adapting well to new irregular fraud patterns when the dataset is highly imbalanced. Instead, unsupervised learning can have a better capability to find anomalous and irregular patterns in new transaction. Despite this, unsupervised techniques also have the disadvantage of not being able to give state-of-the-art detection results. We propose a suite of unsupervised and deep learning techniques to implement an anti-money laundering and fraud detection system to resolve this limitation. The system leverages three deep learning models: autoencoder (AE), variational autoencoder (VAE), and a generative adversarial network. We preprocess the given dataset to separate the Transaction Date attribute into its base components to capture time-related fraud patterns. Also, Wasserstein Generative Adversarial Network (WGAN) is used to generate fraud transactions, which are then mixed with the base dataset to form a more balanced mixed dataset. These two datasets are used to train the AE and VAE models. We built two versions of the AE model (single-loss and multi-loss) besides a novel method of calculating the anomaly score threshold, called Recall-First Threshold (RFT), which helps enhance the model's performance. Experimental results demonstrated that the False Positive Rate (FPR) drops down to as low as 7% in the proposed multi-loss AE model. In comparison, we achieved an accuracy of 93%, with 100% of the fraud transactions recalled successfully.

Selective harmonic elimination (SHE) technique is used in power inverters to eliminate specific lower-order harmonics by determining optimum switching angles that are used to generate Pulse Width Modulation (PWM) signals for multilevel inverter (MLI) switches. Various optimization algorithms have been developed to determine the optimum switching angles. However, these techniques are still trapped in local optima. This study proposes an opposition-based quantum bat algorithm (OQBA) to determine these optimum switching angles. This algorithm is formulated by utilizing habitual characteristics of bats. It has advanced learning ability that can effectively remove lower-order harmonics from the output voltage of MLI. It can eventually increase the quality of the output voltage along with the efficiency of the MLI. The performance of the algorithm is evaluated with three different case studies involving 7, 11, and 17-level three-phase MLIs. The results are verified using both simulation and experimental studies. The results showed substantial improvement and superiority compared to other available algorithms both in terms of the harmonics reduction of harmonics and finding the correct solutions.

The rise in crime rates in many parts of the world, coupled with advancements in computer vision, has increased the need for automated crime detection services. To address this issue, we propose a new approach for detecting suspicious behavior as a means of preventing shoplifting. Existing methods are based on the use of convolutional neural networks that rely on extracting spatial features from pixel values. In contrast, our proposed method employs object detection based on YOLOv5 with Deep Sort to track people through a video, using the resulting bounding box coordinates as temporal features. The extracted temporal features are then modeled as a time-series classification problem. The proposed method was tested on the popular UCF Crime dataset, and benchmarked against the current state-of-the-art robust temporal feature magnitude (RTFM) method, which relies on the Inflated 3D ConvNet (I3D) preprocessing method. Our results demonstrate an impressive 8.45-fold increase in detection inference speed compared to the state-of-the-art RTFM, along with an F1 score of 92%,outperforming RTFM by 3%. Furthermore, our method achieved these results without requiring expensive data augmentation or image feature extraction.