Extract Salient Features Research Articles

Multi-task learning with cross-task consistency for improved depth estimation in colonoscopy

Colonoscopy screening is the gold standard procedure for assessing abnormalities in the colon and rectum, such as ulcers and cancerous polyps. Measuring the abnormal mucosal area and its 3D reconstruction can help quantify the surveyed area and objectively evaluate disease burden. However, due to the complex topology of these organs and variable physical conditions, for example, lighting, large homogeneous texture, and image modality estimating distance from the camera (aka depth) is highly challenging. Moreover, most colonoscopic video acquisition is monocular, making the depth estimation a non-trivial problem. While methods in computer vision for depth estimation have been proposed and advanced on natural scene datasets, the efficacy of these techniques has not been widely quantified on colonoscopy datasets. As the colonic mucosa has several low-texture regions that are not well pronounced, learning representations from an auxiliary task can improve salient feature extraction, allowing estimation of accurate camera depths. In this work, we propose to develop a novel multi-task learning (MTL) approach with a shared encoder and two decoders, namely a surface normal decoder and a depth estimator decoder. Our depth estimator incorporates attention mechanisms to enhance global context awareness. We leverage the surface normal prediction to improve geometric feature extraction. Also, we apply a cross-task consistency loss among the two geometrically related tasks, surface normal and camera depth. We demonstrate an improvement of 15.75% on relative error and 10.7% improvement on δ1.25 accuracy over the most accurate baseline state-of-the-art Big-to-Small (BTS) approach. All experiments are conducted on a recently released C3VD dataset, and thus, we provide a first benchmark of state-of-the-art methods on this dataset.

Multi-task OCTA image segmentation with innovative dimension compression

Optical Coherence Tomography Angiography (OCTA) plays a crucial role in the early detection and continuous monitoring of ocular diseases, which relies on accurate multi-tissue segmentation of retinal images. Existing OCTA segmentation methods typically focus on single-task designs that do not fully utilize the information of volume data in these images. To bridge this gap, our study introduces H2C-Net, a novel network architecture engineered for simultaneous and precise segmentation of various retinal structures, including capillaries, arteries, veins, and the fovea avascular zone (FAZ). At its core, H2C-Net consists of a plug-and-play Height-Channel Module (H2C) and an Enhanced U-shaped Network (GPC-Net). The H2C module cleverly converts the height information of the OCTA volume data into channel information through the Squeeze operation, realizes the lossless dimensionality reduction from 3D to 2D, and provides the "Soft layering" information by unidirectional pooling. Meanwhile, in order to guide the network to focus on channels for training, U-Net is enhanced with group normalization, channel attention mechanism, and Parametric Rectified Linear Unit (PReLU), which reduces the dependence on batch size and enhances the network's ability to extract salient features. Extensive experiments on two subsets of the publicly available OCTA-500 dataset have shown that H2C-Net outperforms existing state-of-the-art methods. It achieves average Intersection over Union (IoU) scores of 82.84 % and 88.48 %, marking improvements of 0.81 % and 1.59 %, respectively. Similarly, the average Dice scores are elevated to 90.40 % and 93.76 %, exceeding previous benchmarks by 0.42 % and 0.94 %. The proposed H2C-Net exhibits excellent performance in OCTA image segmentation, providing an efficient and accurate multi-task segmentation solution in ophthalmic diagnostics. The code is publicly available at: https://github.com/IAAI-SIT/H2C-Net.

Automated EEG-based language detection using directed quantum pattern technique

Electroencephalogram (EEG) signals contain complex useful information about brain activities. These EEG signals are noisy, highly varying and nonstationary in nature. Hence, extracting meaningful information from these signals is challenging. The existing machine learning systems struggle to capture the minute changes from the signals and yield high performance.This study introduces a novel quantum-inspired feature extraction technique called Directed Quantum Pattern (DQP), designed to address these challenges by using a lattice structure to capture directional binary features. These directions (paths) are computed using a maximum function providing a dynamic and adaptive feature representation.This paper presents a novel DQP-LangNet developed using DQP for automated classification of two- languages using EEG signals. We have proposed a hybrid approach, combining DQP, statistical features, and multi-level discrete wavelet transform (MDWT) to extract salient features similar to the deep learning approach. The EEG dataset consisting of 14 channels, produces 7 feature vectors per channel, yielding 98 feature vectors. Neighborhood component analysis and Chi-square (Chi2) feature selection approaches generated 196 feature vectors.In addition to the innovative feature extraction a new classification structure called “t” is proposed k-nearest neighbor (tkNN) and support vector machine (tSVM) classifiers are employed. Using the proposed tkNN and tSVM classifiers, 392 (=196×2) classifier-based outcomes are obtained. To further improve classification performance, we applied the iterative majority voting (IMV) technique to automatically select the best result.Our DQP-based model achieved a classification accuracy of 95.68 %using EEG language dataset with leave-one-subject-out (LOSO) cross-validation strategy. Also, an explainable feature engineering (XFE) structure of DQP-LangNet is employed to obtain channel-specific explainable results. Our proposed DQP-LangNet model can be employed for other applications in neuroscience.

GMRD

The rapid development of social media has allowed people to access information through multiple channels, but social media has also become a breeding ground for rumors. Rumor detection models can effectively assess the credibility of information. However, current research mainly relies on text or combined text and image features, which may not be sufficient to capture complex feature information. Therefore, this paper proposes a rumor detection model based on the graph convolutional network (GCN) and multi-modal features. The proposed model constructs a knowledge graph (KG) and leverages the GCN to extract complex relationships between its nodes. Then, an interactive attention network is adopted to deeply integrate features. Furthermore, ResNet101 is utilized to extract salient features from images, addressing the challenges related to fully utilizing additional feature information and capturing text and image features at a deeper level to some extent. Multiple experiments conducted on datasets from Twitter and Weibo platforms demonstrate the efficacy of the proposed approach.

YOSMR: A Ship Detection Method for Marine Radar Based on Customized Lightweight Convolutional Networks

In scenarios such as nearshore and inland waterways, the ship spots in a marine radar are easily confused with reefs and shorelines, leading to difficulties in ship identification. In such settings, the conventional ARPA method based on fractal detection and filter tracking performs relatively poorly. To accurately identify radar targets in such scenarios, a novel algorithm, namely YOSMR, based on the deep convolutional network, is proposed. The YOSMR uses the MobileNetV3(Large) network to extract ship imaging data of diverse depths and acquire feature data of various ships. Meanwhile, taking into account the issue of feature suppression for small-scale targets in algorithms composed of deep convolutional networks, the feature fusion module known as PANet has been subject to a lightweight reconstruction leveraging depthwise separable convolutions to enhance the extraction of salient features for small-scale ships while reducing model parameters and computational complexity to mitigate overfitting problems. To enhance the scale invariance of convolutional features, the feature extraction backbone is followed by an SPP module, which employs a design of four max-pooling constructs to preserve the prominent ship features within the feature representations. In the prediction head, the Cluster-NMS method and α-DIoU function are used to optimize non-maximum suppression (NMS) and positioning loss of prediction boxes, improving the accuracy and convergence speed of the algorithm. The experiments showed that the recall, accuracy, and precision of YOSMR reached 0.9308, 0.9204, and 0.9215, respectively. The identification efficacy of this algorithm exceeds that of various YOLO algorithms and other lightweight algorithms. In addition, the parameter size and calculational consumption were controlled to only 12.4 M and 8.63 G, respectively, exhibiting an 80.18% and 86.9% decrease compared to the standard YOLO model. As a result, the YOSMR displays a substantial advantage in terms of convolutional computation. Hence, the algorithm achieves an accurate identification of ships with different trail features and various scenes in marine radar images, especially in different interference and extreme scenarios, showing good robustness and applicability.

Enhancing COVID-19 detection through integrated CNN-LSTM models: Addressing the challenge of varied radiological presentations

In an era where COVID-19's varied radiological presentations challenge diagnostic precision, this paper introduces a novel fusion of Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) models to refine detection capabilities. The paper capitalizes on an expansive radiography database, utilizing lung segmentation and data augmentation to bolster the training set. Advanced CNN models extract salient features, which are then classified through meticulously optimized machine learning algorithms. The synergy between diverse CNN architectures and Bayesian Optimization markedly enhances hyperparameter tuning, culminating in a significant uplift in diagnostic accuracy. The integration of LSTM models further allows for nuanced severity assessments, paving the way for more personalized and effective pandemic management. This research heralds a new frontier in medical diagnostics, promising a leap in the accuracy of pandemic detection.

CNN-Based Machine Tool Monitoring with STFT Image Analysis

Abstract: In the realm of machine tool operation, effective condition monitoring holds predominant importance to ensure operational reliability and safety. Leveraging deep learning methodologies, particularly convolutional neural networks (CNN), for defect identification has gained significant attention. However, inherent challenges persist, including the extraction of salient features and potential information loss while extracting the features from raw vibration signals. In response, this study proposes an intelligent approach for condition monitoring in machine tools, integrating short-time Fourier transform and convolutional neural networks (STFT-CNN). The process entails using STFT to convert one-dimensional vibration signals into time-frequency pictures, which are then fed into the STFT-CNN model to acquire and identify fault features. Furthermore, the study explores optimizing STFT parameters, such as window type, window width, and translation overlap width, across various typical window functions to improve the effectiveness of the transformation process. Within the STFT-CNN model, the utilization of stacked double convolutional layers aims to augment the model's nonlinear expression capacity, thereby facilitating robust fault diagnosis capabilities in machine tool condition monitoring applications

Which images to label for few-shot medical image analysis?

The success of deep learning methodologies hinges upon the availability of meticulously labeled extensive datasets. However, when dealing with medical images, the annotation process for such abundant training data often necessitates the involvement of experienced radiologists, thereby consuming their limited time resources. In order to alleviate this burden, few-shot learning approaches have been developed, which manage to achieve competitive performance levels with only several labeled images. Nevertheless, a crucial yet previously overlooked problem in few-shot learning is about the selection of template images for annotation before learning, which affects the final performance. In this study, we propose a novel TEmplate Choosing Policy (TECP) that aims to identify and select “the most worthy” images for annotation, particularly within the context of multiple few-shot medical tasks, including landmark detection, anatomy detection, and anatomy segmentation. TECP is composed of four integral components: (1) Self-supervised training, which entails training a pre-existing deep model to extract salient features from radiological images; (2) Alternative proposals for localizing informative regions within the images; and (3) Representative Score Estimation, which involves the evaluation and identification of the most representative samples or templates. (4) Ranking, which rank all candidates and select one with highest representative score. The efficacy of the TECP approach is demonstrated through a series of comprehensive experiments conducted on multiple public datasets. Across all three medical tasks, the utilization of TECP yields noticeable improvements in model performance.

Design of urban road fault detection system based on artificial neural network and deep learning.

In urban traffic management, the timely detection of road faults plays a crucial role in improving traffic efficiency and safety. However, conventional methods often fail to fully leverage the information from road topology and traffic data. To address this issue, we propose an innovative detection system that combines Artificial Neural Networks (ANNs), specifically Graph Convolutional Networks (GCN), Bidirectional Gated Recurrent Units (BiGRU), and self-attention mechanisms. Our approach begins by representing the road topology as a graph and utilizing GCN to model it. This allows us to learn the relationships between roads and capture their structural dependencies. By doing so, we can effectively incorporate the spatial information provided by the road network. Next, we employ BiGRU to model the historical traffic data, enabling us to capture the temporal dynamics and patterns in the traffic flow. The BiGRU architecture allows for bidirectional processing, which aids in understanding the traffic conditions based on both past and future information. This temporal modeling enhances our system's ability to handle time-varying traffic patterns. To further enhance the feature representations, we leverage self-attention mechanisms. By combining the hidden states of the BiGRU with self-attention, we can assign importance weights to different temporal features, focusing on the most relevant information. This attention mechanism helps to extract salient features from the traffic data. Subsequently, we merge the features learned by GCN from the road topology and BiGRU from the traffic data. This fusion of spatial and temporal information provides a comprehensive representation of the road status. By employing a Multilayer Perceptron (MLP) as a classifier, we can effectively determine whether a road is experiencing a fault. The MLP model is trained using labeled road fault data through supervised learning, optimizing its performance for fault detection. Experimental evaluations of our system demonstrate excellent performance in road fault detection. Compared to traditional methods, our system achieves more accurate fault detection, thereby improving the efficiency of urban traffic management. This is of significant importance for city administrators, as they can promptly identify road faults and take appropriate measures for repair and traffic diversion.

Deep learning cigarette defect detection method based on saliency feature guidance

AbstractCigarette defect detection is important in industrial production. Existing methods extract features for defect detection manually or using deep learning. However, due to the small size of cigarette defects, these methods are unable to effectively extract discriminative features, limiting detection performance. Hence, a deep learning‐based method called significant feature‐guided cigarette defect detection (SFGCD) is proposed, which combines saliency feature extraction methods with deep learning to enhance feature representation and improve detection. First, edge saliency features are extracted using the proposed target gradient saliency feature extraction (GSFE) strategy. Then, a dense multi‐level feature fusion network is designed to combine the original features with the saliency features obtained from the target gradient saliency feature extraction strategy. This network enriches feature representation and improves detection by fusing original and saliency features at different levels and scales. Experimental results demonstrate that the proposed method achieves a higher accuracy of 0.02 mean average precision (MAP) value and a detection speed of 5 frames per second (FPS) on the authors' own labeled cigarette defect dataset compared to existing state‐of‐the‐art methods.

Enhancing Out-of-distribution Generalization on Graphs via Causal Attention Learning

In graph classification, attention- and pooling-based graph neural networks (GNNs) predominate to extract salient features from the input graph and support the prediction. They mostly follow the paradigm of “learning to attend,” which maximizes the mutual information between the attended graph and the ground-truth label. However, this paradigm causes GNN classifiers to indiscriminately absorb all statistical correlations between input features and labels in the training data without distinguishing the causal and noncausal effects of features. Rather than emphasizing causal features, the attended graphs tend to rely on noncausal features as shortcuts to predictions. These shortcut features may easily change outside the training distribution, thereby leading to poor generalization for GNN classifiers. In this article, we take a causal view on GNN modeling. Under our causal assumption, the shortcut feature serves as a confounder between the causal feature and prediction. It misleads the classifier into learning spurious correlations that facilitate prediction in in-distribution (ID) test evaluation while causing significant performance drop in out-of-distribution (OOD) test data. To address this issue, we employ the backdoor adjustment from causal theory—combining each causal feature with various shortcut features, to identify causal patterns and mitigate the confounding effect. Specifically, we employ attention modules to estimate the causal and shortcut features of the input graph. Then, a memory bank collects the estimated shortcut features, enhancing the diversity of shortcut features for combination. Simultaneously, we apply the prototype strategy to improve the consistency of intra-class causal features. We term our method as CAL+, which can promote stable relationships between causal estimation and prediction, regardless of distribution changes. Extensive experiments on synthetic and real-world OOD benchmarks demonstrate our method’s effectiveness in improving OOD generalization. Our codes are released at https://github.com/shuyao-wang/CAL-plus .

A Unified Visual and Linguistic Semantics Method for Enhanced Image Captioning

Image captioning, also recognized as the challenge of transforming visual data into coherent natural language descriptions, has persisted as a complex problem. Traditional approaches often suffer from semantic gaps, wherein the generated textual descriptions lack depth, context, or the nuanced relationships contained within the images. In an effort to overcome these limitations, we introduce a novel encoder–decoder framework called A Unified Visual and Linguistic Semantics Method. Our method comprises three key components: an encoder, a mapping network, and a decoder. The encoder employs a fusion of CLIP (Contrastive Language–Image Pre-training) and SegmentCLIP to process and extract salient image features. SegmentCLIP builds upon CLIP’s foundational architecture by employing a clustering mechanism, thereby enhancing the semantic relationships between textual and visual elements in the image. The extracted features are then transformed by a mapping network into a fixed-length prefix. A GPT-2-based decoder subsequently generates a corresponding Chinese language description for the image. This framework aims to harmonize feature extraction and semantic enrichment, thereby producing more contextually accurate and comprehensive image descriptions. Our quantitative assessment reveals that our model exhibits notable enhancements across the intricate AIC-ICC, Flickr8k-CN, and COCO-CN datasets, evidenced by a 2% improvement in BLEU@4 and a 10% uplift in CIDEr scores. Additionally, it demonstrates acceptable efficiency in terms of simplicity, speed, and reduction in computational burden.

AGV monocular vision localization algorithm based on Gaussian saliency heuristic

To address the issues of poor detection accuracy and the large number of target detection model parameters in existing AGV monocular vision location detection algorithms, this paper presents an AGV vision location method based on Gaussian saliency heuristic. The proposed method introduces a fast and accurate AGV visual detection network called GAGV-net. In the GAGV-net network, a Gaussian saliency feature extraction module is designed to enhance the network’s feature extraction capability, thereby reducing the required output for model fitting. To improve the accuracy of target detection, a joint multi-scale classification and detection task header are designed at the stage of target frame regression to classification. This header utilizes target features of different scales, thereby enhancing the accuracy of target detection. Experimental results demonstrate a 12% improvement in detection accuracy and a 27.38 FPS increase in detection speed compared to existing detection methods. Moreover, the proposed detection network significantly reduces the model’s size, enhances the network model’s deployability on AGVs, and greatly improves detection accuracy.

Modeling vehicle U-turning behavior near intersections: A deep learning approach based on TCN and multi-head attention

In U-turn bays near intersections, the conflict between U-turning vehicles and those going straight-ahead results in traffic accidents since straight-ahead vehicles cannot reliably anticipate the behavior of oncoming U-turning vehicles. However, previous studies on modeling U-turning behavior do not effectively capture the spatial–temporal interaction between the U-turning and surrounding vehicles. To address this issue, a deep-learning framework based on a temporal convolutional network (TCN) and multi-head attention mechanism is developed. The TCN is utilized to capture long-term dependencies of vehicles in the shared left- and U-turn lanes by extracting vehicle historical motion features. The self-attention mechanism extracts salient features related to the U-turn intentions, classifying the vehicles into left- and U-turning vehicles based on their driving intentions. A parallel TCN and spatial multi-head attention structure is constructed to model vehicle–vehicle interactions to further predict the future trajectory of U-turning vehicles. Finally, the obtained features are input into a Transformer-based decoder module and trajectory generator to predict the future displacement and body orientation of U-turning vehicles. The model is validated via comparison with state-of-the-art models and the observed trajectories under various scenarios. Ablation studies are conducted to quantify the efficacy of each module. Further, the effect of the surrounding homogenous and heterogeneous vehicles on U-turning vehicles in four different U-turn scenarios is quantified using spatial–temporal variation graphs and attention matrices.

Distributed feature representations of natural stimuli across parallel retinal pathways

How sensory systems extract salient features from natural environments and organize them across neural pathways is unclear. Combining single-cell and population two-photon calcium imaging in mice, we discover that retinal ON bipolar cells (second-order neurons of the visual system) are divided into two blocks of four types. The two blocks distribute temporal and spatial information encoding, respectively. ON bipolar cell axons co-stratify within each block, but separate laminarly between them (upper block: diverse temporal, uniform spatial tuning; lower block: diverse spatial, uniform temporal tuning). ON bipolar cells extract temporal and spatial features similarly from artificial and naturalistic stimuli. In addition, they differ in sensitivity to coherent motion in naturalistic movies. Motion information is distributed across ON bipolar cells in the upper and the lower blocks, multiplexed with temporal and spatial contrast, independent features of natural scenes. Comparing the responses of different boutons within the same arbor, we find that axons of all ON bipolar cell types function as computational units. Thus, our results provide insights into the visual feature extraction from naturalistic stimuli and reveal how structural and functional organization cooperate to generate parallel ON pathways for temporal and spatial information in the mammalian retina.

Advancing Legal Citation Text Classification A Conv1D-Based Approach for Multi-Class Classification

The escalating volume and intricacy of legal documents necessitate advanced techniques for automated text classification in the legal domain. Our proposed approach leverages Convolutional Neural Networks (Conv1D), a neural network architecture adept at capturing hierarchical features in sequential data. The incorporation of max-pooling facilitates the extraction of salient features, while softmax activation enables the model to handle the multi-class nature of legal citation categorization. By addressing the limitations identified in previous studies, our model aims to advance the state-of-the-art in legal citation text classification, offering a robust and efficient solution for automated categorization in the legal domain. Our research contributes to the ongoing evolution of NLP applications in the legal field, promising enhanced accuracy and adaptability in the automated analysis of legal texts.

A meta-learning method for few-shot bearing fault diagnosis under variable working conditions

Intelligent fault diagnosis in various industrial applications has rapidly evolved due to the recent advancements in data-driven techniques. However, the scarcity of fault data and a wide range of working conditions pose significant challenges for existing diagnostic algorithms. This study introduces a meta-learning method tailored for the classification of motor rolling bearing faults, addressing the challenges of limited data and diverse conditions. In this approach, a deep residual shrinkage network is employed to extract salient features from bearing vibration signals. These features are then analyzed in terms of their proximity to established fault prototypes, enabling precise fault categorization. Moreover, the model’s generalization in few-shot scenarios is enhanced through the incorporation of a meta-learning paradigm during training. The approach is evaluated using two well-known public bearing datasets, focusing on varying speeds, loads, and high noise environments. The experimental results indicate the superior diagnostic accuracy and robustness of our method compared with those of existing studies.

Identifying Malware Packers through Multilayer Feature Engineering in Static Analysis

This research addresses a critical need in the ongoing battle against malware, particularly in the form of obfuscated malware, which presents a formidable challenge in the realm of cybersecurity. Developing effective antivirus (AV) solutions capable of combating packed malware remains a crucial endeavor. Packed malicious programs employ encryption and advanced techniques to obfuscate their payloads, rendering them elusive to AV scanners and security analysts. The introduced research presents an innovative malware packer classifier specifically designed to adeptly identify packer families and detect unknown packers in real-world scenarios. To fortify packer identification performance, we have curated a meticulously crafted dataset comprising precisely packed samples, enabling comprehensive training and validation. Our approach employs a sophisticated feature engineering methodology, encompassing multiple layers of analysis to extract salient features used as input to the classifier. The proposed packer identifier demonstrates remarkable accuracy in distinguishing between known and unknown packers, while also ensuring operational efficiency. The results reveal an impressive accuracy rate of 99.60% in identifying known packers and 91% accuracy in detecting unknown packers. This novel research not only significantly advances the field of malware detection but also equips both cybersecurity practitioners and AV engines with a robust tool to effectively counter the persistent threat of packed malware.

Hand Movement Recognition and Salient Tremor Feature Extraction With Wearable Devices in Parkinson’s Patients

Tremor is one of the earliest signs of Parkinson’s disease (PD), which seriously disrupts patients’ daily lives. It is important to study upper limb tremors quantitatively to control PD progression. In this study, surface electromyography (sEMG) signals from wearable devices are used to recognize rest, posture, and kinetic tremor action from 6 upper-limb clinical actions and to quantify features of tremors. A multivariable time series classification model (MTSCM) based on fully convolutional networks and a long short-term memory network is proposed to recognize tremor actions. MTSCM achieves a high degree of accuracy both on the left-hand and right-hand data sets for tremor actions. An improved Hilbert-Huang transform (HHT) method is proposed to decompose the inertial signals of tremor actions to obtain tremor components. Using the improved HHT, tremor and motion components can be decomposed effectively. In addition, 53 features are extracted from inertial and sEMG signals, and a canonical correlation analysis is used to determine the correlation between features and MDS-UPDRS scores. Several of the relevant characteristics are related to MDS-UPDRS scores, notably the dominant frequency and amplitude of the tremor component are significantly correlated (p < 0.01) with tremor scores. Detecting upper-limb clinical tremors in PD patients using wearable sensors is feasible according to our findings.

Multilingual Speaker Recognition using Mel-frequency Cepstral Coefficients and Gaussian Mixture Model

Introduction: People can recognize a speaker with the help of their voice via mobile or digital devices. Method: To obtain this congenital human being ability, authentication techniques based on speaker biometrics like automated speaker recognition (ASR) have been proposed. An ASR identifies speakers by speech signals analysis and salient feature extraction from their voices. Result: This will become an important part of recent research in the voice biometrics field. This paper proposes multilingual speaker recognition with the help of MFCC as feature extraction and GMM as classification techniques using various available datasets such as TIMIT, librespeech, etc. Conclusion: The results achieved from the given datasets enhance the recognition rate of 70.98% with MFCC.