Training generative adversarial networks (GANs) for noise-to-image synthesis is a challenge task, primarily due to the instability of GANs’ training process. One of the key issues is the generator’s sensitivity to input data, which can cause sudden fluctuations in the generator’s loss value with certain inputs. This sensitivity suggests an inadequate ability to resist disturbances in the generator, causing the discriminator’s loss value to oscillate and negatively impacting the discriminator. Then, the negative feedback of discriminator is also not conducive to updating generator’s parameters, leading to suboptimal image generation quality. In response to this challenge, we present an innovative GANs model equipped with a learnable auxiliary module that processes auxiliary noise. The core objective of this module is to enhance the stability of both the generator and discriminator throughout the training process. To achieve this target, we incorporate a learnable auxiliary penalty and an augmented discriminator, designed to control the generator and reinforce the discriminator’s stability, respectively. We further apply our method to the Hinge and LSGANs loss functions, illustrating its efficacy in reducing the instability of both the generator and the discriminator. The tests we conducted on LSUN, CelebA, Market-1501, and Creative Senz3D datasets serve as proof of our method’s ability to improve the training stability and overall performance of the baseline methods.

Pre-harvest estimation of fruit production is crucial for fruit storage and price analysis in the planting of fruit trees. However, prior research consistently displays low accuracy because of problems with small objects, leaf occlusion, and fruit overlap, and they emphasize large networks that are unrealistic in the real world. In this study, we emphasize the use of smartphones to evaluate citrus fruit production. We suggest a simple method for detecting objections based on the YOLOv5 algorithm compressed by knowledge distillation. To extract the visual features, we first use mobilenetV2 as the foundation of YOLOv5. To learn the reliable detection features, we also incorporate an attention mechanism into YOLOv5. As such, we can obtain embedded Yolo served as a student model, which is learned via knowledge distillation. As such we can take the lightweight student model as the final detection model. Finally, we take the embedded Yolo to detect the citrus fruits and take a linear regression model to predict the number of counted fruits and the production is estimated. Experiments show that the proposed method can accurately count fruits and approximate the production.

The few-shot image classification task aims to train a model to correctly classify unlabeled samples when only a few example images are available. Most current metric-based approaches consider only a single task and ignore the variability between tasks. In this paper, we extract potential information of images from both local and global levels by training feature extractors sensitive to global and local features. Then, the multilevel features are optimized further by employing a self-attention mechanism to assign suitable weights based on the target tasks’ characteristics. Extensive experiments on several benchmark datasets show that the proposed model has better generalization ability for different tasks, and our strategy improves by 4%~6% over the baseline, showing satisfactory results. Our code is available at: https://github.com/XiangLi0503/MTnet.

Non-Orthogonal Multiple Access (NOMA) allows for the sharing of communication link resources among multiple users, which increases spectrum efficiency. In this paper, we consider the NOMA-based mobile edge computing (MEC) networks with unmanned aerial vehicle (UAV), where convenient computation offloading services for ground devices (GDs) with NOMA is provided. The main focus is to investigate the computation capacity in terms of computation throughput, which is characterized by the total size of completed tasks achieved through air-ground collaboration. To guarantee the fairness of GDs, the minimum computation throughput for all GDs is maximized by jointly designing the UAV trajectory, channel relationship coefficient, and computation resource allocation, where the formulated problem is non-convex with closely coupled mixed-integer design variables. A novel penalty based iterative algorithm is proposed, where penalty term is employed to penalize non-integer solution and an inexact block coordinate decent method is adopted to avoid strong locality of the optimized solution, where the convergence is also proved. We conduct extensive simulations and show that our joint design algorithm outperforms other benchmark schemes.

In tiny object detection, the main challenges are tiny objects’ weak feature responses and possible semantic disappearance in deep networks. To address the problems, we proposed an Instance-level, Scale-adaptive, Shape-preserved, and Semantic-consistent Supervision (I4S) module for better locating tiny objects. It models across-scale feature responses of an instance as an elliptic cone, whose axis indicates the instance’s semantic center in different scales. By the cone supervision on across-scale feature maps, it not only exploits the classification semantic from multi-scale feature maps, but also preserves and explores the information of instance shape in consecutive scales. Experiment results on public datasets proved that our method can effectively improve the location accuracy and significantly reduce the missed detection rate comparing with the method of directly fusing multi-scale features.

Detecting objects in aerial images is an active yet challenging task due to the arbitrary orientation of aerial targets and the lack of details for small objects. Observing that geometric priors, e.g., some kinds of points and lines with specific geometrical properties, are helpful to determine the oriented bounding box, we proposed a Point-Line-Region (PLR) supervision module embedded in the Feature Pyramid Networks (FPN) to learn the robust geometrical features, which are conducive to evaluate and locate the orientation, vertexes, and boundary of the bounding box. Our method improves feature representations of aerial targets by utilizing the supervision of across-category geometrical semantics rather than their category semantics in the previous works. Extensive comparison experiments on the DOTA dataset prove that remarkable accuracy gains of mAP, about 2.1%, are achieved by integrating the PLR module in different architectures of networks.

This work focuses on a novel agricultural application of monitoring and managing summer shoots on citrus fruit trees. To enhance the management efficiency and reliability, a novel scheme based on aerial image analysis was presented to implement intelligent supervision. To this end, we proposed an end-to-end network containing three main modules, i.e. semantic circle detection, circle-inner segmentation, and decision classification. They are respectively in charge of detecting multiple trees, evaluating each tree’s shoot quantity, and predicting management decision for each tree in aerial images. Detailed experiment analysis validates the proposed scheme and network on the aerial images collected from practical citrus orchards.

Abstract Worst-case resource usage provides a useful guidance in the design, configuration and deployment of software, especially when it runs under a context with limited amount of resources. Static resource-bound analysis can provide sound upper bounds of worst-case resource usage but may provide too conservative, even unbounded, results. In this paper, we present a resource-usage-aware fuzzing approach to estimate worst-case resource usage. The key idea is to guide the fuzzing process using resource-usage amount together with resource-usage relevant coverage. Moreover, we leverage semantic patch to make use of static analysis information (including control-flow, function-call, etc.) to instrument the original program, for the sake of aiding the subsequent fuzzing. We have conducted experiments to estimate worst-case resource usage of various resources in real-world programs, including heap memory, stack depths, sockets, user-defined resources, etc. The preliminary experimental results show the promising ability of our approach in estimating worst-case resource usage in real-world programs, compared with two state-of-the-art fuzzing tools (AFL and MemLock).

Due to its high mobility, flexible deployment, and low cost, unmanned aerial vehicles (UAVs) have attracted wide attention in wireless communication in recent years. However, the delay requirements (e.g., video streaming, online game, etc.) may limit the UAV's mobility. In this paper, we consider a three-dimensional (3D) UAV communication network, where a UAV is employed to fly flexibly in 3D space to serve ground users with delay requirements. To characterize the fundamental tradeoff between throughput and delay, we introduce the minimum required rate for users and aim to maximize the minimum weighted sum of throughput and required rate for each user, via joint optimization of the 3D UAV trajectory as well as communication time and rate allocation. The formulated problem is a non-convex optimization problem, which is generally intractable. By decomposing the formulated problem into two subproblems, we propose an iterative algorithm by block coordinate descent and difference of two convex (D.C.) optimization as well as successive convex approximation (SCA) techniques. Finally, extensive simulation results show that our proposed solution outperforms baseline schemes and unveils the interesting insights and tradeoff between throughput and delay over 3D UAV communication networks.

In agricultural pest management based on computer vision, numerous species of tiny pests need to be detected in images. However, such tiny detection objects are usually missed when adopting deep detection networks. To improve the detection of tiny pests, this paper presented an adaptive tiny object detection network based on the CenterNet framework. Firstly, a branch with a learnable gating function is integrated into the backbone, and supervised learning is performed on it so that tiny pests’ high-resolution feature maps with category and location semantics are exploited, and the learned gating function adaptively controls the combination of such feature maps and the backbone. Moreover, we proposed a size-adaptive weighting method to improve the CenterNet’s detection loss function. In training, a higher weight will be assigned to an instance if its size is smaller or its prediction center is farther from the ground truth. Extensive experiments on multiple datasets verify that our two contributions, i.e. the adaptive-gating branch, and the size-adaptive weighting method, are both help to enhance tiny pests’ weak feature responses and their discriminations, and further improve the IoU accuracies in detection.

Recent years have witnessed an exponential growth of model scale in deep learning-based recommender systems---from Google's 2016 model with 1 billion parameters to the latest Facebook's model with 12 trillion parameters. Significant quality boost has come with each jump of the model capacity, which makes us believe the era of 100 trillion parameters is around the corner. However, the training of such models is challenging even within industrial scale data centers. We resolve this challenge by careful co-design of both optimization algorithm and distributed system architecture. Specifically, to ensure both the training efficiency and the training accuracy, we design a novel hybrid training algorithm, where the embedding layer and the dense neural network are handled by different synchronization mechanisms; then we build a system called Persia (short for parallel recommendation training system with hybrid acceleration) to support this hybrid training algorithm. Both theoretical demonstrations and empirical studies with up to 100 trillion parameters have been conducted to justify the system design and implementation of Persia. We make Persia publicly available (at github.com/PersiaML/Persia) so that anyone can easily train a recommender model at the scale of 100 trillion parameters.

Unmanned aerial vehicle (UAV) enabled communication has emerged as an appealing technology in wireless sensor networks (WSNs) for efficient data collection. This paper studies the energy issues for data collection in UAV enabled WSNs. It is revealed that a fundamental tradeoff exists between the energy consumption of UAV and that of all sensor nodes (SNs). To characterize such a tradeoff, an optimization problem is formulated to minimize the weighted sum of the energy consumption of UAV and SNs, via jointly optimizing the UAV trajectory, mission completion time, as well as the wake-up scheduling for all SNs. As the formulated problem is a non-convex problem with infinite variables over time, it is difficult to be optimally solved. To tackle this issue, the original problem is transformed into a discretized equivalent with path discretization method, and then a locally optimal solution is obtained by applying the successive convex approximation and block coordinate descent techniques. Simulations are conducted to corroborate our study and show the flexible tradeoff achieved by the proposed design for energy balance between UAV and SNs.

Distance metric learning suppresses the intraclass variation while preserving the inter-class variation between two feature vectors. However, these two types of information are mixed in the feature vectors that need to be separated based on learning from the training data. The limited training data may not be able to well separate these two types of information and hence limits the effectiveness of metric learning. This letter proposes to exploit off-feature information to help suppress the intraclass variation of the feature vectors. For face recognition, some identity-independent information such as pose, expression, and occlusion is extracted from source images and utilized as the off-feature information to enhance the performance of distance metric learning. In training, the algorithm learns how to incorporate the off-feature information to suppress the intraclass variation of features. In recognition, the similarity score of a face image pair is determined by its distance in feature space and that of off-feature space. Extensive experiments demonstrate that the proposed off-feature information incorporated metric learning is helpful to suppress the intraclass variation of feature vectors, which visibly enhances the existing metric learning algorithms.

Volcano monitoring is of great interest to public safety and scientific explorations. However, traditional volcanic instrumentation such as broadband seismometers are expensive, power hungry, bulky, and difficult to install. Wireless sensor networks (WSNs) offer the potential to monitor volcanoes on unprecedented spatial and temporal scales. However, current volcanic WSN systems often yield poor monitoring quality due to the limited sensing capability of low-cost sensors and unpredictable dynamics of volcanic activities. In this article, we propose a novel quality-driven approach to achieving real-time, distributed, and long-lived volcanic earthquake detection and timing. By employing novel in-network collaborative signal processing algorithms, our approach can meet stringent requirements on sensing quality (i.e., low false alarm/missing rate, short detection delay, and precise earthquake onset time) at low power consumption. We have implemented our algorithms in TinyOS and conducted extensive evaluation on a testbed of 24 TelosB motes as well as simulations based on real data traces collected during 5.5 months on an active volcano. We show that our approach yields near-zero false alarm/missing rate, less than one second of detection delay, and millisecond precision earthquake onset time while achieving up to six-fold energy reduction over the current data collection approach.

This paper presents the design, deployment, and evaluation of a real-world sensor network system in an active volcano - Mount St. Helens. In volcano monitoring, the maintenance is extremely hard and system robustness is one of the biggest concerns. However, most system research to date has focused more on performance improvement and less on system robustness. In our system design, to address this challenge, automatic fault detection and recovery mechanisms were designed to autonomously roll the system back to the initial state if exceptions occur. To enable remote management, we designed a configurable sensing and flexible remote command and control mechanism with the support of a reliable dissemination protocol. To maximize data quality, we designed event detection algorithms to identify volcanic events and prioritize the data, and then deliver higher priority data with higher delivery ratio with an adaptive data transmission protocol. Also, a light-weight adaptive linear predictive compression algorithm and localized TDMA MAC protocol were designed to improve network throughput. With these techniques and other improvements on intelligence and robustness based on a previous trial deployment, we air-dropped 13 stations into the crater and around the flanks of Mount St. Helens in July 2009. During the deployment, the nodes autonomously discovered each other even in-the-sky and formed a smart mesh network for data delivery immediately. We conducted rigorous system evaluations and discovered many interesting findings on data quality, radio connectivity, network performance, as well as the influence of environmental factors.

Volcano monitoring is of great interest to public safety and scientific explorations. However, traditional volcanic instrumentation such as broadband seismometers are expensive, power-hungry, bulky, and difficult to install. Wireless sensor networks (WSNs) offer the potential to monitor volcanoes at unprecedented spatial and temporal scales. However, current volcanic WSN systems often yield poor monitoring quality due to the limited sensing capability of low-cost sensors and unpredictable dynamics of volcanic activities. Moreover, they are designed only for short-term monitoring due to the high energy consumption of centralized data collection. In this paper, we propose a novel quality-driven approach to achieving real-time, in-situ, and long-lived volcanic earthquake detection. By employing novel in-network collaborative signal processing algorithms, our approach can meet stringent requirements on sensing quality (low false alarm/missing rate and precise earthquake onset time) at low power consumption. We have implemented our algorithms in TinyOS and conducted extensive evaluation on a testbed of 24 TelosB motes as well as simulations based on real data traces collected during 5.5 months on an active volcano. We show that our approach yields near-zero false alarm/missing rate and less than one second of detection delay while achieving up to 6-fold energy reduction over the current data collection approach.

We present a lightweight lossless compression algorithm for realtime sensor networks. Our proposed adaptive linear filtering compression (ALFC) algorithm performs predictive compression using adaptive linear filtering to predict sample values followed by entropy coding of prediction residuals, encoding a variable number of samples into fixed-length packets. Adaptive prediction eliminates the need to determine prediction coefficients a priori and, more importantly, allows compression to dynamically adjust to a changing source. The algorithm requires only integer arithmetic operations and thus is compatible with sensor platforms that do not support floating-point operations. Significant robustness to packets losses is provided by including small but sufficient overhead data to allow each packet to be independently decoded. Real-world evaluations on seismic data from a wireless sensor network testbed show that ALFC provides more effective compression and uses less resources than an alternative recent work of lossless compression, S-LZW. Experiments in a multi-hop sensor network also show that ALFC can significantly improve raw data throughput and energy efficiency. We also implement the algorithm in our real sensor network, and show that our linear prediction based compression algorithm significantly improves data reliability and network efficiency.

This paper presents the design and deployment experience of an air-dropped wireless sensor network for volcano hazard monitoring. The deployment of five self-contained stations into the rugged crater of Mount St. Helens only took one hour with a helicopter. The stations communicate with each other through an amplified 802.15.4 radio and establish a self-forming and self-healing multihop wireless network. The transmit distance between stations was up to 8 km with favorable topography. Each sensor station collects and delivers real-time continuous seismic, infrasonic, lightning, GPS raw data to a gateway. The main contribution of this paper is the design of a robust sensor network optimized for rapid deployment during periods of volcanic unrest and provide real-time long-term volcano monitoring. The system supports UTC-time-synchronized data acquisition with 1 ms accuracy, and is remotely configurable. It has been tested in the lab environment, the outdoor campus, and the volcano crater. Despite the heavy rain, snow, and ice as well as gusts exceeding 160 km per hour, the sensor network has achieved a remarkable packet delivery ratio above 99 percent with an overall system uptime of about 93.8 percent over the 1.5 months evaluation period after deployment. Our initial deployment experiences with the system demonstrated to discipline scientists that a low-cost sensor network system can support real-time monitoring in extremely harsh environments.

This paper presents the design and deployment experience of an air-dropped wireless sensor network for volcano hazard monitoring. The deployment of five stations into the rugged crater of Mount St. Helens only took one hour with a helicopter. The stations communicate with each other through an amplified 802.15.4 radio and establish a self-forming and self-healing multi-hop wireless network. The distance between stations is up to 2 km. Each sensor station collects and delivers real-time continuous seismic, infrasonic, lightning, GPS raw data to a gateway. The main contribution of this paper is the design and evaluation of a robust sensor network to replace data loggers and provide real-time long-term volcano monitoring. The system supports UTCtime synchronized data acquisition with 1ms accuracy, and is online configurable. It has been tested in the lab environment, the outdoor campus and the volcano crater. Despite the heavy rain, snow, and ice as well as gusts exceeding 120 miles per hour, the sensor network has achieved a remarkable packet delivery ratio above 99 % with an overall system uptime of about 93.8 % over the 1.5 months evaluation period after deployment. Our initial deployment experiences with the system have alleviated the doubts of domain scientists and prove to them that a low-cost sensor network system can support real-time monitoring in extremely harsh environments.

We present a lightweight lossless compression algorithm for realtime sensor networks. Our proposed adaptive linear filtering compression (ALFC) algorithm performs predictive compression, using adaptive linear filtering to predict sample values followed by entropy coding of prediction residuals, encoding a variable number of samples into fixed-length packets. Adaptive prediction eliminates the need to determine prediction coefficients a priori and, more importantly, allows compression to dynamically adjust to a changing source. The algorithm requires only integer arithmetic operations and thus is compatible with sensor platforms that do not support floating-point operations. Significant robustness to packets losses is provided by including small but sufficient overhead data to allow samples in each packet to be independently decoded. Real-world evaluations on seismic data from a wireless sensor network testbed show that ALFC provides more effective compression and uses less resources than some other lossless compression approaches such as S-LZW. Experiments in a multi-hop sensor network also show that ALFC can significantly improve raw data throughput and energy efficiency.

Earlier sensor network MAC protocols focus on energy conservation in low-duty cycle applications, while some recent applications involve real-time high-data-rate signals. This motivates us to design an innovative localized TDMA MAC protocol to achieve high throughput and low congestion in data collection sensor networks, besides energy conservation. TreeMAC divides a time cycle into frames and frame into slots. Parent determines children's frame assignment based on their relative bandwidth demand, and each node calculates its own slot assignment based on its hop-count to the sink. This innovative 2-dimensional frame-slot assignment algorithm has the following nice theory properties. Firstly, given any node, at any time slot, there is at most one active sender in its neighborhood (including itself). Secondly, the packet scheduling with TreeMAC is bufferless, which therefore minimizes the probability of network congestion. Thirdly, the data throughput to gateway is at least 1/3 of the optimum assuming reliable links. Our experiments on a 24 node test bed demonstrate that TreeMAC protocol signi ficantly improves network throughput and energy efficiency, by comparing to the TinyOS's default CSMA MAC protocol and a recent TDMA MAC protocol Funneling-MAC.