Kun Liu

Synthetic Aperture Radar (SAR) is widely used in military and civilian applications due to its all-weather, day-and-night imaging capability. However, interpreting SAR images is challenging for both experts and non-experts. Inspired by deep learning-based style transfer, researchers have employed Generative Adversarial Networks (GANs) to convert SAR images into more intuitive optical ones. Yet, current loss functions and evaluation metrics focus mainly on pixel-level differences, overlooking the structural coherence required for target recognition and downstream tasks. To address this, we propose a semantic segmentation-driven framework for target-level SAR-to-optical image translation. Compatible with various supervised models, it incorporates segmentation loss and uses SAR segmentation maps as additional inputs to preserve target structure. Experiments on custom datasets, built from Sentinel 1-2 imagery with road binary segmentation labels, as well as public datasets, confirm the method's effectiveness across different base translation models. The source code and the datasets used will be published at the following URL https://github.com/NWPU-LHH/SOIT-Seg-Driven.

Few-shot semantic segmentation (FSS) in remote sensing imagery remains challenging due to complex scenes, large image sizes, frequent multi-class coexistence, and significant intra-class variation. These factors often lead to severe overfitting to base classes and hinder generalization to novel classes. To address these issues, we propose a Mamba-Enhanced Transformer framework with Selective Prototype Aggregation (SPA), which simultaneously suppresses base-class interference and enhances novel-class representation. Specifically, the Dynamic Feature Aggregation (DFA) module integrates features from a base-class learner and a meta-learner, effectively enhancing inter-class distinction and mitigating feature contamination in multi-class coexistence scenarios. We further introduce the Mamba-Enhanced Transformer (MET), which combines Mamba blocks and Transformer self-attention to capture long-range dependencies and multi-scale context while maintaining low computational complexity. Mamba blocks independently extract support and query features, preventing cross-branch interference and improving robustness. By selectively aggregating support prototypes with query self-support prototypes, SPA strengthens feature discriminability and leverages query-specific contextual cues to reduce intra-class variance. Our model demonstrates strong performance across multiple datasets, achieves 50.49% mIoU on the iSAID dataset and 30.82% on the LoveDA dataset, representing improvements of 3.95% and 3.89%, respectively, over the Base model. Ablation experiments further validate the superior performance of SPA.

Optical remote sensing and Synthetic Aperture Radar(SAR) remote sensing are crucial for earth observation, offering complementary capabilities. While optical sensors provide high-quality images, they are limited by weather and lighting conditions. In contrast, SAR sensors can operate effectively under adverse conditions. This letter proposes a GAN-based SAR-to-optical image translation method named Seg-CycleGAN, designed to enhance the accuracy of ship target translation by leveraging semantic information from a pre-trained semantic segmentation model. Our method utilizes the downstream task of ship target semantic segmentation to guide the training of image translation network, improving the quality of output optical-styled images. The potential of foundation-model-annotated datasets in SAR-to-optical translation tasks is revealed. This work suggests broader research and applications for downstream-task-guided frameworks. The code and link to download proposed HRSID-DIOR dataset will be available at https://github.com/NPULHH/.

In semi-supervised learning (SSL) for medical image classification, model performance is often hindered by the scarcity of labeled data and the complexity of unlabeled data. This paper proposes an enhanced SSL approach to address these challenges by effectively utilizing unlabeled data through a combination of pseudo-labeling and contrastive learning. The key contribution of our method is the introduction of a Dynamic Sample Reweighting strategy to select reliable unlabeled samples, thereby improving the model’s utilization of unlabeled data. Additionally, we incorporate multiple data augmentation strategies based on the Mean Teacher (MT) model to ensure consistent outputs across different perturbations. To better capture and integrate multi-scale features, we propose a novel feature fusion network, the Medical Multi-scale Feature Fusion Network (MedFuseNet), which enhances the model’s ability to classify complex medical images. Finally, we introduce a pseudo-label guided contrastive learning (PGC) loss function that improves intra-class compactness and inter-class separability of the model’s feature representations. Extensive experiments on three public medical image datasets demonstrate that our method outperforms existing SSL approaches, achieving 93.16% accuracy on the ISIC2018 dataset using only 20% labeled data, highlighting the potential of our approach to advance medical image classification under limited supervision.

The development of medical image classification models necessitates a substantial number of labeled images for model training. In real-world scenarios, sample sizes are typically limited and labeled samples often constitute only a small portion of the dataset. This paper aims to investigate a collaborative similarity learning strategy that optimizes pseudo-labels to enhance model accuracy and expedite its convergence, known as the joint similarity learning framework. By integrating semantic similarity and instance similarity, the pseudo-labels are mutually refined to ensure their quality during initial training. Furthermore, the similarity score is utilized as a weight to guide samples away from misclassification predictions during the classification process. To enhance the model’s generalization ability, an adaptive consistency constraint is introduced into the loss function to improve performance on untrained datasets. The model achieved a satisfactory accuracy of 93.65% at 80% labeling ratio, comparable to supervised learning methods’ performance. Even with very low labeling ratio (e.g., 5%), the model still attained an accuracy of 74.28%. Comparison with other techniques such as Mean Teacher and FixMatch revealed that our approach significantly outperforms them in medical image classification tasks through improving accuracy by approximately 2%, demonstrating this framework’s leadership in medical image classification.

Deep learning has substantially improved the state-of-the-art in object detection and image classification. Deep learning usually requires large-scale labelled datasets to train the models; however, due to the restrictions in medical data sharing and accessibility and the expensive labelling cost, the application of deep learning in medical image classification has been dramatically hindered. In this study, we propose a novel method that leverages semi-supervised adversarial learning and pseudo-labelling to incorporate the unlabelled images in model learning. We validate the proposed method on two public databases, including ChestX-ray14 for lung disease classification and BreakHis for breast cancer histopathological image diagnosis. The results show that our method achieved highly effective performance with an accuracy of 93.15% while using only 30% of the labelled samples, which is comparable to the state-of-the-art accuracy for chest X-ray classification; it also outperformed the current methods in multi-class breast cancer histopathological image classification with a high accuracy of 96.87%.

In practice, the statistics of cargo loading capacity needs to be realized by counting the loading actions. Object detection algorithm may not meet the requirements of such tasks under the influence of environmental factors. And the current temporal action localization methods may ignore the spatial information of objects and identify the inaccurate action time boundary of multiple objects in the cargo loading task. To fill this technical gap, this paper designs a cargo loading identification framework. The core of this framework is the Target Area Rise-descend (TAR) algorithm, which is based on the object detection model and uses the regularity of object location and object size to realize real-time recognition of the cargo loading process. We ensure the real-time and robustness of the algorithm by some means like nap mechanism and time regularity. The garbage truck recycling task is taken as an example to complete the elaboration of the proposed method. Experiments show that the accuracy of the proposed method is higher than 99%, and the real-time performance is higher than 20fps on Neural Network Computing Unit (NPU). Considering the limitations of intelligent application scenarios and resources, self-powered camera is applied. For mechanical motion similar to the scene in this paper, the proposed algorithm can use extremely low computing power to capture the motion process and achieve data statistics.

This article conducts a retrospective analysis of 500 patients with posttraumatic elbow dysfunction admitted to our department from March 2019 to September 2020. The average time from injury to operation is 11 months (2–20 months). We adopt a personalized treatment method to completely remove the hyperplastic adhesion tissue and heterotopic ossification around the joint, remove part of the joint capsule and ligament, and release it to achieve maximum function. After the operation, an external fixator was used to stabilize the loosened elbow joint, and the patient was guided to perform rehabilitation exercises with the aid of a hinged external fixator, and celecoxib was used to prevent heterotopic ossification. Mayo functional scoring system was used to evaluate the curative effect before and after surgery. The rapid realization of ultrasound imaging under the framework of compressed sensing is studied. Under the premise of ensuring the quality of ultrasound imaging reconstruction, the theory of ultrasound imaging is improved, and a plane wave acoustic scattering ultrasound echo model is established. On this basis, the theory of compressed sensing is introduced, the mathematical model of compressed sensing reconstruction is established, and the fast iterative shrinkage thresholding algorithm (FISTA) of compressed sensing reconstruction is improved to reduce the computational complexity and the number of iterations. This article uses FISTA directly to reconstruct medical ultrasound images, and the reconstruction results are not ideal. Therefore, a simulation model of FISTA training and testing was established using the standard image library. By adding different intensities of noise to all images in the image library, the influence of noise intensity on the quality of FISTA reconstructed images is analyzed, and it is found that the FISTA model has requirements for the quality of the images to be reconstructed and the training set images. In this paper, Rob’s blind deconvolution restoration algorithm is used to preprocess the original ultrasound image. The clarity of the texture details of the restored ultrasound image is significantly improved, and the image quality is improved, which meets the above requirements. This paper finally formed a reconstruction model suitable for ultrasound images. The reconstruction strategy verified by the ultrasound images provided by the Institute of Ultrasound Imaging of a medical university has achieved a significant improvement in the quality of ultrasound images.

Ship target classification plays an important role in tasks such as maritime traffic control, maritime target tracking, and military reconnaissance. The complex ocean environment often causes obscuration of the ship targets, thus resulting in low accuracy of the obscured targets. This article presents a novel target classification algorithm—improved InceptionV3 and center loss convolution neural network (IICL-CNN)—based on the well-established inception network to improve the accuracy of obscured targets. This algorithm features a new objective function, which is designed to learn common features of both the clear samples and the obscured samples and, in the meantime, reduce the intraclass distance among the obscured samples. Experiments were performed on an optical remote sensing image dataset which consisted of 48 000 ship images in nine categories. The proposed method demonstrated superior performance on the obscured ship targets compared to the original InceptionV3 model. On average, the accuracy was 4.23%, 5.98%, and 17.48% higher on the ship targets that were occluded by levels of 30%, 50%, and 70%, respectively. Our experimental results showed that the proposed IICL-CNN could effectively improve the accuracy of the ship targets at various occlusion levels.