Particle image velocimetry (PIV) plays a significant role in experimental fluid mechanics, which aims to extract the velocity fields from successive particle image pairs. Deep learning (DL) techniques have been proposed to solve such a fluid motion estimation problem. However, the existing DL methods put emphasis on accuracy while ignoring the problem of model redundancy and low computational efficiency. In this article, we propose a novel lightweight convolutional neural network called LightPIVNet for PIV estimation, which is targeted to improve the advanced optical flow model Recurrent All-Pairs Field Transforms (RAFT). Furthermore, considering the real fluid scene, we generated a new dataset PIV-Dataset-II to train our network, which increases the amount and variety of flow fields. Our approach has been verified and analyzed on synthetic and experimental particle images. The experimental results indicate that our method achieves state-of-the-art performance among all the DL methods and is much superior to those classical traditional methods, such as WIDIM and HS optical flow. Meanwhile, our model LightPIVNet reduces parameters by 40.2% and improves inference time by 14.3% compared to the current best DL model PIV-LiteFlowNet-en.

In frequency division duplex model, the massive multiple-input multiple-output (MIMO) systems rely on feedback of channel state information (CSI) to perform precoding operations. This can increase the potential transmission gain of the system. The use of large scale antennas increases the feedback overhead of CSI exponentially. Therefore, CSI must be compressed for feedback. However, convolution based feedback methods lack long-range dependency modeling of the CSI, resulting in limited effects at high compression ratios. In this letter, we propose a novel neural network based on the convolutional transformer architecture to improve the performance of CSI compression and reconstruction at high compression rates. The experimental results show that the average accuracy of ours is improved by 5.39% over the state-of-the-art method at a high compression rate of 1/64. At the same time, the overall performance of the system has been improved.

The research and application areas of transformers have been extensively enlarged due to the success of vision transformers (ViTs). However, due to the lack of local content acquisition capabilities, the pure transformer architectures cannot be trained directly on small datasets. In this work, we first propose a new hybrid model by combining the transformer and convolution neural network (CNN). The proposed model improves the classification ability on small datasets. This is accomplished by introducing more convolution operations in the transformer’s two core sections: 1) Instead of the original multi-head attention mechanism, we design a convolutional parameter sharing multi-head attention (CPSA) block that incorporates the convolutional parameter sharing projection in the attention mechanism; 2) the feed-forward network in each transformer encoder block is replaced with a local feed-forward network (LFFN) block that introduces a sandglass block with more depth-wise convolutions to provide more locality to the transformers. We achieve state-of-the-art results when training from scratch on 4 small datasets as compared with the transformers and CNNs without extensive computing resources and auxiliary training. The proposed strategy opens up new paths for the application of transformers on small datasets.

Computer-aided algorithm plays an important role in disease diagnosis through medical images. As one of the major cancers, lung cancer is commonly detected by computer tomography. To increase the survival rate of lung cancer patients, an early-stage diagnosis is necessary. In this paper, we propose a new structure, multi-level cross residual convolutional neural network (ML-xResNet), to classify the different types of lung nodule malignancies. ML-xResNet is constructed by three-level parallel ResNets with different convolution kernel sizes to extract multi-scale features of the inputs. Moreover, the residuals are connected not only with the current level but also with other levels in a crossover manner. To illustrate the performance of ML-xResNet, we apply the model to process ternary classification (benign, indeterminate, and malignant lung nodules) and binary classification (benign and malignant lung nodules) of lung nodules, respectively. Based on the experiment results, the proposed ML-xResNet achieves the best results of 85.88% accuracy for ternary classification and 92.19% accuracy for binary classification, without any additional handcrafted preprocessing algorithm.