The real-time collection of bus passenger object detection is an essential part of developing a smart bus system. The difficulty of object detection mainly lies in the objective factors, such as: clothing, hair style and accessories, light, etc. Traditional object detection methods with the artificial feature extraction suffers from insufficient strength in expression, generalization, and recognition rate. The object detection method based on deep learning mainly uses the convolutional neural network in deep learning to learn features from a large set of data. The learned features can describe the rich information inherent in the data, and improve the expression ability of the features as well as the recognition accuracy. Due to too many parameters of the Convolutional neural network (CNN) model, the amount of calculation is too large to be operated on the vehicle terminal. To reduce calculation burden and improve the operation speed, we employs the depthwise separable convolution method to optimize the convolutional layer of tiny YOLO network model. It decomposes a complete convolution operation into depthwise convolution and pointwise convolution, thus reducing the parameter amount of the CNN and improving the operation speed. The experiment results reveal that the speed of bus passenger object detection detected by our improved model is 4 times faster than the previous one but with the nearly same detection accuracy.

As one of the main sources of remote sensing big data, synthetic aperture radar (SAR) can provide all-day and all-weather Earth image acquisition. However, speckle noise in SAR images brings a notable limitation for its big data applications, including image analysis and interpretation. Deep learning has been demonstrated as an advanced method and technology for SAR image despeckling. Most existing deep-learning-based methods adopt supervised learning and use synthetic speckled images to train the despeckling networks. This is because they need clean images as the references, and it is hard to obtain purely clean SAR images in real-world conditions. However, significant differences between synthetic speckled and real SAR images cause the domain gap problem. In other words, they cannot show superior performance for despeckling real SAR images as they do for synthetic speckled images. Inspired by recent studies on self-supervised denoising, we propose an advanced SAR image despeckling method by virtue of Bernoulli-sampling-based self-supervised deep learning, called SSD-SAR-BS. By only using real speckled SAR images, Bernoulli-sampled speckled image pairs (input–target) were obtained as the training data. Then, a multiscale despeckling network was trained on these image pairs. In addition, a dropout-based ensemble was introduced to boost the network performance. Extensive experimental results demonstrated that our proposed method outperforms the state-of-the-art for speckle noise suppression on both synthetic speckled and real SAR datasets (i.e., Sentinel-1 and TerraSAR-X).

Synthetic aperture radar (SAR) images are often affected by a type of multiplicative noise known as “speckle” due to their active imaging characteristics.This property complicates the processing and interpretation of SAR images. While deep learning techniques have demonstrated success in despeckling many models are tailored to specific noise levels. This specificity can limit a model's ability to generalize to real SAR images with varying noise levels, potentially leading to over-smoothing or over-focusing on specific details. To address these challenges, we present the Adaptive Noise Estimation and Despeckling Network (ANED-Net). This network consists of a noise-level estimation phase and a noise-level-guided non-blind denoising phase. During the non-blind denoising phase, we develop a Noise Feature-Guided Denoising Network (NFGDN). This network integrates a hierarchical encoder-decoder denoising module based on the Transformer block (T-unet) and a Denoising Enhancement Control (DEC) block. Together, they skillfully capture both local and global dependencies inherent in SAR images, facilitating effective noise removal. Furthermore, we also introduce a Deep-attention mechanism to counteract the attentional collapse observed when the Transformer is extended in depth, enhancing the network's feature extraction capability and strengthening the model's denoising performance. Extensive tests on synthetic and real images show that ANED-Net is robust to different noise scenarios. It effectively mitigates speckle noise even at unspecified levels, and outperforms many established methods.

In this letter, we aim to address a synthetic aperture radar (SAR) despeckling problem with the necessity of neither clean (speckle-free) SAR images nor independent speckled image pairs from the same scene, and a practical solution for SAR despeckling (PSD) is proposed. First, an adversarial learning framework is designed to generate speckled-to-speckled (S2S) image pairs from the same scene in the situation where only single speckled SAR images are available. Then, the S2S SAR image pairs are employed to train a modified despeckling Nested-UNet model using the Noise2Noise (N2N) strategy. Moreover, an iterative version of the PSD method (PSDi) is also presented. Experiments are conducted on both synthetic speckled and real SAR data to demonstrate the superiority of the proposed methods compared with several state-of-the-art methods. The results show that our methods can reach a good tradeoff between feature preservation and speckle suppression.

Oriented remote sensing object detection (ORSOD) has gained increasing significance in both military and civilian applications due to the necessity of accurately identifying objects with varying shapes and orientations in remote sensing data. Traditional ORSOD methods often employ fixed label assignment strategies to differentiate between positive and negative samples. However, most of them frequently overlook the impact of object shape on sample quality, leading to an imbalanced distribution of positive samples and exacerbating the inconsistency between classification and regression tasks, thereby limiting detection performance. To address these challenges, we propose a novel shape-dependent assignment (SDA) method that dynamically differentiates positive and negative samples based on object shape. It introduces a new metric for evaluating sample box quality by considering angular differences relative to ground truth (GT) boxes and adjusts the sample scoring threshold according to the aspect ratio of each GT box. In addition, we present a DIoU-adaptive weighting (DAW) module that enhances the interaction between classification and regression tasks by leveraging the distance-IoU metric. This approach not only balances the quantity of samples but also improves their quality, enabling more effective training schemes for samples of varying qualities. We validate our proposed methods through extensive experiments on three challenging ORSOD datasets: DOTA-1.0, HRSC2016, and UCAS-AOD. The results demonstrate that our approach achieves significant improvements, especially for objects with large aspect ratios.

Sparse matrix-vector multiplication (SpMV) plays a critical role in a wide range of linear algebra computations, particularly in scientific and engineering disciplines. However, the irregular memory access patterns, extensive memory usage, high bandwidth requirements, and underutilization of parallelism hinder the computational efficiency of SpMV on GPUs. In this paper, we propose a novel approach called block-wise dynamic mixed-precision (BDMP) to address these challenges. Our methodology involves partitioning the original matrix into uniformly sized blocks, with each block’s size determined by considering architectural characteristics and accuracy requirements. Additionally, we dynamically assign precision to each block using a precision selection method that takes into account the value distribution of the original sparse matrix. We develop two distinct SpMV computation algorithms for BDMP: BDMP-PBP (Precision-based partitioning) and BDMP-TCKI (Tailored compression and kernel implementation). BDMP-PBP partitions the matrix into two independent matrices for separate computations based on block precision, offering flexibility for integration with other optimization techniques. Meanwhile, BDMP-TCKI focuses on achieving significant thread-level parallelism and memory utilization by tailoring an appropriate compressed storage format and kernel implementation for each block. We compare BDMP with NVIDIA’s cuSPARSE library and three state-of-the-art SpMV methods, including SELLP, MergeBase, and BalanceCSR, using matrices from the University of Florida’s SuiteSparse dataset collection. BDMP-PBP and BDMP-TCKI show average speedups up to 2.64 $$\times $$× and 2.91 $$\times $$× on Turing RTX 2080Ti, and up to 2.99 $$\times $$× and 3.22 $$\times $$× on Ampere A100. The results demonstrate that BDMP enables the optimization of computation speed without compromising the precision necessary for reliable results.