Facial expression recognition plays a key role in promoting the development of comprehensive intelligence and building friendly human-computer interaction. Due to the interference of feature noise in expression data, ...
详细信息
ISBN:
(数字)9798331521950
ISBN:
(纸本)9798331521967
Facial expression recognition plays a key role in promoting the development of comprehensive intelligence and building friendly human-computer interaction. Due to the interference of feature noise in expression data, the lightweight facial expression recognition model with fewer parameters is difficult to learn more expression features through simple training, which limits the improvement of its recognition performance. An efficient facial expression recognition network based on Spot-adaptive Knowledge Distillation is proposed in this paper. Inspired by VoVNetV2, the network designed in this paper is lightweightly improved using Depthwise Separable Convolution and the parameter-free SimAM attention mechanism, reducing the number of parameters to 0.21 M. To further improve the recognition accuracy of the model, Spot-adaptive Knowledge Distillation is employed to improve the characterization ability of the model. The recognition accuracies of the student network designed in this paper on the KDEF and RAF-DB datasets are 93.05% and 81.17% respectively after spot-adaptive distillation.
作者：
Nanxin HuangChi XuSchool of Automation
China University of Geosciences Hubei Key Laboratory of Advanced Control and Intelligent Automation for Complex Systems Engineering Research Center of Intelligent Technology for Geo-Exploration Ministry of Education Wuhan China
Driven by advancements in industrial production and artificial intelligence, the need for pose estimation of new ob-jects in areas like robotic manipulation and virtual reality is increasing. We introduce a zero-shot ...
详细信息
ISBN:
(数字)9798331521950
ISBN:
(纸本)9798331521967
Driven by advancements in industrial production and artificial intelligence, the need for pose estimation of new ob-jects in areas like robotic manipulation and virtual reality is increasing. We introduce a zero-shot object pose estimation approach that identifies the poses of objects excluded from the training dataset, removing the requirement for re-modeling. The method is built around a multi-level features fusion framework de-signed to enhance generalization. First, a trainable feature extraction module filters and selects multi-level features extracted by the backbone network. Unlike traditional convolutional ker-nels, we incorporate a dynamic convolution kernel to enhance the feature extraction capability. Second, in the feature fusion module, we adopt a dynamic weight generation strategy to perform weighted fusion of multi-level features. This method enhances template matching by effectively describing similarities between unseen objects (those absent from the training set) and templates, leveraging robust and adaptive feature representations to narrow the gap with seen objects. Experimental results demonstrate that our approach achieves state-of-the-art performance on two popu-lar benchmark datasets, LineMod and LineMod-Occlusion, proves that our method has better generalization than previous models.
Geological drilling process, owing to complex geological environment and harsh downhole conditions, generates data including characteristics such as pressure, rotational speed, and depth, which are frequently high-dim...
详细信息
The rate of penetration (ROP) is a critical indi-cator for evaluating drilling efficiency. Developing an accurate ROP model is essential for optimizing drilling performance and addressing process control challenges. H...
详细信息
This paper addresses the stability and stabilization issues of Takagi-Sugeno (T-S) fuzzy systems under sampled-data control. In this paper, efforts are dedicated to developing a stability criterion and control strateg...
详细信息
The issue of H∞ state estimation for neural networks with time-varying delays is investigated in this study. Firstly, an augmented Lyapunov-Krasovskii functional (LKF) with two delay-product-type terms is constructed...
详细信息
In contemporary Multi-Agent Reinforcement Learning (MARL), effectively enhancing the expressive capacity of value functions has been a persistent research focus. Many studies have employed value decomposition methods;...
详细信息
The sccheduling for pushing plan during the coking process critically affects the efficiency and stability of production. However, the complexity with mutiple-stage during production makes it difficult to design an ef...
详细信息
ISBN:
(数字)9798331521950
ISBN:
(纸本)9798331521967
The sccheduling for pushing plan during the coking process critically affects the efficiency and stability of production. However, the complexity with mutiple-stage during production makes it difficult to design an efficient coke pushing plan. To address this issue, this paper proposes a scheduling method based on the particle swarm optimization algorithm. Firstly, the fuzzy c-means clustering is utilized to categorize actual operating conditions as either normal or abnormal, thereby facilitating the scheduling of pushing plan under disparate conditions. Subsequently, the scheduling problem for the pushing plan is transformed into a traveling salesman problem, and scheduling models under various conditions are established. Finally, to accelerate the convergence and enhance the algorithm's global search capability, an adaptive inertia adjustment strategy is employed to dynamically regulate the velocity and position of particles. The proposed method has been implemented in the coking process. Through the analysis of application results, the completion coefficient of pushing plan has been increased by 4.25%, demonstrating that the proposed has advantages in scheduling the pushing plan during the actual coking process.
How to realize the high power factor, high efficiency, miniaturization and high power density of AC-DC converter is the key problem of battery charging applications. In this paper, an isolated AC-DC converter is propo...
详细信息
ISBN:
(数字)9798331521950
ISBN:
(纸本)9798331521967
How to realize the high power factor, high efficiency, miniaturization and high power density of AC-DC converter is the key problem of battery charging applications. In this paper, an isolated AC-DC converter is proposed. The input voltage is rectified by the first stage AC-DC conversion, and the output is stabilized by the series resonant converter at the secondary stage. A sinusoidal modulation method is adopted to realize the constant current output characteristic independent of load. Then through the output current control loop and the PFC control loop, the constant current output and input high power factor are realized. The design of high switching frequency and resonant frequency provides conditions for high power density and miniaturization of converter design. Finally, the simulation of rated power of 960W was built by Matlab-simulink, which verified the rationality of the converter design.
Accurately and promptly detecting the pipeline anomaly is crucial to the safe operation of pipeline systems, while a difficulty lies in that many existing methods require massive data for training models. However, pip...
详细信息
ISBN:
(数字)9798331521950
ISBN:
(纸本)9798331521967
Accurately and promptly detecting the pipeline anomaly is crucial to the safe operation of pipeline systems, while a difficulty lies in that many existing methods require massive data for training models. However, pipelines are running under normal state for the most of the time, and labeled pipeline anomaly data is usually scarce. Among the commonly used sensors, vibration sensors are widely utilized in pipeline detection because of their advantages such as easy installation and high sensitivity. However, the vibration signal shows non-stationary characteristics when anomalies occur, and are contaminated by noises, making it difficult to represent the actual state with features extracted from either the time or frequency domain. Accordingly, this paper proposes a pipeline anomaly detection method based on the KPCA (kernel principal component analysis) and cosine distance prototypical network. First, features are extracted from original signals; then, the feature dimension is reduced by KPCA; last, the cosine distance is introduced to the prototypical network for anomaly detection. The effectiveness of the proposed method is demonstrated by case studies involving experimental data.
暂无评论