Publications

Embedded Artificial Intelligence (EAI) integrates AI technologies with resource-constrained embedded systems, overcoming the limitations of cloud AI in aspects such as latency and energy consumption, thereby empowering edge devices with autonomous decision-making and real-time intelligence. This review provides a comprehensive overview of this rapidly evolving field, systematically covering its definition, hardware platforms, software frameworks and tools, core algorithms (including lightweight models), and detailed deployment processes. It also discusses its widespread applications in key areas like autonomous driving and smart Internet of Things (IoT), as well as emerging directions. By analyzing its core challenges and innovative opportunities in algorithms, hardware, and frameworks, this review aims to provide relevant researchers and developers with a practical guidance framework, promoting technological innovation and adoption.

Embedded Artificial Intelligence (EAI) integrates AI technologies with resource-constrained embedded systems, overcoming the limitations of cloud AI in aspects such as latency and energy consumption, thereby empowering edge devices with autonomous decision-making and real-time intelligence. This review provides a comprehensive overview of this rapidly evolving field, systematically covering its definition, hardware platforms, software frameworks and tools, core algorithms (including lightweight models), and detailed deployment processes. It also discusses its widespread applications in key areas like autonomous driving and smart Internet of Things (IoT), as well as emerging directions. By analyzing its core challenges and innovative opportunities in algorithms, hardware, and frameworks, this review aims to provide relevant researchers and developers with a practical guidance framework, promoting technological innovation and adoption.

Accurate storm surge forecasting is vital for saving lives and avoiding economic and infrastructural damage. Failure to accurately predict storm surge can have catastrophic repercussions. Advances in machine learning models show the ability to improve accuracy of storm surge prediction by leveraging vast amounts of historical and realtime data such as weather and tide patterns. This paper proposes a bidirectional attention-based LSTM storm surge architecture (BALSSA) to improve prediction accuracy. Training and evaluation utilized extensive meteorological and tide level data from 77 typhoon incidents in Hong Kong and Macao between 2017 and 2022. The proposed methodology is able to model complex non-linearities between large amounts of data from different sources and identify complex relationships between variables that are typically not captured by traditional physical methods. BALSSA effectively resolves the problem of long-term dependencies in storm surge prediction by the incorporation of an attention mechanism. It enables selective emphasis on significant features and boosts the prediction accuracy. Evaluation has been conducted using real-world datasets from Macao to validate our storm surge prediction model. Results show that accuracy and robustness of predictions were significantly improved by the incorporation of attention mechanisms in our models. BALSSA captures temporal dynamics effectively, providing highly accurate storm surge forecasts (MAE: 0.0126, RMSE: 0.0003) up to 72 h in advance. These findings have practical significance for disaster risk reduction strategies, saving lives through timely evacuation and early warnings. Experiments comparing BALSSA variations with other machine learning algorithms consistently validate BALSSA’s superior predictive performance. It offers an additional risk management tool for civil-protection agencies and governments, as well as an ideal solution for enhancing storm surge prediction accuracy, benefiting coastal communities.

In this article, we consider in improving the throughput of existing multicast routing protocols by using multiple multicast trees in ad hoc wireless networks. To achieve our goal, we propose virtual layer, a multi-layer container for multicast trees, that provides an scalable routing structure with transmission gain in multicast routing. Without modification of multicast routing protocols, we enable a layer concept into the protocols so that multiple multicast trees can be constructed in a multicast group. Although virtual layer improves throughput, excessive data traffic will be introduced into a network that may cause network congestion. To resolve this situation, we consider a random coding scheme, a practical network coding implementation that encodes data packets for routing in disjoint-path tree while avoiding congestion. We observed in our simulations that a transmission gain trades off the excessive traffic that causes network congestion.

Visible infrared person re-identification (VI Re-ID) is of particular importance for an intelligent safe-guard system, aiming to retrieve the same pedestrian from non-overlapping visible and infrared cameras. The VI Re-ID task is extremely challenging due to significant modality differences, high-sample noise, occlusions, etc. To address these issues, we explore a dual-space aggregation learning (DSAL) method that combines instance batch normalization (IBN) and residual shrinkage (RS) into a baseline model for feature learning and compression at the channel-level. The random erasing (RE) data augmentation method has been applied to preprocess the data. Experiments on two datasets demonstrate that: 1) IBN reduces shallow layer appearance differences and can bridge the gap between heterogeneous modalities; 2) The RS adaptive soft threshold sets the zero-domain features to zero to eliminate noise and clutter information, thereby enhancing the robustness of the network to noise; 3) RE data augmentation method significantly improves the model's generalization ability. Particularly, the design of DSAL can be seamlessly embedded into other CNN frameworks as a bottleneck variant without additional computation costs. Compared with the strong baseline, on SYSU-MM01, Rank-1, mAP, and mINP significantly improved by 10.66%, 7.78%, and 5.91%, respectively. On RegDB, Rank-1, mAP, and mINP significantly improved by 16.40%, 13.83%, and 19.07%, respectively.

Robots with computer vision and text recognition functions are widely used in industrial production, especially in highly automated factories. However, most robots have an excellent ability to recognize printed characters and show low accuracy in recognition of handwritten characters. Therefore, this paper considers recognizing handwritten text in the intelligent processing of handwritten documents. Its high accuracy prediction results are closely related to the effectiveness of manuscript input, intelligent translation, and intelligent scoring. Handwritten text is more difficult to recognize because it contains sequential information, and the images are more complex than single-character images. This paper proposes a new handwritten Chinese text recognition (HCTR) framework based on existing classical convolutional neural network (CNN) and recurrent neural network (RNN) algorithms. We use a handwritten Chinese text dataset from CASIA-HWDB containing numbers and symbols close to real application scenarios to train the model and compare the performance of various models, such as MobileNetV1 and MobileNetV2, with the proposed model. From the analysis of experimental results, it can be found that the proposed method can achieve higher performance with fewer parameters. In addition, we optimize the dropout rates of input blocks and obtain the best CER of our method is 6.11%.

The Building Safe Water Use Plan promoted by the Macao Marine and Water Bureau aims to encourage property management entities to regularly maintain building water supply systems to ensure the safety and stability of drinking water. However, traditional laboratory testing methods are often time-consuming and labor-intensive, making real-time and efficient water quality monitoring challenging. To address this issue, this study proposes a Raspberry Pi-based multi-sensor system for rapid water quality detection and improved monitoring efficiency. This system integrates multiple sensors to measure key water quality parameters, such as pH, total dissolved solids (TDSs), temperature, and turbidity, while recording data in real-time. The data were continuously collected over a period of five months (July to November 2024). The collected data were analyzed and validated using machine learning algorithms, including Isolation Forest, Random Forest, Logistic Regression, and Local Outlier Factor. Among these models, Random Forest exhibited the best overall performance, achieving an accuracy of 98.10% and an F1 score of 98.99%. These results show that the dataset demonstrates high reliability in anomaly detection and classification tasks, accurately identifying deviations in water quality. This approach not only enhances the efficiency of water quality monitoring but also provides technological support for urban drinking water safety management.

Modality differences and intra-modality variations make the visible-infrared person re-identification (VI-ReID) task highly challenging. Most existing methods f

Storm surge elevates the water level, resulting in floods that cause significant harm to our lives and property. Predicting possible occurrences of storm surge using machine learning technology has long piqued the interest of the different communities. However, it is observed that the accuracy rate of predictions is insufficient. This paper aims to develop a framework for evaluating the performance of various machine learning models. Various algorithms are being used to forecast storm surge incidents in Hong Kong and Kaohsiung. The inclusion of trend changes in atmosphere pressure and wind components throughout the implementation of the algorithms to the dataset gathered over the past is one of the innovative aspects of this work. These inclination characteristics are very vital and crucial to capture important correlations and features for the potential occurrence of storm surges. This study yields two major findings. First, even the basic method of an ensemble algorithm produces better results than any single algorithm. Second, determining the best performing ensemble models for those storm surge events is achievable based on the technique findings for the data analyzed.

Image-based deep learning method for plant disease diagnosing is promising but relies on large-scale dataset. Currently, the shortage of data has become an obstacle to leverage deep learning methods. Few-shot learning can generalize to new categories with the supports of few samples, which is very helpful for those plant disease categories where only few samples are available. However, two challenging problems are existing in few-shot learning: (1) the feature extracted from few shots is very limited; (2) generalizing to new categories, especially to another domain is very tough. In response to the two issues, we propose a network based on the Meta-Baseline few-shot learning method, and combine cascaded multi-scale features and channel attention. The network takes advantage of multi-scale features to rich the feature representation, uses channel attention as a compensation module efficiently to learn more from the significant channels of the fused features. Meanwhile, we propose a group of training strategies from data configuration perspective to match various generalization requirements. Through extensive experiments, it is verified that the combination of multi-scale feature fusion and channel attention can alleviate the problem of limited features caused by few shots. To imitate different generalization scenarios, we set different data settings and suggest the optimal training strategies for intra-domain case and cross-domain case, respectively. The effects of important factors in few-shot learning paradigm are analyzed. With the optimal configuration, the accuracy of 1-shot task and 5-shot task achieve at 61.24% and 77.43% respectively in the task targeting to single-plant, and achieve at 82.52% and 92.83% in the task targeting to multi-plants. Our results outperform the existing related works. It demonstrates that the few-shot learning is a feasible potential solution for plant disease recognition in the future application.