Wood surface defect inspection is a key step in production-line quality control. However, supervised detectors typically require large-scale annotations and often generalize poorly to unseen defect types. In this paper, we formulate wood defect inspection as an unsupervised anomaly detection problem and propose CoMemRet, a framework based on multiview contrastive learning and memory bank retrieval. CoMemRet trains a multiview contrastive encoder in the Lab color space to decouple luminance structure from chromatic variation, and introduces a patch-level contrastive constraint to enhance representations of fine-grained texture anomalies. The extracted features are further refined by feature gating and fusion modules to obtain more stable and compact embeddings. During anomaly detection, CoMemRet constructs a memory bank from a coreset of normal features and computes anomaly scores from the k-NN distances between test features and the normal-feature coreset. A clustering-guided coreset subsampling strategy is further introduced to improve the efficiency of coreset construction and k-NN retrieval while preserving coverage of normal patterns. Experiments on the newly collected Wood Surface Anomaly Detection Dataset (WSAD) and the public MVTec AD (wood) benchmark demonstrate competitive performance and improved generalization, suggesting suitability for production-line inspection.

Interactive medical image segmentation plays a crucial role in medical diagnosis, treatment planning, and in-terventional procedures. With the rapid advancement of artificial intelligence and deep learning technologies, this field has not only seen significant improvements in the accuracy and efficiency of segmenting anatomical structures and pathological areas from medical images, but it has also made substantial strides in the development of user-centric interactive tools. The introduction of the Segmentation Anything Model (SAM) represents a significant expansion of prompt-driven approaches within the domain of image segmentation, introducing a plethora of previously untapped functionalities. However, due to the substantial differences between natural and medical images, the generalization capabilities of interactive segmentation models across different imaging techniques remain limited, necessitating the ongoing development of targeted interactive segmentation models. In this work, we provide a comprehensive overview aimed at extending the efficacy of SAM to various medical imaging modalities, encompassing diverse image analyses and domains. Additionally, we explore potential research directions for SAM in different medical imaging fields. Despite these advancements, the accuracy and robustness of segmentation models still largely depend on high-quality annotated data, which is costly and time-consuming to acquire. These research efforts herald significant forthcoming advancements in interactive medical image segmen-tation technology, with the potential to greatly improve patient outcomes and streamline medical procedures, thereby driving further development and innovation in the healthcare industry.

Accurate and reproducible assessment of myocardial conditions is essential for the diagnosis of prior infarctions, cardiomyopathies, and inflammatory diseases. Although cardiac magnetic resonance imaging (CMR) is regarded as the gold standard for evaluating myocardial anatomy and function, manual segmentation remains labor-intensive and susceptible to variability. The demand for high-resolution images, dense predictions, and comprehensive segmentation across all phases of the cardiac cycle introduces significant challenges for Cine-CMR analysis. To address these issues, we present EML-SAM, an innovative interactive segmentation model based on the Segment Anything Model (SAM). EML-SAM incorporates an early-mid-late (EML) multi-scale fusion strategy, effectively mitigating the premature dilution of interaction information, thereby reducing segmentation errors and enhancing feature utilization. Furthermore, the model utilizes a multi-scale linear attention (MLA) transformer to establish an efficient global receptive field. Comprising 12 Transformer layers, the model integrates multi-scale attention with multi-scale perceptrons, striking a balance between computational efficiency and capacity. The proposed methodology offers a robust and efficient solution for high-resolution Cine-CMR segmentation, delivering accurate and real-time segmentation performance throughout the various stages of temporal changes within the cardiac cycle.

Fabric defect detection is critical for quality control in the textile industry. However, current challenges include the scarcity of high-quality datasets, the difficulty of detecting small defects in complex environments, and suboptimal detection performance for elongated defects. To address these issues, we propose an improved Real-Time Detection Transformer (RT-DETR) model specifically designed for fabric defect detection in industrial settings. A custom fiber defect dataset was created with random augmentations for improved diversity. We designed the Global Attention-based Instance Feature Interaction (GA-AIFI) module to enhance local feature extraction, improving the model's detection capability for small defects. Additionally, to address the prevalence of elongated defects in industrial fabric production, we introduced the Aspect Ratio-aware Distance-IoU (AR-DIoU) loss function, which further improves bounding box localization accuracy. Experiments conducted on both the custom dataset and the MVTec dataset demonstrate significant improvements, with [email protected] increasing by 3.5% and 3.6%,respectively, confirming the model's effectiveness.

Detection of wood surface defects is one critical step in the plywood process. YOLO series detection models need to find a good tradeoff between detection speed and accuracy. Therefore, this paper improves the YOLOv8 model by replacing the backbone, neck, and head with the FasterNet, Enhanced Feature Pyramid Network (EFPN), and Multi-Scale Feature Decoupling Head (MSDH), respectively. The FasterNet reduces the number of model parameters by using Partial Convolution (PConv). To remedy the accuracy loss caused by the FasterNet, the EFPN adds auxiliary branches to reduce information loss between different feature layers and uses Dilated Reparameter Modules (DRM) to extract more multi-scale features. Besides, the MSDH reorganizes the features on different scales to enhance their semantic relationships. Experimental results show that FasterNet reduces the number of parameters by 43%, the EFPN and the MSDH increase the mAP (mean Average Precison) by 2.2% and 1.4%, respectively, compared to the original Yolov8 model.

Field-road mode mining (FRMM) has gained increasing attention because of its crucial role in machinery management. As the Global Navigation Satellite System (GNSS) is widely applied in agricultural machinery, many methods based on motion and spatial-temporal features of GNSS trajectory data have been proposed for FRMM. However, these methods ignore the density and parallel features of GNSS points. The two features are useful for FRMM because field points usually have a higher density and more parallel points than road points. Therefore, a neural network driven by density and parallel features is proposed for accurate FRMM. Firstly, a statistical method is designed to extract the density feature (i.e., the number of its neighbor points) and parallel feature (i.e., the number of approximately parallel points in its neighbors) of each point. Then, the two features and eight motion features (e.g., speed, direction, etc.) are fed into a neural network to extract valuable latent features. Finally, a linear classifier is used to identify the field and road categories of GNSS points based on the latent features. Experimental results show that our method outperforms state-of-the-art methods and achieves the accuracy of 91.69% and 86.44% on public Wheat and Paddy datasets, respectively.

The Global Navigation Satellite System Network Reference Stations (GNSS-NRS) are pivotal in modern po-sitioning and navigation applications. However, GNSS-NRS face significant communication challenges, including instability during rapid user movement and inconsistent coverage by different operators, which often results in data transmission interruptions and delays. Moreover, the conventional Multipath Transmission Control Protocol (MPTCP) scheduling strategy does not accommodate the diversity of the data types managed by GNSS-NRS. This default approach treats all data uniformly, even under deteriorating network conditions. To overcome these limitations, this paper introduces a ‘Multiple Reception’ strategy based on MPTCP that ensures timely and reliable GNSS data transmission. This novel strategy differentiates between observational and control data based on their real-time criticality and applies tailored transmission techniques. Under adverse conditions, it utilizes multiple transmission paths and selectively receives control data, thus prioritizing the transfer of critical information. We conducted simulation experiments using MiniNet to compare the ‘Multiple Reception’ method against conventional and redundant strategies, demonstrating that it significantly reduces data loss rates in GNSS-NRS.

The seedling stage is important in the growth and development of maize, and it is also a critical period that affects maize yield and quality. Accurately recognizing this stage of maize is challenging, as current maize seedling detection methods struggle with small sizes and complex environments. This paper proposes a rapid maize seedling detection method that emphasizes spatial feature extraction in the receptive field and cross-dimensional interaction—receptive field triplet attention YOLO (RFT-YOLO). First, we design a convolutional module for cross-dimensional information interaction within the receptive field. Second, we introduce an advanced selective fusion network, which boosts the multi-scale fusion capability of small object features while reducing the GFLOPs of the model. Additionally, we employ Inner-IoU to diminish sensitivity to positional biases of small objects and expedite the convergence of the loss function. Finally, we create a dataset of maize seedlings to conduct our experiments. Experimental results demonstrate that the RFT-YOLO significantly enhances performance, reducing the number of parameters by 35.3% and GFLOPs by 14.24%, compared to the baseline model. Moreover, the mean average precision (mAP) improves from 89.6% to 91.5%. These improvements confirm the model's effectiveness in detecting maize seedlings.

Change detection is a crucial task in the field of remote sensing, focusing on identifying areas where significant changes occur between two remote sensing images captured at different time points. However, existing change detection methods based on CNNs or Transformers face two main challenges. Firstly, CNNs often lack the ability to effectively model long-range dependencies, resulting in diminished recognition performance for targets sharing the same semantics but having different features. Secondly, Transformers leverage self-attention mechanisms to capture long-range dependencies adeptly but struggle to capture local details and multi-scale features effectively. In response to these challenges, we propose a CNN and Transformer Aggregation Network (CTA-Net) for change detection in remote sensing images. Specifically, we devise two encoders based on Transformers and CNNs, respectively, facilitating the generation of complementary global and local features during the encoding phase. In the decoding phase, we design a pyramid-structured complementary decoder to aggregate multi-level complementary features from the CNN and Transformer branches. Furthermore, we propose a novel skip connection strategy to exploit the relationship between decoded features and multi-level complementary features, thereby enhancing the model's multi-scale invariance. Extensive experiments conducted on multiple benchmark datasets validate the effectiveness of our proposed CTA-Net model.