Automatic speech recognition (ASR) is a fundamental technology in the field of artificial intelligence. End-to-end (E2E) ASR is favored for its state-of-the-art performance. However, E2E speech recognition still faces speech spatial information loss and text local information loss, which results in the increase of deletion and substitution errors during inference. To overcome this challenge, we propose a novel Enhancing Hybrid Architecture with Fast Attention and Capsule Network (termed En-HACN), which can model the position relationships between different acoustic unit features to improve the discriminability of speech features while providing the text local information during inference. Firstly, a new CNN-Capsule Network (CNN-Caps) module is proposed to capture the spatial information in the spectrogram through the capsule output and dynamic routing mechanism. Then, we design a novel hybrid structure of LocalGRU Augmented Decoder (LA-decoder) that generates text hidden representations to obtain text local information of the target sequences. Finally, we introduce fast attention instead of self-attention in En-HACN, which improves the generalization ability and efficiency of the model in long utterances. Experiments on corpora Aishell-1 and Librispeech demonstrate that our En-HACN has achieved the state-of-the-art compared with existing works. Besides, experiments on the long utterances dataset based on Aishell-1-long show that our model has a high generalization ability and efficiency.

3D face recognition (FR) has become increasingly widespread due to the illumination invariance and pose robustness of 3D face data. Most existing 3D FR methods can only achieve excellent performance on complete and smooth faces. However, low-quality 3D FR with missing facial features still suffers from insufficient discriminative feature extraction for visible face regions. In this paper, we propose a dual-stream multi-scale fusion network (DSNet) for low-quality 3D FR. First, in the first stream, we design a new multi-scale local and global feature fusion network, which consists of an enhanced shallow feature extraction module, an enhanced deep feature extraction module, and a layered multi-scale feature correlation fusion module, aiming to obtain more discriminative details and category information of the facial visible region, reducing the interference of similar features and the redundancy of the same features. Second, we also introduced a capsule network as the second stream to enhance the expression of 3D facial spatial position information, thereby further improving the performance of low-quality 3D FR with missing facial features. We conduct extensive experiments on low-quality datasets (Lock3DFace, KinectFaceDB, and IIIT-D) and cross-quality datasets synthesized by Bosphorus. These results show that our proposed DSNet can achieve state-of-the-art recognition performance and exhibit excellent performance on low-quality 3D faces with missing facial features.

Few-shot learning (FSL) uses prior knowledge and supervised experience to effectively classify hyperspectral images (HSIs), thereby reducing the cost of large numbers of labeled samples. However, existing few-shot methods ignore the correlation between cross-domain feature channels, and the feature representation ability is insufficient. To address above issue, this paper proposes a novel Residual Channel Attention Based Sample Adaptation Few-Shot Learning for Hyperspectral Image Classification(RCASA-FSL) for hyperspectral image classification (HSIC), which can capture and enhance cross-domain dependencies through multi-layer residual connection and random-based feature recalibration. Specifically, a Deep Residual Feature Channel Attention Mechanism (DRFCAM) is designed to obtain cross-domain dependencies by residual concatenation, and further the residual structure is stacked for mining depth discrimination information. Furthermore, a new Random-based Feature Recalibration Module (RFRM) is proposed to reassign the feature weights via random matrix, which fully explore feature weight relationships to guide the sample adaptation process. Besides, we design a joint loss function with combining the FSL loss and domain adaptive loss for further optimization model. Experiments conducted on several standard hyperspectral datasets demonstrate that the proposed RCASA-FSL is superior to other FSL techniques in both quantitative and qualitative aspects.

Patch features obtained by fixed convolution kernel have become the main form in hyperspectral image (HSI) classification processing. However, the fixed convolution kernel limits the weight learning of channels, which results in the potential connections between pixels not being captured in patches, and seriously affects the classification performance. To tackle the above issues, we propose a novel Adaptive Pixel Attention Network, which can improve HSI classification by further mining the connections between pixels in patch features. Specifically, a Spectral–Spatial Superposition Enhancement module is first proposed for enhancing the spectral–spatial information of 3D input cubes via complementing the 1D spectral vectors by zero and reflection filling operations. More importantly, we also propose a new Adaptive Pixel Attention mechanism, which explores Cosine and Euclidean similarity to adaptively explore the distance and angle relationship between pixels of different scale convolution patch features. Moreover, the Cross-Layer Information Complement module is designed to form a contextual interaction by integrating the output features of different convolution layers, which can prevent the omission of discriminative information and further improve the network performance. Experimental results on four widely used HSI datasets IP, UP, HU, and KSC show that the proposed network is superior to other state-of-the-art classification models in accuracy, and it also has a better efficiency than other 3D works.

Coherent Doppler Wind Lidar (CDWL) is an effective tool for measuring the atmospheric wind field. However, CDWL is affected by various noises, which can reduce the usable value of the received echo signal. This paper proposes a novel joint denoising algorithm based on SVD-ICEEMDAN-SCC-MF to remove noises in CDWL detection. The SVD-ICEEMDAN-SCC-MF consists of singular value decomposition (SVD), improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN), Spearman correlation coefficient (SCC), and median filtering (MF). Specifically, the SVD first separates the signal from the noise by retaining the main feature (large singular value) and removing the remained components (small singular value) to achieve the initial signal reconstruction. Then, ICEEMDAN is used for decomposition to distinguish the intrinsic mode function (IMF) of the signal and the noise. The SCC of the retained components is calculated to determine the correlation of the reconstructed signal. Furthermore, low correlation components of the reconstructed signal are denoised again by median filtering (MF). Finally, the complete denoised signal is obtained by combining the components after MF and the high correlation components in the previous stage. The validity of the SVD-ICEEMDAN-SCC-MF is verified in simulated and real data, and the denoising effect is significantly better than other algorithms. In simulation cases, the $SN{R}_{out}$ of the proposed method is improved by 20.5117 dB at most, from −5 dB to 15.5117 dB, and the $RMSE$ is only 0.5174. After denoising the power spectrum of the real CDWL signal, the detection range is extended from 3 km to more than 3.6 km.