Image dehazing has been used in many applications, such as underwater imaging, autonomous, aerial, and drone vehicles, as well as surveillance and security. While it is common today to use combinations of CPUs and GPUs in heterogeneous computing environments, we present in this work the development of a parallel hybrid approach using CUDA and OpenMP to enhance the efficiency of the dark channel prior algorithm for haze removal. Implementation techniques and result discussions regarding program improvements may be used to support parallel developers. Preliminary studies, results, and experiments on hybrid CUDA and OpenMP have shown promising outcomes.
Optimising the execution of Java bytecode on embedded systems is essential for various applications, considering the limitations of their resources, such as storage and processing power. In this regard, using the conventional profilers available to analyse bytecode, identifying basic blocks, determining frequencies, and defining new superinstructions for every basic block is cumbersome and time-consuming. This paper presents a novel approach to enhance the performance of embedded systems through the integration of superinstructions into Java bytecode. The proposed approach performs bytecode analysis, identifies basic blocks along with the specific candidate instructions suitable as superinstructions through an efficient profiler, and subsequently replaces these instructions with optimised superinstructions. The instruction set enhancement and reduced stack operations result in better execution performance of the applications running on embedded systems. The approach implemented in our prototype framework, JEOPT, has been used to optimise the execution of complex mathematical benchmarks including matrix multiplication and LU factorisation. The experimental results demonstrate that our proposed approach significantly improves the execution time of matrix multiplication and LU factorisation up to 4.65% and 2.12%, respectively.
Mixed integer linear programming (MILP) model is one of the best tools to search differential trails for bit-oriented block ciphers. Based on this method, it is easy to evaluate block ciphers against differential attack. This paper focuses on the research to search for better distinguishers of block cipher PRESENT-80 and Keccak-f[400]. The 14/15/16-round differential distinguisher of PRESENT-80 with the probability of no less than 2(-71)/2(-75)/2(-79) is obtained. A 17-round differential path (the start with two active S-boxes) with no less than 2(-76) is shown. For Keccak-f[400], the 3/4/5-round differential characteristics with the probabilities of 2(-156)/2(-284)/2(-385) are presented. Better differential cryptanalysis results can be obtained based on the above superior distinguishers. This paper provides a reference to find better and more accurate differential paths to attack block ciphers, and presents a detailed calculation process of them. The limitation of this paper lies in the fact that the current distinguishers have not been used for specific cryptanalysis.
Longer basic blocks can provide greater opportunity for the instruction-level parallelism of loops on very long instruction word (VLIW) architectures. This paper proposes a compilation method of loop linearisation (LLU) for loops with fixed iteration number. The method can eliminate the loop structure from corresponding input loops by rearranging the instructions of loop iterations, and it removes all the instructions for loop control. These make the subsequent instruction scheduling more effective. This paper investigates the problems of auto analysing the number of loop iterations and the corresponding code generation method, and proposes the algorithms to solve them. Experimental research is conducted on the FT-M7002 hardware platform. The results show that compared to loop unrolling, the proposed method achieves a minimum average performance improvement of 7% and a maximum improvement of 28%. For data-intensive computing circumstances, this method can be used to accelerate the response speed of related applications.
This paper presents an IoT-enabled post-harvest monitoring and decision support system for perishable crops, such as onions, using real-time environmental sensing and a Bayesian network model to forecast spoilage. The system integrates multiple sensors to continuously monitor storage conditions and crop health. The data collected are fed into the Bayesian network, enabling probabilistic reasoning about spoilage risk and profitability. In parallel, machine learning models linear regression (LR), support vector regression, and artificial neural networks were trained to predict weight loss, with LR achieving the lowest mean absolute error (1.61). The system demonstrated a 94% probability of preventing rot and a 79% probability of ensuring profitability under optimised conditions. Furthermore, intelligent alarms and blower control based on real-time sensor input significantly reduced both rot risk and energy consumption. The framework can be adapted to various perishable agricultural commodities.
Though vector digital signal processors (DSPs) have abundant hardware resources for data processing, word alignment is typically required by their vector memory access and data transmission components. However, in some DSP applications which process half-word data blocks, word alignment requirement may be not satisfied because there is an interval also in half-word between adjacent valid data elements. Thus, for the highly effective processing of half-word data blocks in a vectorised way under the word alignment requirement, this paper proposes a DMA and half-word vector shuffling (DHWVS) co-processing method. The DHWVS co-processing method is classified according to the adjacent data stride and the alignment characteristic of data block address. Besides, the DMA function and vector shuffle function of vector DSP are used to reduce the demand of memory accessing. The experiment results show that the performance of the algorithms optimised by our method is significantly improved, and the maximum speedup ratio reaches 16.1.
In response to the problem that existing teaching quality assessment techniques cannot comprehensively evaluate teaching quality from multiple dimensions, this study uses class distance and approximate complete binary tree generation strategies to reduce the variance of sample data, and establishes an intelligent classroom teaching quality evaluation model based on distance binary tree support vector machine. The results showed that the classification accuracy of the proposed algorithm on three datasets was 97.32%, 98.5%, and 97.46%, respectively, and the training time was 23.2 ms, 23.7 ms, and 23.8 ms. The accuracy of the proposed algorithm was 0.71, and the area under curve was 0.92. When the data reached 400, the training and testing time of the proposed algorithm was 16.3 ms and 4.0 ms, respectively. The research results contribute to improving the scientific and reliable evaluation of teaching quality, thereby helping teaching managers better understand teaching effectiveness and improve teaching quality.
This paper addresses the challenges of privacy-preserved federated learning (PPFL) in the internet of things (IoT) by introducing Multi-SparseAgg, an innovative framework designed for efficient, secure multi-round aggregation. Existing secure aggregation protocols in PPFL struggle with scalability, latency, and the communication overhead associated with frequent model updates across numerous, resource-limited IoT devices. Multi-SparseAgg tackles these challenges by employing sparse neural networks optimised with binary masks, significantly reducing communication costs without sacrificing model accuracy. A one-time setup phase generates reusable secrets, eliminating the need for costly reinitialisation in each round and enabling robust aggregation even with intermittent client participation. Experimental results on a benchmark dataset demonstrate that Multi-SparseAgg significantly lowers communication costs on the client side by 8.2% to 69.1% and on the server side by 7.1% to 28.6%, compared to baseline methods. It also reduces computational overhead by 21.1% to 77.6% while preserving model accuracy and ensuring fast convergence.
This work proposes a novel sophisticated mathematical model for the device. Initially, pre-acquired experimental data are collected, and a database is constructed. A variety of factors, such as gate length, electric field, gate width, drain-source voltage, and gate-source voltage, are included in the previously obtained experimental data. The B-Spline interpolation technique is used for the data augmentation process. An improved MobileNet model that defines three distinct models is then used to train and evaluate the supplemented data. The improved MobileNet model helps the network improve the training speed and accuracy. The training is executed iteratively till a minimal error is obtained between actual and predicted values. This work improves predictive accuracy by utilising experimental data enabling more dependable simulations that are essential for enhancing device performance and design. This extends beyond improved FinFET modelling; it sets a precedent for integrating advanced machine learning techniques in semiconductor technology.
Due to its rapid spread and associated symptoms, conventional methods of prevention and treatment have often proven inadequate to manage SARS-CoV-2, as conventional approaches show limited effectiveness in many scenarios. The ability of the virus to transmit rapidly, even to asymptomatic individuals, and the severity of symptoms in some patients have put significant pressure on healthcare systems. Against this backdrop, the work described aimed to develop an expert system for personalised recognition and treatment of SARS-CoV-2. A clear methodology was defined, supported by conceptual diagrams to outline the logical flow of the problem and proposed solutions. This methodology is based on key components: a Bayesian network to calculate the probability of infection by analysing the patient's symptoms, contacts, and geographic context; and a semantic component to determine the most appropriate treatment using the patient's clinical and personal information, such as allergies or individual risk factors. Once the solution was defined logically, we moved on to formalising the components and designing the workflows, which were implemented using appropriate technologies and open data. Testing of the system carried out through real data simulations, confirmed the system's ability to provide customised patient responses.
This work proposes an AI-IOT based transmission line fault detection comprising cloud server and artificial neural network. In this proposed work, current sensing panel along with Arduino panel are incorporated to capture the data from the transmission line and sent to central control system using WiFi module. Additionally, this topology enables real-time data transmission, enabling continuous monitoring of device faults from any remote location by server or application. However, the faulty and non-faulty data (in terms of current) are further trained using artificial neural network-based model; so that any data which are captured through IoT network can be directly analysed to predict whether it is faulty or not. Thus, a compact IoT cloud server-based system with minimal hardware requirement can be designed with incorporation of ANN model. The performance has been shown in terms of data received via Arduino, as well as IoT cloud server. Finally, performance analysis in terms of fault detection accuracy and computation time has been presented to justify the superiority of the proposed network compared with the existing works available in the literature.
Cloud computing significantly improves the performance of robots in data processing and storage, but still faces problems such as high computational loads and high energy requirements for local robots. To address these issues, a protocol-fused physical framework integrating multiple sensors is proposed to simplify sensoring data integration and device deployment. Cloud robotics intelligent workload optimisation has also been achieved through accurate sensoring data collection based on this framework. First, a middleware called ProtoFusion is introduced to manage the robot's local services, facilitating protocol conversion and transmission of multimodal sensory information. Next, the cloud robot's physical framework, based on ProtoFusion, enables sensing, perception, and control. Finally, ProtoFusion's task division (e.g., receiving, sending, and controlling) is scheduled using uC/OS-III, optimising system resource utilisation. The effectiveness of the optimisation is verified experimentally. Resource efficiency was improved, energy consumption was reduced and system reliability was enhanced.
Cloud computing is a promising technique to conquer the resource limitations of a single mobile device. To relieve the work load of mobile users, computation-intensive tasks are proposed to be offloaded to the remote cloud or local Cloudlet. However, these solutions also face some challenges. It is difficult to support data intensive and delay-sensitive applications in the remote cloud, while the local Cloudlets often have limited coverage. When both of these methods cannot be supported, another option is to relieve the load of a single device by taking advantage of resources of surrounding smart-phones or other wireless devices. To facilitate the efficient operation of the third option, we propose a novel pervasive mobile cloud framework to provide an incentive mechanism to motivate mobile users to contribute their sources for others to borrow and an efficient mechanism to enable multi-site computation partition. More specifically, we formulate the problem as a Stackelberg game, and prove that there exists a unique Nash equilibrium for the game. Based on the unique Nash equilibrium, we propose an offloading protocol to derive the mobile users' strategies. Through extensive simulations, we evaluate the performance and validate the theoretical properties of the proposed economy-based incentive mechanism.
Medical image fusion is dedicated to extracting structural and functional information from medical images. However, existing medical image fusion methods usually rely on convolutional operations and ignore long-distance information transmission. To address this problem, we propose a dilated convolutional attention transformer network (DCATNet) for medical image fusion. Specifically, to enhance the long-term dependence of the network on the input image, a transformer (TF) module is built. At the same time, a dilated convolutional channel attention (DCCA) module is built to realise the accurate extraction of feature and multi-scale information. This module combines the CPA module with the expansion convolution to enhance the robustness of the model. This enables the proposed method to handle the complexities of long-distance information transfer without losing important contextual and structural details. Experimental results demonstrate that the DCATNet outperforms competitors and proves its potential in medical image fusion for long-distance information transfer processes. Meanwhile, the results highlight the potential of DCATNet to advance medical image fusion, and it can lead to better clinical outcomes and more accurate diagnoses.
