Shahrukh Agha

Deep Neural Networks (DNN) have achieved state-of-the-art performance in medical image classification and are increasingly being used for disease diagnosis. However, these models are quite complex and that necessitates the need to reduce the model complexity for their use in low-power edge applications that are becoming common. The model complexity reduction techniques in most cases comprise of time-consuming operations and are often associated with a loss of model performance in proportion to the model size reduction. In this paper, we propose a simplified model complexity reduction technique based on reducing the number of channels for any DNN and demonstrate the complexity reduction approaches for the ResNet-50 model integration in low-power devices. The model performance of the proposed models was evaluated for multiclass classification of CXR images, as normal, pneumonia, and COVID-19 classes. We demonstrate successive size reductions down to 75%, 87%, and 93% reduction with an acceptable classification performance reduction of 0.5%, 0.5%, and 0.8% respectively. We also provide the results for the model generalization, and visualization with Grad-CAM at an acceptable performance and interpretable level. In addition, a theoretical VLSI architecture for the best performing architecture has been presented.

Motion Estimation (ME) and the two-dimensional (2D) discrete cosine transform (2D-DCT) are both computationally expensive parts of HEVC standard, therefore real-time performance of the HEVC may not be free from glitches. To address this issue, this study deploys the graphics processing units (GPUs) to perform the ME and 2D-DCT tasks. In this concern, authors probed into four levels of parallelism (i.e., frame, macroblock, search area, and sum of the absolute difference (SAD) levels) existing in ME. For comparative analysis, authors involved full search (FS), test zone search (TZS) of HEVC, and hierarchical diamond search (EHDS) ME algorithms. Similarly, two levels of parallelism (i.e., macroblock and sub-macroblock) are also explored in 2D-DCT. Notably, the least computationally complex multithreaded Loeffler DCT algorithm is utilized for computing 2D-DCT. Experimental results show that ME processing task corresponding to 25 frames, with each frame of size (3840×2160) pixels, is accomplished in 0.15 seconds on the NVIDIA GeForce GTX 1080, whereas the 2D-DCT task along with the image reconstruction and differencing corresponding to 25 frames took 0.1 seconds. Collectively, both ME and 2D-DCT tasks are processed in 0.25 seconds, which still leaves enough room for the encoder’s remaining parts to be executed within one second. Due to this enhancement, the resultant encoder can safely be used in real-time applications.

In this work, a meta-heuristic optimization based method, known as the Firefly Algorithm (FA), to achieve the maximum power point (MPP) of a solar photo-voltaic (PV) system under partial shading conditions (PSC) is investigated. The Firefly Algorithm outperforms other techniques, such as the Perturb & Observe (P&O) method, proportional integral derivative (PID, and particle swarm optimization (PSO). These results show that the Firefly Algorithm (FA) tracks the MPP accurately compared with other above mentioned techniques. The PV system performance parameters i.e., convergence and tracking speed, is improved compared to conventional MPP tracking (MPPT) algorithms. It accurately tracks the various situations that outperform other methods. The proposed method significantly increased tracking efficiency and maximized the amount of energy recovered from PV arrays. Results show that FA exhibits high tracking efficiency (>99%) and less convergence time (<0.05 s) under PSCs with less power oscillations. All of these methods have been validated in Matlab simulation software.

In this work high throughput and low power implementations of two dimensional Discrete Fourier Transform, Discrete Cosine Transform and MPEG Motion Estimation, for battery powered real time applications, using configurable multiprocessors, are proposed and compared in terms of speed, power, area and type of application. The results show the significance of multiprocessing in reducing the computational complexity of the application depending upon the type of application and the speed, area or power constraint.

Iris recognition technology recognizes a human based on his/her iris pattern. However, the accuracy of the iris recognition technology depends on accurate iris localization. Localizing a pupil region in the presence of other low-intensity regions, such as hairs, eyebrows, and eyelashes, is a challenging task. This study proposes an iris localization technique that includes a localizing pupillary boundary in a sub-image by using an integral projection function and two-dimensional shape properties (e.g., area, geometry, and circularity). The limbic boundary is localized using gradients and an error distance transform, and the boundary is regularized with active contours. Experimental results obtained from public databases show the superiority of the proposed technique over contemporary methods.