RISC-V is an open, modular instruction set architecture (ISA) that has gained increasing adoption in fields such as image processing, machine vision, and deep learning inference due to its scalability and flexibility. This paper provides a comprehensive review of the latest research on RISC-V, with a focus on its vector extensions, custom instruction set optimizations, and the design and implementation of related hardware accelerators. The study employs a hardware-software co-optimization approach, featuring the design of a lightweight convolutional neural network (CSANet) and an artificial intelligence image signal processing (AI-ISP) accelerator to enhance image construction and inference efficiency. The experimental methods include performance evaluations based on FPGA hardware platforms, with a quantitative analysis of computational bottlenecks in the CSANet inference process, optimizing convolution operations and data flow handling. This research provides comprehensive technical support for the application of RISC-V in high-efficiency image processing and deep learning inference, especially in low-power and edge-computing scenarios.

Convolutional Neural Networks (CNNs) have broad application prospects in computer vision, image processing, and other fields. However, their characteristics, such as high computational speed and large data volume, pose significant challenges for hardware implementation. This project aims to improve computational efficiency and reduce energy consumption by summarizing hardware-based convolutional neural network algorithms to meet real-time requirements. Additionally, it will summarize research methods for accelerating convolutional computation, pooling, and fully connected layers using different hardware platforms such as ASIC, FPGA, and GPU. This paper focuses on optimizing data flow, parallel processing, and memory architecture to reduce computation latency and energy consumption. To ensure network accuracy, methods such as quantization and pruning are employed to reduce model size and computational complexity. Literature indicates that custom hardware designs for convolutional neural networks significantly enhance performance and energy efficiency compared to traditional software implementations. This review aims to provide an efficient hardware acceleration method for practical deep learning algorithms and promote development in fields such as smart terminals and edge computing.

With the continuous development of the new generation of microelectronics, there are several different clock domains in the complex digital systems. In order to iron out the complications caused by data transfer and storage in different clock domains, the designer used asynchronous FIFO (first in first out) in the design to realize cross-clock communication, data buffer. By analyzing and studying a number of results on asynchronous FIFO, this paper provides a comprehensive overview of the results of asynchronous FIFO up to now. This paper mainly summarizes the study of using Gray code to solve the substable problem in asynchronous FIFO, using the empty-full flag bit technique for the empty-full judgment of reading and writing, and summarizes how to make the structure of cyclic Asynchronous FIFO, and regulating the Asynchronous FIFO depth to improve their performance. For the practical application of asynchronous FIFO-based, this paper summarizes the optimization techniques of UART communication protocol based on asynchronous FIFO. the use of handshake signals for asynchronous FIFO design based on handshake-synchronous cross-clock-domain transmission technology is also a widely used technique for asynchronous FIFO.

Nowadays, large to a variety of factories and aerospace fields, small to mechanical toys, robotic arm has widely used in all works of life. Because of the interest of robotic arms, during the summer vacation, we did some research on the robot arm in order to better understand the operating principle of the robot arm and how to execute the task. This paper discusses the body frame and kinetic principles of a six-degree swivel manipulator arm. We confine to the High-Level APIs that ROS provides and focus this article on the analysis of mechanical design in a robotic arm using inverse kinematics theories. These theories not only help us recognize what goes on in the movement patterns of a robot arm but also for its well-considered control. There are unavoidable slight deviations in the calculation of inverse kinematics and the catching of objects. Even though, through some theories and program, we have already manipulate the robotics to finished some tasks.

Speech recognition technology is user-friendly and enables machines to understand and process human language, converting spoken language into text. As a critical component in numerous applications, this technology facilitates natural, hands-free interaction, enabling individuals to communicate and operate devices seamlessly, thereby enhancing the convenience and accessibility of everyday life. Additionally, speech synthesis assists users in multitasking and offers benefits to the visually impaired. Translation applications enable users of different languages to communicate with each other through one-to-one language conversion in the program. Speech recognition technology has evolved from rule-based methods to modern deep learning models. This paper explores the development history of speech recognition systems, focusing on analyzing its key technical milestones and challenges. Through a combination of historical analysis and technical insights, this paper examines how algorithms such as deep learning and neural networks can significantly improve speech recognition accuracy. The paper concluded that while deep learning has significantly boosted performance, hurdles such as managing diverse accents and environmental noise persist, indicating that there is still potential for future advancements.

This paper studies an adaptive event-triggered (AET) control method for sampled-data systems to enhance communication efficiency and conserve resources. Initially, a dynamic event-triggering mechanism is introduced to minimize redundant data transmissions while maintaining system performance. Then, a state feedback control law is designed by optimizing the event-triggered threshold condition, with the system's stability verified using a Lyapunov function. The gain parameters are computed through a cone complementarity linearization algorithm, ensuring computational efficiency and robustness. Additionally, the proposed approach addresses the trade-off between communication resource utilization and control performance by dynamically adjusting the triggering conditions based on system states. This ensures a reduction in redundant transmissions without compromising system reliability. Finally, this method provides a systematic and effective framework for improving communication efficiency, with potential applications in industrial automation and resource-constrained networked control systems.

As live music became popular, many music platforms emerged, and the music industry experienced a dramatic development. To further develop the music industry, music trend prediction is necessary for music companies and composers. The paper aims to provide a detailed description of the current situation in the music industry and showcase the procedures and some common methods used by other researchers to predict music trends. Those methods are all based on training models, including big data algorithms making mathematical models, feature extraction from the songs, utilizing Self-Embedding Attention Layer (SEAL) framework and Graph Neural Networks (GNNs), and using the influence data set and Integrative Collective Music (ICM). These methods could predict the music trends on certain platforms or data sets. Meanwhile, they also have some drawbacks, such as a lack of data access, especially data with high quality, and the static characteristic since the trend of pop music always changes. Concerning the models created by those authors, some possible future perspectives, like establishing a public data set and further support for the music industry, are proposed, attempting to summarize the field of music trend prediction in current society.

In recent years, convolutional neural networks (CNNs) have achieved significant progress in the field of image recognition. However, their generalization capability and model complexity still require optimization in practical applications. This paper aims to further explore and optimize the performance of CNNs in image recognition tasks. First, the paper reviews the applications and techniques of CNNs in areas such as agricultural image recognition, animal recognition, smart city construction, and facial recognition. Second, the basic structure and training process of CNNs are introduced. A corresponding CNN model is constructed and trained using the MNIST handwritten dataset, achieving a test set accuracy of 98%, which demonstrates the effectiveness of CNNs in image recognition tasks. Finally, the feasibility and limitations of using advanced methods, such as residual networks (ResNet) and batch normalization, to address more complex image recognition problems are discussed.

Object detection, which detects and classifies objects, is widely used in different fields. One such field is in the medical field, and more specifically it can be used to detect molecules in microscopy images to improve blood test efficiency. Nowadays, plentiful hospital labs are still manually counting blood cells which is time-consuming, and it is likely to make manual errors. A solution to this problem is to enable accurate object detection specifically designed for identifying molecules. To achieve this goal, the comparison analysis of different YOLO models, one of the most popular models in object detection, on a blood cell dataset is valuable as it evaluates the models’ accuracy, which provides insight into their strengths and weaknesses. The paper considers three types of YOLO models, YOLOv5nu, YOLOv8n, and YOLO11n models, on detecting and classifying red blood cells (RBC), white blood cells (WBCs), and Platelets. The experiment results show that all three models have high precision and recall rates, which means that they can identify most of the molecules accurately. This indicates a promising future of integrating object detection on blood cell count to speed up medical analysis.

Nowadays, along with the trend that more and more devastating tropical cyclones are happening all over the world, people’s are facing serious threat. Since traditional models have trouble giving more accurate prediction results, ML models are introduced to provide a more effective way. This overview briefly summarizes the history of ML and the causes of TCs, giving some algorithms of ML that were applied to TC prediction. It also included two actual examples of ML methods that had made a success on predicting TCs. Despite the fact that challenges exist in data quality and computational resources, machine learning models have proven to have huge potential in improving the accuracy and efficiency of tropical cyclone predictions. What’s more, this overview provides some possible prospects of the field, too, including model optimization, risk assessment, and interdisciplinary collaboration. These urgently-needed advancements are essential for improving the resilience of coastal rigions abilities to deal with TCs and the disaster caused by it.