Search Results (47)

Single-view-based anomaly detection approaches present challenges due to the lack of context, particularly for multi-label problems. In this work, we demonstrate the efficacy of using multiview image data for improved classification using a hierarchical learning approach. Using 170,958 images from the International Mouse Phenotyping Consortium (IMPC) repository, a specimen-wise multiview dataset comprising 54,046 specimens was curated. Next, two hierarchical classification frameworks were developed by customizing ConvNeXT and a convolutional autoencoder (CAE) as CNN backbones, respectively. The customized architectures were trained at three hierarchy levels with increasing anatomical granularity, enabling specialized layers to learn progressively more detailed features. At the top level (L1), multiview (MV) classification performed about the same as single views, with a high mean AUC of 0.95. However, using MV images in the hierarchical model greatly improved classification at levels 2 and 3. The model showed consistently higher average AUC scores with MV compared to single views such as dorsoventral or lateral. For example, at Level 2 (L2), the model divided abnormal cases into three subclasses, achieving AUCs of 0.65 for DV, 0.76 for LV, and 0.87 for MV. Then, at Level 3 (L3), it further divided these into ten specific abnormalities, with AUCs of 0.54 for DV, 0.59 for LV, and 0.82 for MV. A similar performance was achieved by the CAE-driven architecture, with mean AUCs of 0.87, 0.88, and 0.89 at Level 2 (L2) and 0.74, 0.78, and 0.81 at Level 3 (L3), respectively, for DV, LV, and MV views. The overall results demonstrate the advantage of multiview image data coupled with hierarchical learning for skeletal abnormality detection in a multi-label context. Full article

The black box nature of artificial neural networks limits the understanding of internal mechanisms and processes that happen inside hidden layers. The introduction of deep neural networks and efficient layer-wise training methods has enabled researchers to study how features are learnt through different layers of neural networks. However, there has been limited research on mapping input features to neural network weights in order to understand how features are represented in the layers. This research proposes a novel component model to establish the relationship between input features and neural network weights. This will aid in optimizing transfer learning models by only extracting relevant weights instead of all the weights in the layers. The proposed model is evaluated using standard IRIS and a set of modified IRIS datasets. Classification experiments are conducted, and the results are evaluated to verify the quality of the dataset. A visualization of input features and components through the proposed model is presented using t-SNE to indicate the impact of changes in the input features. From the results, it is concluded that the proposed component model provides core knowledge in the form of weights representing the input features that are learnt through training. The proposed work will aid in designing component-based transfer learning, which would improve the speed. Also, the components could be used as pretrained testing models for similar work with large datasets. Full article

Background/Objectives: Ultrasound imaging is widely employed to assess kidney health and diagnose renal diseases. Accurate segmentation of renal structures in ultrasound images plays a critical role in the diagnosis and treatment of related kidney diseases. However, challenges such as speckle noise and low contrast still hinder precise segmentation. Methods: In this work, we propose an encoder–decoder architecture, named MAT-UNet, which incorporates two distinct attention mechanisms to enhance segmentation accuracy. Specifically, the multi-convolution pixel-wise attention module utilizes the pixel-wise attention to enable the network to focus more effectively on important features at each stage. Furthermore, the triple-branch multi-head self-attention mechanism leverages the different convolution layers to obtain diverse receptive fields, capture global contextual information, compensate for the local receptive field limitations of convolution operations, and boost the segmentation performance. We evaluate the segmentation performance of the proposed MAT-UNet using the Open Kidney US Data Set (OKUD). Results: For renal capsule segmentation, MAT-UNet achieves a Dice Similarity Coefficient (DSC) of 93.83%, a 95% Hausdorff Distance (HD95) of 32.02 mm, an Average Surface Distance (ASD) of 9.80 mm, and an Intersection over Union (IOU) of 88.74%. Additionally, MAT-UNet achieves a DSC of 84.34%, HD95 of 35.79 mm, ASD of 11.17 mm, and IOU of 74.26% for central echo complex segmentation; a DSC of 66.34%, HD95 of 82.54 mm, ASD of 19.52 mm, and IOU of 51.78% for renal medulla segmentation; and a DSC of 58.93%, HD95 of 107.02 mm, ASD of 21.69 mm, and IOU of 43.61% for renal cortex segmentation. Conclusions: The experimental results demonstrate that our proposed MAT-UNet achieves superior performance in multiple renal structure segmentation in ultrasound images. Full article

Single image super-resolution is the inverse problem of reconstructing a high-resolution image from its low-resolution counterpart. Although recent Transformer-based architectures leverage global context integration to improve reconstruction quality, they often overlook frequency-specific characteristics, resulting in the loss of high-frequency information. To address this limitation, we propose the Dual Frequency Adaptive Network (DFAN). DFAN first decomposes the input into low- and high-frequency components via Stationary Wavelet Transform. In the low-frequency branch, Swin Transformer layers restore global structures and color consistency. In contrast, the high-frequency branch features a dedicated module that combines Directional Convolution with Residual Dense Blocks, precisely reinforcing edges and textures. A frequency fusion module then adaptively merges these complementary features using depthwise and pointwise convolutions, achieving a balanced reconstruction. During training, we introduce a frequency-aware multi-term loss alongside the standard pixel-wise loss to explicitly encourage high-frequency preservation. Extensive experiments on the Set5, Set14, BSD100, Urban100, and Manga109 benchmarks show that DFAN achieves up to +0.64 dBpeak signal-to-noise ratio, +0.01 structural similarity index measure, and −0.01learned perceptual image patch similarity over the strongest frequency-domain baselines, while also delivering visibly sharper textures and cleaner edges. By unifying spatial and frequency-domain advantages, DFAN effectively mitigates high-frequency degradation and enhances SISR performance. Full article

Microbiome data analysis is essential for understanding the role of microbial communities in human health. However, limited data availability often hinders research progress, and synthetic data generation could offer a promising solution to this problem. This study aims to explore the use of machine learning (ML) to enrich an unbalanced dataset consisting of microbial operational taxonomic unit (OTU) counts of 148 samples, belonging to 61 patients. In detail, 34 samples are from 16 adenomatous polyps (AP) patients, while 114 samples are from 46 colorectal cancer (CRC) patients. Synthesis of AP and CRC samples was conducted using the Synthetic Data Vault Python library, employing a Gaussian Copula synthesiser. Subsequently, the synthesised data quality was evaluated using a logistic regression model in parallel with an optimised support vector machine algorithm (polynomial kernel). The data quality is considered good when neither of the two algorithms can discriminate between real and synthetic data, showing low accuracy, F1 score, and precision values. Furthermore, additional statistical tests were employed to confirm the similarity between real and synthetic data. After data validation, layer-wise relevance propagation (LRP) was performed on a deep learning classifier to extract important OTU features from the generated dataset, to discriminate between CRC patients and those affected by AP. Exploiting the acquired features, which correspond to unique bacterial taxa, ML classifiers were trained and tested to estimate the validity of such microorganisms in recognising AP and CRC samples. The simplified version of the original OTU table opens up opportunities for further investigations, especially in the realm of extensive data synthesis. This involves a deeper exploration and augmentation of the condensed data to uncover new insights and patterns that might not be readily apparent in the original, more complex form. Digging deeper into the simplified data may help us better grasp the biological or ecological processes reflected in the OTU data. Transitioning from this exploration, the synergy of ML and synthetic data enrichment holds promise for advancing microbiome research. This approach enhances classification accuracy and reveals hidden microbial markers that could prove valuable in clinical practice as a diagnostic and prognostic tool. Full article

In this study, a multi-physical and thermal–fluid–solid coupling model was developed to simulate the autoclave curing process of composite materials, aiming to explore the influence mechanism of the external flow field on the curing process. First, the extended layerwise method (XLWM) and finite volume method were adopted to simulate the composite laminates and heating airflows, respectively. Then, the thermo-chemical–mechanical-seepage analysis was carried out for the composite laminates. Considering the interaction between the airflows and laminates, a weak coupling method was proposed to solve the thermal–fluid–solid coupling problem, which consists of two parts: unidirectional coupling and bidirectional coupling. In numerical examples, the results of the two coupling schemes were compared, which indicated that the bidirectional coupling scheme consumed fewer computing resources but achieved similar accuracy. Full article

An important hurdle in medical diagnostics is the high-quality and interpretable classification of audio signals. In this study, we present an image-based representation of infant crying audio files to predict abnormal infant cries using a vision transformer and also show significant improvements in the performance and interpretability of this computer-aided tool. The use of advanced feature extraction techniques such as Gammatone Frequency Cepstral Coefficients (GFCCs) resulted in a classification accuracy of 96.33%. For other features (spectrogram and mel-spectrogram), the performance was very similar, with an accuracy of 93.17% for the spectrogram and 94.83% accuracy for the mel-spectrogram. We used our vision transformer (ViT) model, which is less complex but more effective than the proposed audio spectrogram transformer (AST). We incorporated explainable AI (XAI) techniques such as Layer-wise Relevance Propagation (LRP), Local Interpretable Model-agnostic Explanations (LIME), and attention mechanisms to ensure transparency and reliability in decision-making, which helped us understand the why of model predictions. The accuracy of detection was higher than previously reported and the results were easy to interpret, demonstrating that this work can potentially serve as a new benchmark for audio classification tasks, especially in medical diagnostics, and providing better prospects for an imminent future of trustworthy AI-based healthcare solutions. Full article

This paper introduces an progressive pruning method for Light DeHaze Networks, focusing on a static scene captured by a fixed camera environments. We develop a progressive pruning algorithm that aims to reduce computational complexity while maintaining dehazing quality within a specified threshold. Our key contributions include a fine-tuning strategy for specific scenes, channel importance analysis, and an progressive pruning approach considering layer-wise sensitivity. Our experiments demonstrate the effectiveness of our progressive pruning method. Our progressive pruning algorithm, targeting a specific PSNR(Peak Signal-to-Noise Ratio) threshold, achieved optimal results at a certain pruning ratio, significantly reducing the number of channels in the target layer while maintaining PSNR above the threshold and preserving good structural similarity, before automatically stopping when performance dropped below the target. This demonstrates the algorithm’s ability to find an optimal balance between model compression and performance maintenance. This research enables efficient deployment of high-quality dehazing algorithms in resource-constrained environments, applicable to traffic monitoring and outdoor surveillance. Our method paves the way for more accessible image dehazing systems, enhancing visibility in various real-world hazy conditions while optimizing computational resources for fixed camera setups. Full article

Accurately detecting the maturity and 3D position of flowering Chinese cabbage (Brassica rapa var. chinensis) in natural environments is vital for autonomous robot harvesting in unstructured farms. The challenge lies in dense planting, small flower buds, similar colors and occlusions. This study proposes a YOLOv8-Improved network integrated with the ByteTrack tracking algorithm to achieve multi-object detection and 3D positioning of flowering Chinese cabbage plants in fields. In this study, C2F-MLCA is created by adding a lightweight Mixed Local Channel Attention (MLCA) with spatial awareness capability to the C2F module of YOLOv8, which improves the extraction of spatial feature information in the backbone network. In addition, a P2 detection layer is added to the neck network, and BiFPN is used instead of PAN to enhance multi-scale feature fusion and small target detection. Wise-IoU in combination with Inner-IoU is adopted as a new loss function to optimize the network for different quality samples and different size bounding boxes. Lastly, ByteTrack is integrated for video tracking, and RGB-D camera depth data are used to estimate cabbage positions. The experimental results show that YOLOv8-Improve achieves a precision (P) of 86.5% and a recall (R) of 86.0% in detecting the maturity of flowering Chinese cabbage. Among them, mAP50 and mAP75 reach 91.8% and 61.6%, respectively, representing an improvement of 2.9% and 4.7% over the original network. Additionally, the number of parameters is reduced by 25.43%. In summary, the improved YOLOv8 algorithm demonstrates high robustness and real-time detection performance, thereby providing strong technical support for automated harvesting management. Full article

This paper presents a mathematical analysis of semantic convergence in transformer-based language models, drawing inspiration from the concept of fractal self-similarity. We introduce and prove a novel theorem characterizing the gradient of embedding similarity across layers. Specifically, we establish that there exists a monotonically increasing function that provides a lower bound on the rate at which the average cosine similarity between token embeddings at consecutive layers and the final layer increases. This establishes a fundamental property: semantic alignment of token representations consistently increases through the network, exhibiting a pattern of progressive refinement, analogous to fractal self-similarity. The key challenge addressed is the quantification and generalization of semantic convergence across diverse model architectures and input contexts. To validate our findings, we conduct experiments on BERT and DistilBERT models, analyzing embedding similarities for diverse input types. While our experiments are limited to these models, we empirically demonstrate consistent semantic convergence within these architectures. Quantitatively, we find that the average rates of semantic convergence are approximately 0.0826 for BERT and 0.1855 for DistilBERT. We observe that the rate of convergence varies based on token frequency and model depth, with rare words showing slightly higher similarities (differences of approximately 0.0167 for BERT and 0.0120 for DistilBERT). This work advances our understanding of transformer models’ internal mechanisms and provides a mathematical framework for comparing and optimizing model architectures. Full article

Computational models of homologous protein groups are essential in sequence bioinformatics. Due to the diversity and rapid evolution of viruses, the grouping of protein sequences from virus genomes is particularly challenging. The low sequence similarities of homologous genes in viruses require specific approaches for sequence- and structure-based clustering. Furthermore, the annotation of virus genomes in public databases is not as consistent and up to date as for many cellular genomes. To tackle these problems, we have developed VOGDB, which is a database of virus orthologous groups. VOGDB is a multi-layer database that progressively groups viral genes into groups connected by increasingly remote similarity. The first layer is based on pair-wise sequence similarities, the second layer is based on the sequence profile alignments, and the third layer uses predicted protein structures to find the most remote similarity. VOGDB groups allow for more sensitive homology searches of novel genes and increase the chance of predicting annotations or inferring phylogeny. VOGD B uses all virus genomes from RefSeq and partially reannotates them. VOGDB is updated with every RefSeq release. The unique feature of VOGDB is the inclusion of both prokaryotic and eukaryotic viruses in the same clustering process, which makes it possible to explore old evolutionary relationships of the two groups. VOGDB is freely available at vogdb.org under the CC BY 4.0 license. Full article

The existing Siamese trackers have achieved increasingly successful results in visual object tracking. However, the interactive fusion among multi-layer similarity maps after cross-correlation has not been fully studied in previous Siamese network-based methods. To address this issue, we propose a novel Siamese network for visual object tracking, named SiamSMN, which consists of a feature extraction network, a multi-scale fusion module, and a prediction head. First, the feature extraction network is used to extract the features of the template image and the search image, which is calculated by a depth-wise cross-correlation operation to produce multiple similarity feature maps. Second, we propose an effective multi-scale fusion module that can extract global context information for object search and learn the interdependencies between multi-level similarity maps. In addition, to further improve tracking accuracy, we design a learnable prediction head module to generate a boundary point for each side based on the coarse bounding box, which can solve the problem of inconsistent classification and regression during the tracking. Extensive experiments on four public benchmarks demonstrate that the proposed tracker has a competitive performance among other state-of-the-art trackers. Full article

The industrial adoption of material extrusion of polymers (MEX/P) is hindered by the geometric quality of manufactured parts. Contact image sensors (CISs), commonly used in flatbed scanners, have been proposed as a suitable technology for layer-wise characterization of contour deviations, paving the way for the application of corrective measures. Nevertheless, despite the high resolution of CIS digital images, the accurate characterization of layer contours in MEX/P is affected by contrast patterns between the layer and the background. Conventional edge-recognition algorithms struggle to comprehensively characterize layer contours, thereby diminishing the reliability of deviation measurements. In this work, we introduce a novel approach to precisely locate contour points in the context of MEX/P based on evaluating the similarity between the grayscale pattern near a particular tentative contour point and a previously defined gradient reference pattern. Initially, contrast patterns corresponding to various contour orientations and layer-to-background distances are captured. Subsequently, contour points are identified and located in the images, with coordinate measuring machine (CMM) verification serving as a ground truth. This information is then utilized by an adaptive edge-detection algorithm (AEDA) designed to identify boundaries in manufactured layers. The proposed method has been evaluated on test targets produced through MEX/P. The results indicate that the average deviation of point position compared to that achievable with a CMM in a metrology laboratory ranges from 8.02 µm to 13.11 µm within the experimental limits. This is a substantial improvement in the reliability of contour reconstruction when compared to previous research, and it could be crucial for implementing routines for the automated detection and correction of geometric deviations in AM parts. Full article

Few-shot segmentation (FSS) is a cutting-edge technology that can meet requirements using a small workload. With the development of China Aerospace Engineering, FSS plays a fundamental role in astronaut working scene (AWS) intelligent parsing. Although mainstream FSS methods have made considerable breakthroughs in natural data, they are not suitable for AWSs. AWSs are characterized by a similar foreground (FG) and background (BG), indistinguishable categories, and the strong influence of light, all of which place higher demands on FSS methods. We design a pixel-wise and class-wise network (PCNet) to match support and query features using pixel-wise and class-wise semantic cues. Specifically, PCNet extracts pixel-wise semantic information at each layer of the backbone using novel cross-attention. Dense prototypes are further utilized to extract class-wise semantic cues as a supplement. In addition, the deep prototype is distilled in reverse to the shallow layer to improve its quality. Furthermore, we customize a dataset for AWSs and conduct abundant experiments. The results indicate that PCNet outperforms the published best method by 4.34% and 5.15% in accuracy under one-shot and five-shot settings, respectively. Moreover, PCNet compares favorably with the traditional semantic segmentation model under the 13-shot setting. Full article

Additive manufacturing technologies collectively refer to a set of layer-wise deposition methods that typically rely on CAD-CAM approaches for obtaining products with a complex shape/geometry and high precision and reliability. If the additive manufacturing of polymers is relatively easy and scalable due to the low temperatures needed to obtain processable inks, using similar technologies to fabricate ceramic products is indeed more challenging and expensive but, on the other hand, allows for obtaining high-quality results that would not be achievable through conventional methods. Furthermore, the implementation of additive manufacturing allows for the addressing of some important concerns related to the environment and sustainability, including the minimization of resource depletion and waste production/disposal. Specifically, additive manufacturing technologies can provide improvements in energy consumption and production costs, besides obtaining less waste material and less CO₂ emissions, which are all key points in the context of the circular economy. After providing an overview of the additive manufacturing methods which are specifically applied to ceramics, this review presents the sustainability elements of these processing strategies, with a focus on both current and future benefits. The paucity of specific available studies in the literature—which are included and discussed in this review—suggests that the research on additive manufacturing sustainability in the field of ceramic materials is in the preliminary stage and that more relevant work still deserves to be carried out in the future to explore this fascinating field at the boundary among ceramics science/technology, production engineering and waste management. Full article