Blood pressure is a critical indicator of human health, making convenient and easy-to-use monitoring methods a significant research focus. Existing blood pressure monitoring techniques often rely on additional sensors or devices, imposing both usage and economic burdens on users. To tackle this problem, we propose MetaBP, a blood pressure monitoring system that leverages the camera on smart devices. The basic idea is to record a video when the fingertip covers the camera to capture the tiny skin color changes caused by the heartbeat and extract pulse signals from the video frames. Furthermore, MetaBP employs a meta-learning approach, enabling the model to quickly adapt to the blood pressure monitoring task with minimal data and develop personalized models. By leveraging meta-learning, the model can efficiently learn the unique features of an individual’s physiological signals, providing accurate and personalized blood pressure estimation even with limited training samples. To obtain accurate measurements, we use a variational mode decomposition (VMD) method to reduce signal noise, ensuring that the pulse signals are both reliable and precise. To evaluate the robustness of MetaBP, we conduct experiments with 30 participants and implement MetaBP on commercial devices with camera parameters. The results demonstrate that MetaBP can accurately estimate systolic and diastolic blood pressure, with mean errors of 1.37 and 0.82 mmHg and standard deviations of 7.39 and 5.83 mmHg. Furthermore, our results demonstrate the feasibility of using widely available smart device cameras for reliable blood pressure monitoring, offering a practical solution for continuous health tracking.

Pest detection is critical for achieving effective pest control. However, the current deep learning-based pest detection algorithm is unsuitable for deployment on resource-limited edge devices due to its extensive computation and long inference time. Although lightweight models have been widely used for practical detection, their insufficient feature extraction capability leads to a decline in detection accuracy. This paper proposes a fast algorithm for crop pest detection based on lightweight feature extraction and knowledge distillation. Firstly, we introduce partial convolution and propose a lightweight feature extraction module, C3Faster, which reduces the model's computation and speeds up model inference while ensuring effective feature extraction. Secondly, we use knowledge distillation to improve the model's detection accuracy by using teacher networks to assist in training. Finally, we created a dataset, CropPest6, consisting of six crop pest categories and conducted experiments. The experimental results demonstrate that our method reduces the detection time, number of parameters, and computation by 17%, 38%, and 44%, respectively, compared to the baseline model. Furthermore, our method achieves 93.9% Precision, 93.6% Recall, and 97.5% mean Average Precision (mAP), demonstrating its practical suitability for fast crop pest detection.

In recent years, wireless networks and applications have grown rapidly and converged across a wide variety of scenarios. More and more applications require wireless networks for high bandwidth and low latency. However, due to the attenuated propagation of wireless signals, bandwidth changes rapidly in a short period. TCP fails to work properly in such an environment and suffers from low network link utilization and high latency. To solve above problems, this paper proposes a receiver-driven congestion control framework, named NUiVa. NUiVa decouples the congestion avoidance phase of sender side congestion control and implements it on the receiver side. In addition, NUiVa uses one-way delay to detect network congestion and controls the sending rate of senders via the receiving window field in the packet header. We confirm that the throughput degradation caused by network flips can be mitigated by NUiVa. And the throughput of data transmission can be further improved by the design of the receiver’s algorithm. The evaluation results show that NUiVa improves the throughput of TCP stream by 10 to 23 percent in most cases and reduces the queuing delay by an average of 29 percent.

Water segmentation is essential for autonomous navigation of Unmanned Surface Vehicles (USV) in inland waterways. However, most of the existing methods are suitable for maritime environments. Due to their poor real-time performance and a high false-positive rate of water surface reflection and water-sky interference in inland waterways, we propose a radarassisted real-time water segmentation network (RaRWS) to solve the above problems. We obtain pseudo-waterlines based on radar data by deleting points, fitting them multiple times, and generating a radar mask. While simplifying the encoder backbone network, we use the Attention Refinement Module (ARM) to fuse radar masks to improve detection and waterside accuracy. In addition, a Feature Fusion Module (FFM) is introduced to help the decoder fuse high- and low-level features and further fuse radar data. RaRWS is tested in both normal and bad weather. The results show that RaRWS can achieve higher performance compared with the current state-of-the-art methods (F1is above 99%), while gaining real-time performance (41.3fps).

With the annual increase of malleable iron (MI) production, MI factories have to employ almost three-quarters of workers to inspect unqualified products. To promote the automation of the quality inspection process of MI products, many convolution neural networks have been proposed to address such defect detection tasks. However, they sometimes fail to achieve the expected detection results due to the different defect presentation patterns on MI surface. In this work, we proposed a lightweight neural network model ES-MobileNet* fused with the attention mechanism, specifically designed for the MI defect detection on edge computing devices. Our ES-MobileNet* has two main parts: the expansion and shrinking bottleneck (ESNet) and the selective kernel (SK). Functionally, ESNet includes two feature extraction branches, one of which is responsible for extracting the features of the internal defects of MI, and the other one is responsible for extracting the features of the edge contours of MI. Then, SK will recalibrate the output features of two branches before feature fusion that suppress interference information in the channel domain of the original feature map. Experiment result shows that compared with MobileNetV2, the computational complexity and parameter of the proposed model are reduced by 28.6% and 20.2% respectively. It has a mAP of 62.45% on dataset IIDD and a FPS of 78.7 at Nvidia Jetson NANO, which can more than enough provide technical support for the surface defect detection of MI at edge computing devices.

Rivers are an essential part of the aquatic environment. The accurate detection and timely cleaning of river floating debris plays a vital role in the landscape and aquatic ecological environment. At present, the detection of river floating debris mainly relies on manually patrolling rivers and fixed camera monitoring, which has the problem of low efficiency and high cost. In this paper, we used Unmanned Aerial Vehicle (UAV) to photograph rivers and then utilized deep learning algorithms to detect floating debris in the images to improve the efficiency of river regulation. However, its application faces some problems, including the lack of datasets, the complex background of UAV images, and the small and sparsely distributed objects. To address the above issues, we used a UAV to capture river images and labeled the floating debris to construct a dataset. Moreover, to enhance the detection capability of river floating debris, we proposed an improved algorithm based on YOLOv5s. Firstly, the algorithm adds a microscale detection layer in the detection phase to improve small targets' detection. Secondly, it introduces an improved CBAM in the feature fusion phase to suppress the effects of useless and complex background information. Finally, in the loss function, a weighting factor is added to objectness loss to raise the loss weight of positive samples to ameliorate where negative samples loss much more than positive samples. The experimental results illustrated that compared to the baseline, our method has superior precision, recall, and [email protected], reaching 87.4%, 85.6%, and 91.8%, respectively. The proposed method can accurately detect river floating debris in UAV images and provide technical support for river regulation.

Acoustic-communication-based intelligent devices attract attention for their simplicity and cost-effectiveness. Still, there is a lack of research on acoustic communication solutions that can simultaneously balance low Bit Error Rate(BER), inaudible sound, low power consumption, high throughput, and operate on embedded devices. This paper proposes an acoustic communication framework based on Chirp Spread Spectrum (CSS) modulation technology, which can accomplish reliable acoustic communication on common MCU such as stm32 and ESP32. Our proposed Custom Rolling Matching Encoding (CRME) protocol matches complex acoustic channel states with a vote-queue algorithm. It performs reliable low-rate acoustic communication in a low signal-to-noise ratio (SNR) environment. Results show that the framework has surpassed technologies such as Bluetooth Low Energy (BLE) and Radio-frequency Identification (RFID) in terms of BER, power consumption for device usage, and communication distance: BER is below 0.5% at a 272 bps data reception rate. Also, the framework can be deployed on embedded devices, consumes less power than RFID, and performs similarly to BLE, providing an alternative to cost-effective acoustic communication.

Adversarial examples fool the neural networks by adding slightly-perturbed noise to the original image, which barriers the usability of deep models. Most of the works focused on the adversarial attack on the classification task. We, in this work, attempt to develop an adversarial example generation method for attacking neural-based image quality assessment (IQA). Specifically, instead of employing conventional additive adversarial noise generation methods, we propose an image content deformation approach, avoiding the loss of adversarial noise after compression. The deformation component is designed as neural layers. The given image is firstly deformed and then undergoes compression; an existing IQA evaluates the compressed image. The deformation layers are trained by back-propagating the differences between the targeted IQA score and the originally-evaluated one. Experimental results demonstrate that the proposed method can produce compression-resistant adversarial images for image quality evaluators. The generated adversarial examples could effectively attack neural-based image quality evaluators with less distortion. The data and code of this work are available at https://github.com/luning409/Attack_IQA.

Bandwidth-sensitive applications rely on the accurate estimation of the bottleneck bandwidth. The real-time bandwidth prediction enables the application to cope with bandwidth fluctuation and adjust the transmission strategy to improve the Quality of Experience (QoE) of user. The traditional bandwidth prediction model hardly considers the bandwidth characteristics in various scenarios, making it challenging to achieve high accuracy. In this paper, we propose ALSTM model, which is based on the Long Short Term Memory (LSTM) recurrent neural network and the attention mechanism for multi-scenario bandwidth prediction. Firstly, we conduct the bandwidth trajectories feature analysis, and then we adopt the Support Vector Machine (SVM) to classify scenarios based on the bandwidth characteristics. Secondly, we apply an attention mechanism to assign weights to the input of the bandwidth series, and the attention feature is utilized to effectively select the feature sequences as input to the LSTM model for the prediction. The experimental results show that the ALSTM reduces the Root Mean Square Error (RMSE) by 20%, and the Mean Average Error (MAE) is improved by 26%. For practical applications, we adopt the pre-trained SVM model for real-time scenario detection, dynamic switch the corresponding ALSTM model, and the switching success rate is up to 86%. In addition, by deploying the proposed bandwidth prediction model ALSTM, the DASH's QoE has increased by more than 25%.

The high-throughput, low-latency, and reliable data delivery are fundamental demands of many networked applications, e.g. BitTorrent and Skype. But the inappropriate congestion control of TCPs, caused by the reactive and coarse- grained congestion feedbacks, brings the low link utilization, high queuing delay and frequent packet loss in high bandwidth-delay product network. To mitigate this issue, TCP variants have been developed. Thereinto, the load factor based congestion control (LFCC), e.g. VCP, BMCC, have shown the powerful capabilities to achieve better performances in terms of high link utilization, low persistent queue length, negligible packet loss, and fairness. However, due to the conservative increase and synchronized feedbacks, LFCC faces the slow convergence of the link utilization and inter-flow fairness. This could incur the large flow completion time of new-coming flows indirectly. To solve the issue of existing LFCCs, an asynchronous congestion control based on hybrid feedbacks, called HFCC, is proposed to achieve the faster convergence while keeping the features of LFCCs in this paper. Specifically, HFCC decreases the congestion window when the bottleneck link is in the high-load region and the flow rate exceeds the fair share of the bottleneck bandwidth, or the bottleneck link is in overload region. Otherwise, HFCC increases the congestion window. Note that an overlay coding method is developed in HFCC. To reduce the flow completion time, HFCC adopts an available bandwidth estimation method to speed up the data delivery in low-load region. The simulation results indicate that HFCC has the better performance and faster convergence than VCP, MLCP, and BMCC.