Facial expression recognition using convolutional neural networks (CNNs) is a prevalent research area, and the network’s complexity poses obstacles for deployment on devices with limited computational resources, such as mobile devices. To address these challenges, researchers have developed lightweight networks with the aim of reducing model size and minimizing parameters without compromising accuracy. The LiteFer method introduced in this study incorporates depth-separable convolution and a lightweight attention mechanism, effectively reducing network parameters. Moreover, through comprehensive comparative experiments on the RAFDB and FERPlus datasets, its superior performance over various state-of-the-art lightweight expression-recognition methods is evident.
A novel compact broadband negative group delay (NGD) circuit with flatness and bandwidth enhancement is presented. The presented negative group delay circuit (NGDC) consists of a high-impedance transmission line connected by two resistors, which are linked together with two coupled lines and a low-impedance transmission line. The flatness of the NGD is enhanced by tuning the characteristic impedance of the transmission lines. In order to verify the method, a compact broadband NGDC with a size of 29.4 mm × 58.1 mm (0.14 λg × 0.29 λg) is designed, fabricated, and measured at the center frequency of 1.0 GHz. The measured results show that an NGD time of −0.49 ns at the center frequency is obtained with return loss and insertion loss of 35.0 dB and 18 dB, respectively. And, the flat-NGD bandwidth reaches 509 MHz (50.9%) over 0.766 to 1.275 GHz with 19% group-delay fluctuation.
A bandstop filter with flat negative group delay (NGD) is proposed. The flatter NGD and reduced insertion loss have been achieved in the passband of the bandstop filter. This paper provides the group delay and insertion loss equations for each component of bandstop filter, then analyzes the effects of different unit values on bandwidth, flatness, and minimum NGD. The experimental results show that the flatness is 5.06% and the NGD is -0.498ns. The insertion loss is 6.72dB, and the relative NGD bandwidth is 85.2%. The simulation and measurement results are basically consistent with the theoretical results. The feasibility of the design is proved.
Deep learning models are difficult to train in image classification due to the sparsity of samples, so few-shot learning that can train a good model with a small number of samples came into being, and the few-shot learning that applies it to the classification task is called few-shot classification. Prototype networks are an effective method for few-shot classification, but the prototype representation obtained by averaging only sparse samples is weak and the final classification results are poor. To address these challenges, a prototype fusion completion network(ProfcNet) is proposed, which enriches the representation of the prototype by using only the data from the prototype itself, uses the fusion module to obtain a more representative sample fusion feature mapping, complements the prototype by using the fusion feature mapping to obtain the final prototype, and gives an inference test that applies to the prototype fusion completion network. Extensive experiments were conducted on four few-shot classification public datasets, and the experimental results show that the method can improve the accuracy up to almost 30% and 20% for single-shot and five-shot, respectively, which is significant compared with the prototype network. In addition to this, only then comparison tests and visualisations were used to validate the accuracy of the different modules.