The last decade has seen increased interest in environmental sound classification (ESC) due to the increased complexity and rich information of ambient sounds. The state-of-the-art methods for ESC are based on transfer learning paradigms that often utilize learned representations from common image-classification problems. This paper aims to determine the effectiveness of employing pre-trained convolutional neural networks (CNNs) for audio categorization and the feasibility of retraining. This study investigated various hyper-parameters and optimizers, such as optimal learning rate, epochs, and Adam, Adamax, and RMSprop optimizers for several pre-trained models, such as Inception, and VGG, ResNet, etc. Firstly, the raw sound signals were transferred into an image format (log-Mel spectrogram). Then, the selected pre-trained models were applied to the obtained spectrogram data. In addition, the effect of essential retraining factors on classification accuracy and processing time was investigated during CNN training. Various optimizers (such as Adam, Adamax, and RMSprop) and hyperparameters were utilized for evaluating the proposed method on the publicly accessible sound dataset UrbanSound8K. The proposed method achieves 97.25% and 95.5% accuracy on the provided dataset using the pre-trained DenseNet201 and the ResNet50V2 CNN models, respectively.

An approach for null broadening beamforming is proposed based on adaptive variable diagonal loading (VDL) and combined with the covariance matrix taper (CMT) approach, aiming at improving the robustness of adaptive antenna null broadening beamforming when array calibration error exists. Hence, it is named VDL-CMT. In this novel approach, the signal-to-noise ratio in the tapered sample covariance matrix is estimated and the VDL factor can be obtained adaptively. Then, the covariance matrix of the CMT approach is loaded with the obtained VDL factor. According to simulation results, in the case of array calibration error, robustness of the VDL-CMT is significantly improved and its performance is better than that of the existing adaptive antenna null broadening beamforming approaches.

Because of the challenges associated with the difficulty of accurately labeling the remote sensing (RS) scene images and the need to identify new scene classes, few-shot learning (FSL) has shown significant advantages in addressing the remote sensing scene classification (RSSC) tasks, leading to a growing interest. However, due to the scale variations of targets and irrelevant complex background in scene images, the current few-shot methods exist the following problems: the problem of the extraction capability of feature extractor in the few-shot mechanism; the problem of the separability of few-shot RS scene images classifier. To solve the above problems, an approach, called few-shot remote sensing scene classification via subspace based on multiscale feature learning is introduced in this work. We firstly design a multiscale feature learning technique to address scale variations of the targets in the scene images. Concretely, different branches are utilized to learn scene features at various scales. And the self-attention mechanism is embedded in each branch to incorporate the understanding of the global information in the different scale features. After that, a multiscale feature fusion operation, incorporating channel attention, will be devised to effectively merge the different scale features, so as to obtain a more precise feature representation of RS scene images. Furthermore, the subspace is utilized to capture the shared characteristics of each category, to reduce the impact of the complex irrelevant backgrounds in the scene images. The results of our experiments conducted on the public available RS scene datasets demonstrate the strong competitiveness of our approach.

Few-shot remote sensing scene classification (FSRSSC) has become a study that has attracted widespread attention and aims to identify new scene classes through one or a few labeled scene images. Nevertheless, due to the existence of unrelated complex background in scene images, local descriptors that offer a more efficient representation than image-level features, will carry semantic information unrelated to the real semantics of the scene images. Concurrently, these irrelevant background local descriptors are also causing a large distribution bias in support and query sets, which leads to the problem of inaccurate feature representation of scene images. To address the above problems, we introduce a local descriptor-based rectification network called LDRNet. Within this network, we first design a local descriptor semantic rectification (LDSR) module. It performs semantic rectification on local descriptors that are unrelated to scene image semantics by obtaining a descriptor-level global-aware semantic representation. Secondly, we introduce a cross-set bias rectification (CSBR) module. It rectifies the query set by obtaining the offset between two sets (query and support) from a more detailed local descriptor perspective. This operation can shorten the distance among the two sets (query and support), thereby obtaining a more accurate representation of scene image features. Furthrmore, we employ a local descriptor-based contrastive loss function to guarantee that the rectified local descriptor semantics are consistent with the corresponding scene image. The comparative experimental result indicates that our LDRNet achieves state-of-the-art performance on three commonly used public datasets.

Speech enhancement (SE) models based on deep neural networks (DNNs) have shown excellent denoising performance. However, mainstream SE models often have high structural complexity and large parameter sizes, requiring substantial computational resources, which limits their practical application. In this paper, a high-efficiency encoder-decoder structure, inspired by the top-down attention mechanism in human brain perception and named human-like perception attention network (HPANet), is proposed for monaural speech enhancement, which is able to emulate brain perceptual attention in noise environments. In HPANet, the raw waveform is first encoded by using attention encoder to capture shallow global features. These features are then downsampled, and multi-scale information is aggregated through top attention module to prevent the loss of crucial information. Next, down attention module integrates features from neighboring layers to reconstruct signal in a top-down manner. Finally, the decoder reconstructs the denoised clean signal. Experiments show that the proposed method effectively reduces model complexity while maintaining competitive performance.