Boundary equilibrium generative adversarial networks (BEGANs) can generate impressively realistic face images, but there is a trade-off between the quality and the diversity of generated images. Based on BEGANs, we propose an effective approach to generate images with higher quality and better diversity. By adding a second loss function (a denoising loss) to the discriminator, the discriminator can learn more useful information about the distribution of real images. Naturally, the ability of discriminator in distinguishing between real and generated images is improved, which further guides the generator to produce more realistic images to confuse the discriminator. We also find that using technique of batch normalization in BEGANs architecture can improve the diversity of generated images. By using batch normalization and adding a denoising loss to the objective of discriminator, we achieve comparative generations on CIFAR-10 and CelebA data sets. In addition, we evaluate the effect of several techniques on BEGANs framework through "Inception-Score", a measure which has been found to correlate well with human assessment of generated samples.

Human Activity Recognition (HAR) has attracted much attention due to the application in consumer electronics. To address the shortcomings of existing HARs, such as no identification, privacy leakage, battery maintenance, low accuracy, and few recognized activities, we have proposed new bound-RFID HARs. Firstly, we have devised body RFID skeleton to sense human activity by collecting tag response records of the skeleton node, which form the human activity dataset suitable for a benchmark. Secondly, we propose data standardization to solve the problem caused by the differences in the size and the amount of the tag response signal feature data. Thirdly, some body RFID skeleton-based HARs are proposed by utilizing machine learning, recurrent neural networks, and graph convolutional networks. In these proposed HARs, the best HAR is an improvement on the state-of-the-art HAR in the computer vision. The experimental results show that the proposed best HAR improves 0.27% than the state-of-the-art HAR by utilizing RSSI, Phase, and Doppler Frequency with only 4.5MB parameters, and achieves a recognition accuracy of 98.52%. Therefore, the devised body RFID skeleton effectively senses human activities, and the proposed HAR recognizes human activities with high accuracy and fine granularity. The dataset and source code are available at http://www.cwnuiot.net/BRS/.

With the increasing number of users and application in the era of the Industrial Internet of Things (I-IoT), computing efficiency and energy consumption become two vital problems. Multi-processor heterogeneous system has been widely used for handling parallel workflow applications to advance the performance of system, in which efficient workflow scheduling and the choice of frequency for each task play key roles. However, most scheduling algorithms just focus on an optimal solution for every single task and may exclude optimal task priority sequence at the task sorting phase. Therefore, this paper proposes a Predictive Energy Consumption Scheduling (PECS) algorithm to match the frequency for each task. Firstly, the alterable time space is introduced to decompose the deadline of the whole workflow application into each task. Secondly, as the basis of allocation, Predictive Energy Consumption Matrix (ECM) is defined to predict the impact of current task allocation on the energy consumption of subsequent tasks. Thirdly, the processor with the lowest predicted energy consumption is matched for each task to form a preliminary scheduling scheme with the constraint of time. Finally, all tasks in the same priority level of a processor are swapped in order, then a scheduling strategy is formed to minimize energy consumption. Experimental results reveal that our proposed algorithm can reduce energy 48.28% and 13.33% averagely compared with other two state of the art algorithms.

As an indispensable material, aluminum requires real-time surface defect detection to effectively control quality. However, defects on the surface of aluminum have the inherent properties of tiny size, diverse morphology, and multiscale variations, which present a significant challenge for defect detection. This article aims to improve the performance and efficiency of a defect detection model and proposes an efficient multiscale perceptual enhancement network for aluminum defect detection (ADDet). First, to compensate for the loss of deep features and mitigate information interference from the background, a multilayer semantic-aware enhancement module (MSEM) is proposed for the feature extraction network. Second, a proximity feature pyramid (PFP) structure in the neck is designed to effectively merge multiscale features with low computational cost and enhance feature representation across spatial coordinates, channels, and scales. Finally, a large kernel-size decoupled head (LKSDH) with a large receptive field and multiple detection branches is constructed in the detection head to improve detection accuracy. Experimental results show that the mean average precision (mAP) of the proposed ADDet is 80.85% with a real-time detection speed of 63.41 frames per second (FPS). Compared to YOLOv7, ADDet improves the mAP by 3.02%, and the number of parameters and FLOPs are reduced by 23.8% and 13%, respectively. Moreover, ADDet is shown to be superior to state-of-the-art detection methods in terms of detection accuracy.

Deep neural network (DNN) models have been widely used in many tasks due to their superior performance. However, DNN models are usually vulnerable to adversarial example attacks, which limits their applications in many safety-critic scenarios. How to effectively detect adversarial examples to enhance the robustness of DNN models has attracted much attention in recent years. Most adversarial example detection methods require modifying or retraining the model, which is impractical and reduces the classification accuracy of normal examples. In this paper, we propose an adversarial example detection approach that does not require modification of the DNN models and meanwhile retains the classification accuracy of normal examples. The key observation is that when we transform the input example with some operations (e.g., masking a pixel with a reference value), feed the transformed example to the target model, and use the output of the intermediate layers to predict the label of the example, the generated label sequences of adversarial examples will be extremely discrepant but the label sequences of normal examples keep nearly unchanged. Motivated by this observation, we design an approach to detect adversarial examples based on the label sequence discrepancy (LSD) of the given examples. The experimental results against five mainstream adversarial attacks on three benchmark datasets demonstrate that LSD outperforms the state-of-the-art solutions in the detection rate of adversarial examples. Moreover, LSD performs well at various confidence levels and exhibits good generalizability between different attacks.