Remote sensing image segmentation is a key task in the field of remote sensing image processing, which can quickly and accurately identify the category and distribution of objects in the image. Although existing Convolutional Neural Networks(CNNs) can extract multi-scale information from images, they often overlook fine details and lack satisfactory interpretability. Therefore, we proposed the idea of combining torsion feature of differential geometric features to extract detailed information in the image. On this basis, we design the Local Torsion Feature Extraction Block (LTB), which can deeply mine the local geometric structure features in the image and further help the model to capture image details. At the same time, we also integrate the SS2D module of VMamba, so as to balance the learning tendency of the model for global information and local information. Finally, through the experimental verification on three public datasets (Vaihingen, Potsdam and LoveDA), our proposed module shows effectiveness. This result not only verifies the potential of torsion features in image processing, but also provides conjectures and inspirations for further exploration of the role of torsion features.

Remote sensing semantic segmentation is a critical technology in the field of remote sensing image processing, with broad applications in environmental monitoring, urban planning, disaster assessment, and resource exploration. Despite the transformative impact of convolutional neural networks (CNNs) on this domain, CNN-based methods often encounter limitations due to their localized receptive fields, which struggle to capture the global context necessary for accurate segmentation in complex remote sensing imagery. In this letter, a novel approach is presented for remote sensing semantic segmentation using a mamba-based model named PPmamba. The PPmamba model integrates Resblock and PPmamba within an encoder-decoder framework to effectively capture both local and global contextual information from high-resolution remote sensing images. Leveraging the strengths of the Mamba architecture, our model employs selective scanning to efficiently process long sequences, overcoming the limitations of traditional CNNs and transformers in handling large-scale images with complex scenes. Extensive experiments on two benchmark datasets (Potsdam and Vaihingen) demonstrate the superiority of our PPmamba model against state-of-the-art models, achieving significant improvements in segmentation results. The codes will be available athttps://github.com/Jerrymo59/PPMambaSeg.

In recent years, the utilization of spectral and spatial information from adjacent pixels has become one of the key focuses in hyperspectral image (HSI) classification research, and the advent of the attention mechanism has introduced innovative approaches to this classification. However, existing applications of the attention mechanism primarily concentrate on the central pixel and its similar counterparts, overlooking cases when the central pixel is a noisy outlier. To remedy this defect, we introduce an attention network with fractional-ordered mixed geometric (ANFMG) features. First, we design differential geometric attention to differentiate between noise, edge, and flat areas. When the central pixel is a noisy outlier, differential geometric attention helps the model minimize its reliance on its own spectral information and reduces the effect of surrounding noise. Second, to achieve optimal classification outcomes with limited training samples, we introduce the specific fractal detail extraction strategy for targeted extraction of detailed information and minimizing redundant input. Furthermore, to focus on the central pixel and its neighboring pixels while mitigating the influence of pixels that exhibit spectral variability, we devise a fractal and similarity measurement fusion (FSMF) module, allocating limited resources to crucial features. The efficacy of our approach is confirmed through experiments on five HSI datasets, with results indicating superior classification performance.

Recent progress on parse tree encoder for sentence representation learning is notable. However, these works mainly en-code tree structures recursively, which is not conducive to parallelization. On the other hand, these works rarely take into account the labels of arcs in dependency trees. To address both issues, we propose Dependency-Transformer, which applies a relation-attention mechanism that works in concert with the self-attention mechanism. This mechanism aims to encode the dependency and the spatial positional relations between nodes in the dependency tree of sentences. By a score-based method, we successfully inject the syntax information without affecting Transformer’s parallelizability. Our model outperforms or is comparable to the state-of-the-art methods on four tasks for sentence representation and has obvious advantages in computational efficiency.

Multispectral (MS) pansharpening involves fusing a low-spatial-resolution MS image and its associated high-spatial-resolution panchromatic image. Recently, convolutional neural network (CNN)-based fusion models have been widely used in pansharpening domain, but most of them treat diversity features equally and also neglect the contribution of multilevel features, thereby impeding the representation ability of CNNs. To deal with these issues, we propose a novel channel similarity attention fusion network (CSAFNet) in this letter, where several channel attention residual dense blocks (CARDBs) are stacked to fully exploit discriminative features, and then the features produced by all CARDBs are combined via a multilevel feature fusion module. Such network enables the network to focus on more informative features and make full use of them. Both visual and quantitative assessments validate the superior performance of the proposed network over the current pansharpening methods with respect to spectral fidelity and spatial enhancement.

In recent years, deep learning has been widely used in the field of medical image processing, such as identification of symptoms, detection of organ. Due to the complexity of medical images, in the model training, there are many parameters when deep learning is used for image classification, and it takes a long time. Therefore, the training process for neural networks needs to be optimized. The stochastic gradient descent with momentum (SGD) is a common optimization algorithm in deep learning, and the particle swarm optimization (PSO) is a classical and effective swarm intelligence optimization algorithm. These two methods have their own advantages and disadvantages. Combining the two algorithms to calculate parameters, this paper proposes a novel particle swarm optimization-stochastic gradient descent with momentum (PSO-SGD) algorithm, which can find the optimal solution of the network more quickly and improve the solution efficiency on the basis of ensuring the classification accuracy. This algorithm is verified on two data sets, namely Blood Cell Images Data Set (BCIDS) and COVED-19 Radiography Data Set (COVED19RDS). Experiments prove the effectiveness of the algorithm.

Intracranial hemorrhage is a serious craniocerebral injury. The automatic segmentation of intracranial hematoma is of great significance for the diagnosis of patients with intracranial hemorrhage. So far, image segmentation methods based on deep learning have made great progress. However, the accuracy and efficiency of hematoma segmentation still can not meet the requirements of clinical application due to the particularity of brain hematoma images, such as sparse samples, uneven sample distribution, diverse hematoma morphology and inconsistent scale. In order to solve these problems, this paper proposes a FewShot segmentation framework based on confrontation. The framework abandons the solution process based on fixed loss function in the existing segmentation model, and realizes confrontation learning through the confrontation between the discrimination model and the segmentation model. It not only solves the problem that hematoma samples are scarce and difficult to label, but also adopts the way of confrontation learning, so that the model pays more attention to the overall characteristics of hematoma rather than local information in the segmentation process, So as to make it more universal. At the same time, this paper designs a multi-scale convolution algorithm, which uses four different scale convolution cores to realize feature convolution and spatial fusion, so that the model can maintain a certain accuracy range when segmenting hematomas of different shapes and sizes, and solves the problem of multi-scale hematoma segmentation. In order to speed up the operation of the segmentation model, a mask is added to the image segmentation model for the first time to replace the reconstruction process from feature to segmented image, which greatly saves the computational cost. In the experiment, the open data set cq500 is used to develop the algorithm, and the image data provided by the First Affiliated Hospital of Chongqing Medical University are used to test. The results show that compared with the most advanced model, this method has the highest comprehensive evaluation index and significantly improves the segmentation efficiency.

Summary Particle swarm optimization (PSO) is an iterative computational methods which is used for obtaining the solutions of practical optimization problems. PSO is however, prone to be ended up obtaining a local optimum. Various strategies are proposed in the related literature to address this issue. Such strategies, however, often reduce the convergence speed of the algorithms. This article proposes a multiion particle swarm optimization (MION‐PSO) algorithm which incorporates three strategies to balance the exploration and exploitation abilities of the algorithm. To improve the exploration abilities, the particles are regarded as ions with repellent/attraction forces among them. The second strategy is a multiion strategy (MIS) in which the population is divided into many subswarms, namely, particle group. Using MIS, the optimal solution within each group is then used to purposefully guide the updates of other individuals. To delete the useless particles without the capability of updating historical best location for a specific period and have a misleading impact on others, without we employ a particle elimination strategy. Twenty‐three benchmark functions and eight baseline algorithms are employed to test the performance of MION‐PSO method. The experimental results show that MION‐PSO method significantly outperforms the baseline algorithms in terms of convergence, and further exhibits high capability in finding the optimal solutions especially on unimodal functions.

Coarse-grained reconfigurable architectures (CGRAs) have drawn increasing attention due to their flexibility and energy efficiency. Data flow graphs (DFGs) are often mapped onto CGRAs for acceleration. The problem of DFG mapping is challenging due to the diverse structures from DFGs and constrained hardware from CGRAs. Consequently, it is difficult to find a valid and high quality solution simultaneously. Inspired from the great progress in deep reinforcement learning (RL) for AI problems, we consider building methods that learn to map DFGs onto spatially programmed CGRAs directly from experiences. We propose RLMap, a solution that formulates DFG mapping on CGRA as an agent in RL, which unifies placement, routing and processing element insertion by interchange actions of the agent. Experimental results show that RLMap performs comparably to state-of-the-art heuristics in mapping quality, adapts to different architecture, and converges quickly.