Feature Extraction Research Articles

Research on citrus leaf disease detection based on improved SEDS-YOLOv8 model with efficient multi-scale attention and multi-level channel compression

Citrus is one of the important economic crops, with a vast planting area and complex terrain and environmental conditions. The growth cycle of citrus is long, and it is prone to a wide variety of diseases and pests. If the types of diseases and pests cannot be accurately identified in a timely manner to take corresponding control measures, it will seriously affect the yield and quality of citrus. This study aims to improve the detection accuracy of leaf diseases and pests, reduce the computational scale of the model, and enhance its deployability. A lightweight disease and pest detection model based on the improved YOLOv8 is proposed, and a disease and pest dataset considering different environmental conditions is established. Firstly, the convolutional module (Conv) in the neck network of YOLOv8 is replaced by GSConv, and the C2f module is replaced by VoV-GSCSP, forming a Slim-neck architecture, which reduces the computational complexity of the model while maintaining high recognition accuracy. At the same time, the C2f module in the backbone network is replaced by the C2f_EMA module that integrates the EMA efficient multi-scale attention mechanism, enhancing the model's feature extraction ability for leaf diseases and pests in complex environments. Additionally, the original detection head is improved through multi-level channel compression to reduce features along the channel dimension. The SEDS-YOLOv8 model is designed through the above methods. Experimental results show that the model's parameters, computational cost, and memory usage are reduced by 63.5%, 72.83%, and 61.9% respectively. The model's precision, recall, and mean average precision are 97.5%, 96.2%, and 98.5% respectively. In terms of performance, the detection frame rate on mobile devices reaches 358.5 frames per second, and the average inference time for a single leaf disease and pest image is 4.4 ms. This proves that the algorithm can significantly reduce the computational load of the network while maintaining high detection performance, meeting the deployment requirements of mobile and embedded devices. **************** ACKNOWLEDGEMENTS**************** Thanks for the data support provided by National-level Innovation Program Project Fund "Research on Seedling Inspection Robot Technology Based on Multi-source Information Fusion and Deep Network" (No.: 202410451009); Jiangsu Provincial Natural Science Research General Project (No.: 20KJB530008); China Society for Smart Engineering "Research on Intelligent Internet of Things Devices and Control Program Algorithms Based on Multi-source Data Analysis" (No.: ZHGC104432); China Engineering Management Association "Comprehensive Application Research on Intelligent Robots and Intelligent Equipment Based on Big Data and Deep Learning" (No.: GMZY2174); Key Project of National Science and Information Technology Department Research Center National Science and Technology Development Research Plan (No.: KXJS71057); Key Project of National Science and Technology Support Program of Ministry of Agriculture (No.: NYF251050).

Adaptive genetic algorithm based deep feature selector for cancer detection in lung histopathological images

Cancer is a global health concern because of a significant mortality rate and a wide range of affected organs. Early detection and accurate classification of cancer types are crucial for effective treatment. Imaging tests on different image modalities such as Histopathology images, provide valuable insights into the cellular and architectural features of tissues, allowing pathologists to make diagnosis, determine disease stages, and guide treatment decisions. They are an essential tool in the study and understanding of diseases, aiding in research, education, and patient care. Convolutional neural network based pretrained deep learning models can be used successfully to detect lung cancer. In this study, we have used a channel attention-enabled deep learning model as a feature extractor followed by an adaptive Genetic Algorithm (GA) based feature selector. Here, we calculate the fitness score of each chromosome (i.e., a candidate solution) using a filter method, instead of a classifier. Further, the GA optimized feature vector is fed to the K-nearest neighbors classifier for final classification. The proposed method shows a promising result with an overall accuracy of 99.75% on the LC25000 dataset, which is a publicly available dataset of lung histopathological images. The source code for this work can be found https://github.com/priyam-03/GA-Feature-Selector-Lung-Cancer.

Deep learning for time series forecasting: a survey

Abstract Time series forecasting (TSF) has long been a crucial task in both industry and daily life. Most classical statistical models may have certain limitations when applied to practical scenarios in fields such as energy, healthcare, traffic, meteorology, and economics, especially when high accuracy is required. With the continuous development of deep learning, numerous new models have emerged in the field of time series forecasting in recent years. However, existing surveys have not provided a unified summary of the wide range of model architectures in this field, nor have they given detailed summaries of works in feature extraction and datasets. To address this gap, in this review, we comprehensively study the previous works and summarize the general paradigms of Deep Time Series Forecasting (DTSF) in terms of model architectures. Besides, we take an innovative approach by focusing on the composition of time series and systematically explain important feature extraction methods. Additionally, we provide an overall compilation of datasets from various domains in existing works. Finally, we systematically emphasize the significant challenges faced and future research directions in this field.

Intelligent Monitoring and Safety Management System Construction of Manhole Cover Hidden Danger Based on YoSwin-Former Model

This paper focuses on the safety management of urban manhole covers. In view of the drawbacks of the traditional management model and the frequent occurrence of manhole cover accidents, an intelligent manhole cover hidden danger identification project is proposed. The project is dedicated to solving the problems of high cost, difficult hidden danger detection, and low fault tolerance in the manhole cover hidden danger detection industry. A data set of 13,420 manhole covers in various states is collected through network search, field investigation, etc., and after data cleaning, standardization, annotation, and partitioning, the YoSwin - Former model, which combines the advantages of YOLOv5 and Swin Transformer, is used for training. This model performs excellently in feature extraction and dealing with complex scenes, effectively improving the accuracy and efficiency of manhole cover hidden danger detection, and has far-reaching significance for ensuring the safety of urban residents' lives and property and promoting the intelligence of urban management.

A study on rolling bearing fault diagnosis using RIME-VMD

To address the challenges of feature extraction in Variational Mode Decomposition (VMD) for rolling bearing fault diagnosis, this paper proposes a feature extraction method optimized by the RIME algorithm, called RIME-VMD. First, under various rolling bearing fault conditions, the RIME algorithm is employed to determine the optimal combination of decomposition components and penalty factors in VMD. Next, the kurtosis values of each decomposed Intrinsic Mode Function (IMF) are calculated, and the component with the most prominent fault features is selected for noise reduction through reconstruction. Finally, the sample entropy of the reconstructed signal is calculated as a fault feature and input into a Support Vector Machine (SVM) for rapid identification and diagnosis of various rolling bearing fault types. Simulation results indicate that, compared to the Whale Optimization Algorithm optimized VMD (WOA-VMD), the RIME algorithm optimized VMD (RIME-VMD) achieves shorter search times and higher search efficiency. It facilitates faster identification of decomposition parameters under various fault conditions, enhancing the robustness of fault signal detection and enabling rapid, efficient identification of rolling bearing faults. The findings of this study offer guidance and reference for future research on rolling bearing fault diagnosis.

Polarity-JaM: an image analysis toolbox for cell polarity, junction and morphology quantification

Cell polarity involves the asymmetric distribution of cellular components such as signalling molecules and organelles within a cell, alterations in cell morphology and cell-cell contacts. Advances in fluorescence microscopy and deep learning algorithms open up a wealth of unprecedented opportunities to characterise various aspects of cell polarity, but also create new challenges for comprehensible and interpretable image data analysis workflows to fully exploit these new opportunities. Here we present Polarity-JaM, an open source package for reproducible exploratory image analysis that provides versatile methods for single cell segmentation, feature extraction and statistical analysis. We demonstrate our analysis using fluorescence image data of endothelial cells and their collective behaviour, which has been shown to be essential for vascular development and disease. The general architecture of the software allows its application to other cell types and imaging modalities, as well as seamless integration into common image analysis workflows, see https://polarityjam.readthedocs.io. We also provide a web application for circular statistics and data visualisation, available at www.polarityjam.com, and a Napari plug-in, each with a graphical user interface to facilitate exploratory analysis. We propose a holistic image analysis workflow that is accessible to the end user in bench science, enabling comprehensive analysis of image data.

Fusion of multi-scale attention for aerial images small-target detection model based on PARE-YOLO

In view of the complex environments and varying object scales in drone-captured imagery, a novel PARE-YOLO algorithm based on YOLOv8 for small object detection is proposed. This model enhances feature extraction and fusion across multiple scales through a restructured neck network. Additionally, it incorporates a lightweight detection head that is optimized for small objects, thereby significantly improving detection performance in cluttered and intricate backgrounds. To further enhance the extraction of small object features, the conventional C2f is replaced with a novel architecture. Moreover, the EMA-GIoU loss function is proposed to mitigate class imbalance and enhance robustness, particularly in scenarios characterized by skewed class distributions. Evaluation on the VisDrone2019 dataset indicates that PARE-YOLO achieves a 5.9% improvement in mean Average Precision (mAP) at a threshold of 0.5, compared to the original YOLOv8 model. In addition, the PARE-YOLO model exhibits significant robustness, achieving a mean Average Precision (mAP) at a threshold of 0.5 values on the HIT-UAV dataset that are 0.8%, 0.5%, and 1.2% higher than those of YOLOv8, YOLOv10, and RT-DETR. These results underscore the effectiveness of PARE-YOLO in addressing the challenges inherent in aerial scenarios. The code will be available online (https://github.com/Sunnyxiao69/PARE-YOLO).

Deep Learning Based Image Aesthetic Quality Assessment- A Review

Image Aesthetic Quality Assessment (IAQA) spans applications such as the fashion industry, AI-generated content, product design, and e-commerce. Recent deep learning advancements have been employed to evaluate image aesthetic quality. A few surveys have been conducted on IAQA models; however, details of recent deep learning models and challenges have not been fully mentioned. This paper aims to fill these gaps by providing a review of deep learning IAQA over the past decade, based on input, process, and output phases. Methodologies for deep learning-based IAQA can be categorized into general and task-specific approaches, depending on the type and diversity of input images. The processing phase involves considerations related to network architecture, learning structures, and feature extraction methods. The output phase generates results such as scoring, distribution, attributes, and description. Despite achieving a maximum accuracy of 91.5%, further improvements in deep learning models are still required. Our study highlights several challenges, including adapting models for task-specific methodology, accounting for environmental factors influencing aesthetics, the lack of substantial datasets with appropriate labels, imbalanced data, preserving image aspect ratio and integrity in network architecture design, and the need for explainable AI to understand the causative factors behind aesthetic judgments.

Deep learning based gasket fault detection: a CNN approach

Gasket inspection is a critical step in the quality control of a product. The proposed method automates the detection of misaligned or incorrectly fitting gaskets, ensuring timely repair action. The suggested method uses deep learning approaches to recognize and evaluate radiator images, with a focus on identifying misaligned or incorrectly installed gaskets. Deep learning algorithms are specific for feature extraction and classification together with a convolutional neural network (CNN) module that allows for seamless connection. A gasket inspection system based on a CNN architecture is developed in this work. The system consists of two sets of convolution layers, followed by two sets of batch normalization layer, two sets of RELU layer, max pooling layer and finally fully connected layer for classification of gasket images. The obtained results indicate that our system has great potential for practical applications in the manufacturing industry. Moreover, our system provides a reliable and efficient mechanism for quality control, which can help reduce the risk of defects and ensure product reliability.

Adaptive Clustering Method for Panel DataBased on Multi-dimensional Feature Extraction

Adaptive Clustering Method for Panel DataBased on Multi-dimensional Feature Extraction

Novel machine learning-driven comparative analysis of CSP, STFT, and CSP-STFT fusion for EEG data classification across multiple meditation and non-meditation sessions in BCI pipeline

This study focuses on classifying multiple sessions of loving kindness meditation (LKM) and non-meditation electroencephalography (EEG) data. This novel study focuses on using multiple sessions of EEG data from a single individual to train a machine learning pipeline, and then using a new session data from the same individual for the classification. Here, two meditation techniques, LKM-Self and LKM-Others were compared with non-meditation EEG data for 12 participants. Among many tested, three BCI pipelines we built produced promising results, successfully detecting features in meditation/ non-meditation EEG data. While testing different feature extraction algorithms, a common neural network structure was used as the classification algorithm to compare the performance of the feature extraction algorithms. For two of those pipelines, Common Spatial Patterns (CSP) and Short Time Fourier Transform (STFT) were successfully used as feature extraction algorithms where both these algorithms are significantly new for meditation EEG. As a novel concept, the third BCI pipeline used a feature extraction algorithm that fused the features of CSP and STFT, achieving the highest classification accuracies among all tested pipelines. Analyses were conducted using EEG data of 3, 4 or 5 sessions, totaling 3960 tests on the entire dataset. At the end of the study, when considering all the tests, the overall classification accuracy using SCP alone was 67.1%, and it was 67.8% for STFT alone. The algorithm combining the features of CSP and STFT achieved an overall classification accuracy of 72.9% which is more than 5% higher than the other two pipelines. At the same time, the highest mean classification accuracy for the 12 participants was achieved using the pipeline with the combination of CSP STFT algorithm, reaching 75.5% for LKM-Self/ non-meditation for the case of 5 sessions of data. Additionally, the highest individual classification accuracy of 88.9% was obtained by the participant no. 14. Furthermore, the results showed that the classification accuracies for all three pipelines increased with the number of training sessions increased from 2 to 3 and then to 4. The study was successful in classifying a new session of EEG meditation/ non-meditation data after training machine learning algorithms using a different set of session data, and this achievement will be beneficial in the development of algorithms that support meditation.

Modified CNN Model for Network Intrusion Detection and Classification System Using Local Outlier Factor-based Recursive Feature Elimination

An intrusion detection system (IDS) is either a part of a software or hardware environment that monitors data and analyses it to identify any attacks made against a system or a network. Traditional IDS approaches make the system more complicated and less efficient, because the analytical properties process is difficult and time-consuming. This is because the procedure is complex. Therefore, this research work focuses on a network intrusion detection and classification (NIDCS) system using a modified convolutional neural network (MCNN) with recursive feature elimination (RFE). Initially, the dataset is balanced with the help of the local outlier factor (LOF), which finds anomalies and outliers by comparing the amount of deviation that a single data point has with the amount of deviation that its neighbors have. Then, a feature extraction selection approach named RFE is applied to eliminate the weakest features until the desired number of features is achieved. Finally, the optimal features are trained with the MCNN classifier, which classifies intrusions like probe, denial-of-service (DoS), remote-to-user (R2U), user-to-root (U2R), and identifies normal data. The proposed NIDCS system resulted in higher performance with 99.3% accuracy and a 3.02 false alarm rate (FAR) as equated to state-of-the-art NIDCS approaches such as deep neural networks (DNN), ResNet, and gravitational search algorithms (GSA).

Metaheuristic-enhanced Deep Learning Model for Accurate Alzheimer's Disease Diagnosis from MRI Imaging

Alzheimer’s Disease (AD) is the neuro-degenerative dementia, where the precise and early recognition of AD is vital for timely treatment to reduce mortality rate. A new automated model is implemented in this work for early discovery of AD in the Magnetic Resonance Imaging (MRI) brain scans. Initially, the input brain scans are taken from the Alzheimer's disease Neuroimaging Initiative (ADNI) database. Further, the acquired raw brain scans are visually improved by employing the binary normalization technique. The denoised brain scans are fed to the pre-trained Convolutional Neural Network (CNN) named GoogleNet for feature extraction. Next, the extracted richer feature values are fed to the Long Short Term Memory (LSTM) network for classifying the brain scan as Normal Control (NC), Mild Cognitive Impairment (MCI) and AD. In this manuscript, a Honey Badger Optimization Algorithm (HBOA) technique is incorporated with the LSTM networks for hyper-parameters optimization, where this procedure helps in diminishing the LSTM network’s complexity and computational time. The experimental results conducted on the ADNI database underscore the HBOA-based LSTM network's effectiveness, showcasing a remarkable mean classification accuracy of 97.83% in multi-class classification. Moreover, the sensitivity of HBOA based LSTM for AD/NC is 96.73% which is high when compared to the existing methodologies such as SVM with radial basis kernel function and NCSINs. This performance surpasses that of other comparative models for AD detection, emphasizing the superior capabilities and potential of the proposed method in the early detection.

A comprehensive analysis of deep learning and transfer learning techniques for skin cancer classification

Accurately and early diagnosis of melanoma is one of the challenging tasks due to its unique characteristics and different shapes of skin lesions. So, in order to solve this issue, the current study examines various deep learning-based approaches and provide an effective approach for classifying dermoscopic images into two categories of skin lesions. This research focus on skin cancer images and provides solution using deep learning approaches. This research investigates three approaches for classifying skin cancer images. (1) Utilizing three fine-tuned pre-trained networks (VGG19, ResNet18, and MobileNet_V2) as classifiers. (2) Employing three pre-trained networks (ResNet-18, VGG19, and MobileNet v2) as feature extractors in conjunction with four machine learning classifiers (SVM, DT, Naïve Bayes, and KNN). (3) Utilizing a combination of the aforementioned pre-trained networks as feature extractors in conjunction with same machine learning classifiers. All these algorithms are trained using segmented images which are achieved by using the active contour approach. Prior to segmentation, preprocessing step is performed which involves scaling, denoising, and enhancing the image. Experimental performance is measured on the ISIC 2018 dataset which contains 3300 images of skin disease including benign and malignant type cancer images. 80% of the images from the ISIC 2018 dataset are allocated for training, while the remaining 20% are designated for testing. All approaches are trained using different parameters like epoch, batch size, and learning rate. The results indicate that combining ResNet-18 and MobileNet pre-trained networks using concatenation with an SVM classifier achieved the maximum accuracy of 92.87%.

Bilingual hate speech detection on social media: Amharic and Afaan Oromo

Due to significant increases in internet penetration and the development of smartphone technology during the preceding couple of decades, many people have started using social media as a communication platform. Social media has grown to be one of the most significant components, with several benefits. However, technology also poses a number of threats, challenges, and barriers, such as hate speech, disinformation, and fake news. Hate speech detection is one of the many ways social media platforms can be accused of not doing enough to thwart hate speech on their platform. People in Bilingual and multinational societies commonly employ a code mix in both spoken and written communication. Among these, Amharic and Afaan Oromo language speakers frequently mix the two languages when conversing and posting on social media. The majority of previous study concentrated on identifying either technological favoured language or monolingual hate speech in Ethiopian languages; however, the availability of Bilingual communication in social media hampers the propagation of hate speech via social media. In this work, a Bilingual hate speech detection for Amharic and Afaan Oromo languages were conducted using four different deep learning classifiers (CNN, BiLSTM, CNN-BiLSTM, and BiGRU) and three feature extraction (Keras word embedding, word2vec, and FastText) techniques. According to the experiment, BiLSTM with FastText feature extraction is an outperforming the other algorithm by achieving a 78.05% accuracy for Bilingual Amharic Afaan Oromo hate speech detection. The FastText feature extraction overcomes the problem of out of vocabulary (OOV). Furthermore, we are working towards including others linguistic features of the languages to detect hate speech and make the resource available to facilitate further research in the area of Bilingual hate speech detection for other under-resourced Ethiopian languages.

Cross-contextual stress prediction: Simple methodology for comparing features and sample domain adaptation techniques in vital sign analysis

Stress significantly impacts individuals, particularly in professions like nursing and driving, leading to severe health risks and accidents. Accurate stress measurement is critical for effective interventions, yet research is hindered by incomplete datasets and inconsistent methodologies, slowing the development of reliable predictive models. This paper introduces a framework for cross-contextual stress prediction, enabling the generation of general stress prediction models adaptable to specific domain challenges. The methodology leverages two general daily life datasets and three domain-specific datasets, employing steps such as dataset selection, feature extraction, significant feature identification, feature preprocessing, fine-tuning, domain adaptation, and application to specific contexts. Through this framework, key vital signs were identified as significant predictors of stress, including electrocardiography (ECG), heart rate (HR), heart rate variability (HRV) - low frequency (LF), electrodermal activity (EDA), body temperature (TEMP), and skin conductance response (SCR). The experiments conducted include: 1) Utilizing HR and HRV-LF through domain adaptation from general to automobile driving datasets; 2) Applying EDA, HR, and TEMP from general to specific nurse activity datasets; and 3) Adapting ECG, HR, and TEMP from general to automobile driving datasets. Results demonstrate the potential of the proposed framework for cross-contextual stress prediction, with HR and HRV-LF identified as pivotal features. When applied to target datasets specific to stress scenarios, the model achieved a 62% F1 score, demonstrating the effectiveness of the feature-based Correlation Alignment (CORAL) technique combined with Random Forest models in transferring learned knowledge across domains. These findings highlight the robustness of the approach in adapting general stress prediction models to specific contexts, paving the way for real-world applications such as stress monitoring in driving and nursing during high-stress periods like COVID-19.

Winter wheat harvest detection via Sentinel-2 MSI images

ABSTRACT Wheat is one of the most important staple crops globally. Timely mapping and monitoring of wheat harvests are essential for efficiently scheduling large-scale harvesters, ensuring the timely completion of the harvest, and maintaining grain quality. Traditional manual survey methods for obtaining wheat harvest information are neither highly accurate nor cost-effective and do not meet the needs of agricultural management departments. This study introduces two novel indices for wheat harvest detection: the optical-band brightness harvest index (OBHI) and the visible-band brightness harvest index (VBHI). The research is structured into three primary components: (1) Extraction of wheat planting areas, utilizing phenological features from multiple growth stages; (2) Extraction of wheat harvesting features, where the proposed OBHI and VBHI are analysed using the box plot method to identify the harvesting characteristics of wheat croplands; and (3) Wheat harvest detection, employing the OBHI, VBHI, and a threshold method to determine the harvest status. The key findings are as follows: (1) Combining the OBHI with a threshold method achieves the highest accuracy in detecting wheat harvest using Sentinel-2 MSI images; (2) Integrating Sentinel-2 multispectral remote sensing imagery with the OBHI threshold method enables real-time monitoring of wheat harvest progress. In the study area, the wheat harvest commenced on 1 June 2023 (0.62%) and was nearly complete by 13 June 2023 (97.94%). The OBHI and VBHI proposed in this study have the potential to assist agricultural management departments in improving the efficiency of wheat harvest supervision. However, further research and validation are necessary to determine the generalizability and applicability of this method.

Reach&Grasp: a multimodal dataset of the whole upper-limb during simple and complex movements

Upper-limb movement characterization is crucial for many applications, from research on motor control, to the extraction of relevant features for driving active prostheses. While this is usually performed using electrophysiological and/or kinematic measurements only, the collection of tactile data during grasping movements could enrich the overall information about interaction with external environment. We provide a dataset collected from 10 healthy volunteers performing 16 tasks, including simple movements (i.e., hand opening/closing, wrist pronation/supination and flexion/extension, tridigital grasping, thumb abduction, cylindrical and spherical grasping) and more complex ones (i.e., reaching and grasping). The novelty consists in the inclusion of several types of recordings, namely electromyographic -both with bipolar and high-density configuration, kinematic-both with motion capture system and a sensorized glove, and tactile. The data is organized following the Brain Imaging Data Structure standard format and have been validated to ensure its reliability. It can be used to investigate upper-limb movements in physiological conditions, and to test sensor fusion approaches and control algorithms for prosthetics and robotic applications.

Bridging branches and attributes: spectral-spatial global-local interaction network for hyperspectral image classification

ABSTRACT The CNN-Transformer joint model stands as the leading architecture for contemporary hyperspectral image (HSI) classification, integrating global and local features through either successive or dual-branch CNN and Transformer networks. However, these methods often fall short in effectively incorporating spatial-spectral information with local-global attributes, resulting in incomplete feature representation. To address these challenges, we propose a spectral-spatial global-local interaction network that transmits global and local features into the spatial and spectral branches, facilitated by cross-interaction operators to ensure adequate feature flow. Initially, CNNs are employed to separately extract shallow features for the spectral and spatial branches. We then introduce a Spectral-Spatial Global-Local Interaction block designed for deep feature extraction, enhancing the flow of spectral and spatial features with local and global attributes using parallel transformers and dynamic convolutions. Transformers model the long-range dependencies of global spectral and spatial features, while dynamic convolutions enhance the context sensitivity of local spectral and spatial representations. Quadruple cross-interaction blocks are proposed to traverse both the spectral-spatial branches and local-global attribute dimensions, facilitating information exchange for complementary HSI representation. Extensive experiments and ablation studies on four public HSI datasets demonstrate the superiority of our proposed method. convolutional neural networks (CNNs); spectral and spatial; global and local; hyperspectral image classification; cross-interaction.

An integrated framework for aspect category-based sentiment analysis using adaptive feature selection and category-aware decision fusion strategies

Purpose Aspect category-based sentiment analysis (ACSA) has been widely used in consumer preference mining and marketing strategy formulation. However, existing studies ignore the variability in features and the intrinsic correlation among diverse aspect categories in ACSA tasks. To address these problems, this paper aims to propose a novel integrated framework. Design/methodology/approach The integrated framework consists of three modules: text feature extraction and fusion, adaptive feature selection and category-aware decision fusion. First, text features from global and local views are extracted and fused to comprehensively capture the potential information in the different dimensions of the review text. Then, an adaptive feature selection strategy is devised for each aspect category to determine the optimal feature set. Finally, considering the intrinsic associations between aspect categories, a category-aware decision fusion strategy is constructed to enhance the performance of ACSA tasks. Findings Comparative experimental results demonstrate that the integrated framework can effectively detect aspect categories and their corresponding sentiment polarities from review texts, achieving a macroaveraged F1 score (Fmacro) of 72.38% and a weighted F1 score (F1) of 79.39%, with absolute gains of 2.93% to 27.36% and 4.35% to 20.36%, respectively, compared to the baselines. Originality/value This framework can simultaneously detect aspect categories and corresponding sentiment polarities from review texts, thereby assisting e-commerce enterprises in gaining insights into consumer preferences, prioritizing product improvements, and adjusting marketing strategies.