The world's population mostly obtains its food from plants. Plant diseases, which may be prevented with continuous observation, are a contributing factor in productivity loss. Monitoring plant diseases manually is time-consuming and prone to mistakes. To overcome this problem, this research endeavours to develop an advanced tomato leaf disease detection system, aiming to address the pressing challenge of crop health deterioration. The methodology involves a comprehensive approach, starting with a robust pre-processing pipeline utilizing Non-Local Means (NLM) for image cleaning, standardizing image sizes, and normalizing pixel values to enhance model convergence. A novel Improved Laplacian of Gaussian (ILoG) segmentation algorithm is proposed for precise leaf region extraction, while advanced texture analysis methods such as Gamma Correlation-based Local Phase Quantization (GC-LPQ) and Local Directional Patterns (LDP) are explored for feature extraction. Additionally, colour-based features, including colour moments, and geometric features like leaf area, perimeter, and shape descriptors, are extracted. The research delves into effective feature selection using the Hybrid FoxWheel Optimization Algorithm (HFWOA), a combination of Binary Waterwheel Plant Optimization Algorithm (BWPOA) and FOX-Inspired Optimization Algorithm (FIOA). The Unified Learning Ensemble (ULE) approach integrates Decision Trees, Support Vector Machines (SVM), and Neural Networks with weighted voting for accurate leaf disease classification. The methodology's efficacy is evaluated through cross-validation, achieving an accuracy of 99.29% with an 80:20 training-testing split. The implementation is publicly available in a GitHub repository https://github.com/psaranyadevi27-ui/UNIFIED-LEARNING-ENSEMBLE-FORROBUST-LEAF-DISEASE-DETECTION
The increasing trend towards the adoption of transformers in edge devices stipulates the need for ultralow power devices that also require high energy efficiency with power budget constraints. This article proposes a near-threshold SRAM compute-in-memory transformer accelerator with adaptive precision scaling for sub-mW edge AI applications. The architecture performs MAC operations directly inside a modified 6T SRAM cell, minimizing data movement between compute and memory. Operating at 0.45 V reduces total power consumption by 58% compared with conventional digital MAC accelerators running at 0.9 V. An adaptive precision precision mechanism has been implemented to switch between the INT4 and INT8 precision dynamically based on the sensitivity of the layer, resulting in an additional 27% savings with less than 1.2% degradation on TinyML transformer benchmarks. The proposed accelerator is implemented in 28 nm CMOS. All reported results (power, energy efficiency, MAC linearity, and noise margins) are obtained from post-layout simulations using TSMC 28 nm LP foundry models and SPEF-extracted parasitics; they are not silicon measurements. The design shows an energy efficiency of 12.4 TOPS/W with a total power of 0.82 mW at 50 MHz.
Image enhancement is a vital field in image processing; that aims to improve the visual quality of images for best interpretation, analysis and decision making. This paper presents the basic structure of deep learning networks, comprehensive overview of the state-of-the-art image enhancement deep learning technique, motivation behind the selection of FPGA as implementation processor, and an overview of hardware component improvements; also presenting challenges to conventional approaches and the solutions to them. Deep learning approaches, particularly convolutional neural networks (CNNs), generative adversarial networks (GANs) and autoencoders, have revolutionized the field by learning complex features and providing adaptive, data-driven enhancement solutions. The review highlights the effectiveness of deep learning approaches in various application domains, such as medical imaging, satellite imagery, low-light photography and video enhancement. With FPGA offering offering basic logic cells with accelerators like LUTs, DSPs, etc. to configure any instruction set; making it a favorable co-processor (truly unlimited processor because it can be configured to perform any instruction); we further explore the challenges in implementing deep learning methods, such as computational complexity, real-time processing and the need to process large datasets. Finally, we discuss emerging trends and future directions in the field such as bringing in published approximators to models used in deep learning algorithms to make them implementable on FPGA. This review aims to provide a valuable resource for researchers and practitioners seeking to understand the landscape of image enhancement methods and their practical implementation.
Hyperspectral images are essential for remote sensing applications in the environment, surveillance, and precision agriculture. In recent years, convolutional neural networks (CNNs) have demonstrated remarkable results in computer vision applications, such as hyperspectral image classification. This research proposes a novel Deep Learning (DL)-based approach to hyperspectral image classification. Initially, spectral and spatial smoothing is performed using Savitzky-Golay and median filters, respectively, in the pre-processing stage. After that, a Multi-scale Residual Deep Convolutional Neural Network (MSRDCNN) and Stacked Bidirectional Long Short-Term Memory (SB-LSTM) are utilized to retrieve spatial and spectral attributes. Then, the Chaotic Dwarf Mongoose Optimization Algorithm (CDMOA) is used to select features. The Modified Prairie Dog Optimization (MPDO) algorithm is used to enhance the performance when a Global Pooling Dilated Convolutional Neural Network (GPDCNN) is employed for categorization. Comprehensive analyses show the proposed approach performs better than the state-of-the-art models.
Miniature Reaction Wheel Brushless Direct Current (MATLABDC) motors are widely employed in satellite attitude control systems because of their compact size and accurate torque control. Direct torque control (DTC) has a fast response time and torque accuracy potential, but the time-averaged torque ripple around zero and the control voltages for each phase result in a complex control effort when used in space applications. To address these issues, this paper proposes an optimized DTC strategy employing a nonfragile proportional integral derivative acceleration controller for enhanced satellite attitude performance. The major objective of this work is to minimize torque ripple and improve the dynamic response of the miniature reaction wheel brushless direct current motor, thereby enhancing speed regulation and reducing torque fluctuations for superior motor performance. A Bitterling Fish Optimization Algorithm (BFOA) is employed to minimize both torque ripple and speed errors of the motor drive system. The proposed system consists of a miniature reaction wheel brushless direct current motor, a nonfragile proportional integral derivative acceleration controller, a pulse width modulation inverter, a hall sensor feedback system and a voltage source inverter, and is implemented in MATLAB. Comparative analysis with existing methods such as the Whale Optimization Algorithm (WOA), Particle Swarm Optimization (PSO) and Genetic Algorithm (GA) confirms the superior performance of the BFOA in optimizing torque control and minimizing speed error. The proposed technique minimizes torque ripple and speed error while significantly improving the dynamic performance and stability of miniature reaction wheel brushless direct current motors under varying operating conditions. Its effectiveness makes it well-suited for high-precision applications such as satellite attitude control.
