Aircraft detection in synthetic aperture radar (SAR) images is a challenging task because of the discreteness, variability, and interference of aircraft scattering features. This article proposes a new hybrid approach of scattering information enhancement (SIE) and an attention pyramid network (APN). It first extracts strong scattering points (SSPs) of aircraft via an adapted Harris-Laplace detector. These SSPs are then clustered into candidate scattering regions by density-based spatial clustering of applications with noise (DBSCAN) and are then modeled with a Gaussian mixture model (GMM). Target scattering clusters are discriminated from background clutter by measuring the Kullback-Leibler divergence (KLD) to the known target templates. These target scattering clusters are enhanced in the preprocessing stage. All the SIE-preprocessed images are then fed into the APN for training and testing. It is composed of the multiscale feature pyramid network (FPN) and the modified convolutional block attention module (CBAM) to cope with the discreteness and variability of aircraft. In addition, focal loss (FL) is adopted to deal with the issue of unbalanced sample distribution and the interference from hard samples. Experiments conducted on the Gaofen-3 and TerraSAR-X data sets demonstrate the effectiveness of the proposed method with an average precision (AP) of 83.25%.

Lack of interpretability and weak generalization ability have become the major challenges with data-driven intelligent SAR-ATR technology, especially in practical applications with azimuth-sparse training samples. A novel insight into SAR image representation with neural networks from a causal perspective is presented in this paper. Firstly, a causal model of SAR image representation conditioned on disentangled semantic factors is proposed. A set of SAR images is considered as a low-dimensional manifold, which is controlled by three semantic factors, namely, intrinsics, diversity, and randomness. A Causal Adversarial auto-Encoder (CAE) for SAR-ATR is then proposed to embody this disentangled representation, which incorporates a number of novel built-in network features. A physically reasonable Cyclic High-frequency information-based Embedding (CHE) method is proposed for azimuth encoding, which ensures the uniformity, continuity, periodicity, and distinctiveness of angle. A Symmetrically Conditional Encoding (SCE) module is established to constrain the semantic consistency of low-dimensional features. Besides, a hybrid loss function is designed, which is composed of latent adversarial loss, reconstruction loss, and task-oriented losses. Both representation and generalization abilities are thoroughly evaluated through qualitative visualization and quantitative comparison experiments on the MSTAR and FUSAR-Ship datasets. Experimental results demonstrate superior representation ability for the disentangled properties via angle-interpolation and target-transformation of SAR images. By using only 12 samples per-class, the proposed CAE can achieve an accuracy of 93.1% for the 10-target SAR-ATR classification task.

Deep neural networks (DNNs) have evolved to be the state-of-the-art technique for machine learning tasks. However, their high computational demands make them difficult to deploy on embedded devices with limited hardware resources and strict power budgets. Most embedded systems perform better with 8-bit data processing, prompting extensive research into 8-bit network quantization to enable faster inference. This paper aims to propose a unified 8-bit inference and training framework to support object detection tasks, striving to balance accuracy and speed for conventional convolutional neural networks (CNNs). Initially, this paper establishes a unified full int8 post-training quantization (PTQ) method using KL_divergence to evaluate the range of parameter distributions and thresholds before and after quantization. This method effectively addresses the quantization issues commonly found in networks with linear activations. For networks with nonlinear activations, this paper introduces a hybrid precision post-training quantization (H-PTQ) method that utilizes hybrid precision to perform forward inference, thereby mitigating quantization errors caused by nonlinear activation functions. Furthermore, quantization-aware training (QAT) typically employs the straight-through estimator (STE) for backward propagation of gradients through the quantization function. However, since STE is an approximate computation, this paper proposes an alternative called $\alpha$ -quantization-aware training ( $\alpha$ -QAT). This method replaces the quantized weights in the loss function with affine combinations of the quantized and full-precision weights, enabling more precise forward and backward propagation to fine-tune the errors introduced by quantization. Finally, this paper conducted evaluations of quantized networks on ARM platforms and performed experiments across multiple datasets. The results indicate that the proposed PTQ, H-PTQ, and $\alpha$ -QAT methods achieved maximum accelerations of 4x, 2.3x, and 3.9x, respectively. Additionally, these methods significantly reduced memory overhead by up to 57.11%, 43.16%, and 91.94%, and achieved model compression rates of up to 51.52%, 48.48%, and 49.70%.

Data augmentation is beneficial when the measured training data are insufficient to train a robust deep model. One of the promising techniques is to use simulated data generated by physics-based engines. For example, few-shot learning of synthetic aperture radar (SAR) targets could be benefited from simulated SAR images. However, the characteristics of the simulated training data significantly affect the performance of the trained model. Therefore, it is of great significance to evaluate the utility of simulated data objectively and effectively. A recognition rate versus substitution rate curve (RSC)-based assessment criterion is proposed, consisting of substitution rate (SR)-based dataset allocation stage and RSC-based evaluation stage. First, the differential dataset allocation is performed under a progressive SR to obtain paired reference and comparison training sets. Then, the reference and comparison classifiers are trained under different SRs using the same network and parameter configuration in the RSC criterion-based evaluation stage. AconvNet and AlexNet are selected as the backbones of the evaluation network. Especially, k-fold cross-validation is applied to alleviate selection bias. The difference between the integrals of RSCs is defined as the RSC score for the simulated dataset. Experiments conducted on the measured and simulated moving and stationary target acquisition and recognition (MSTAR) database demonstrate the rationality and validity of the proposed RSC criterion. Specifically, multisource simulated datasets are adopted, including the adversarial autoencoder-generated and electromagnetic simulation datasets. The proposed RSC criterion shows promising utility evaluation ability, flexibility, and extensibility compared with traditional full-reference image-quality assessment criteria.

To improve the performance of rural power load forecasting, a new power load forecasting method is proposed. Firstly, basic and statistical features are extracted from historical power load data and numerical weather prediction (NWP) data. Then, t-SNE dimensionality reduction is used to extract main features, and K-means clustering is used to distinguish seasonal features. Secondly, the deep learning network CNN-LSTM (fusion of convolutional neural network and long short-term memory) is designed, and the load prediction models for spring and autumn, summer and winter are trained separately. Finally, use a rural power load in Wuxi, China to predict the power load in the next 24 hours and calculate the error. Experimental results show that the forecasting error of the proposed method is lower than that of LSTM, ARIMA and other prediction models.

Data-driven algorithms based on deep neural networks (DNNs) for ship detection in synthetic aperture radar (SAR) images are restricted by limited training samples and complex background interference. Inspired by the sparsity and neighborhood relevance of ships in SAR images, a novel detection algorithm based on self-attention dictionary learning (SADL) is proposed in this paper, which only requires a few samples for training. A self-attention mechanism is injected to learn discriminative features between classes, which can extract inherent information from sequences adaptively. Particularly, a hybrid loss function is tailored for the representations of multi-classes targets using SADL, which consists of the reconstruction error, minimal intra-class error, maximum inter-class error, and exclusiveness error. Further, a SADL-based ship detection method is proposed by building sub-dictionaries of the target and background, respectively. The gradient and intensity information in a fixed neighborhood are used to construct the feature dictionary to suppress complex background interference and provide effective prior knowledge. Experiments conducted on the large-scale SAR ship detection dataset (LS-SSDD-v1.0) demonstrate the effectiveness of the proposed method, which achieves an F1-score of 0.47 on 3000 test images with only 3 training images.