Main Feature Extraction Research Articles

CC-De-YOLO: A Multiscale Object Detection Method for Wafer Surface Defect

Abstract Surface defect detection on wafers is crucial for quality control in semiconductor manufacturing. However, the complexity of defect spatial features, including mixed defect types, large scale differences, and overlapping, results in low detection accuracy. In this paper, we propose a CC-De-YOLO model, which is based on the YOLOv7 backbone network. Firstly, the coordinate attention is inserted into the main feature extraction network. Coordinate attention decomposes channel attention into two one-dimensional feature coding processes, which are aggregated along both horizontal and vertical spatial directions to enhance the network’s sensitivity to orientation and position. Then, the nearest neighbor interpolation in the upsampling part is replaced by the CAR-EVC module, which predicts the upsampling kernel from the previous feature map and integrates semantic information into the feature map. Two residual structures are used to capture long-range semantic dependencies and improve feature representation capability. Finally, an efficient decoupled detection head is used to separate classification and regression tasks for better defect classification. To evaluate our model’s performance, we established a wafer surface defect dataset containing six typical defect categories. The experimental results show that the CCDe-YOLO model achieves 91.0% mAP@0.5 and 46.2% mAP@0.5:0.95, with precision of 89.5% and recall of 83.2%. Compared with the original YOLOv7 model and other object detection models, CC-De-YOLO performs better. Therefore, our proposed method meets the accuracy requirements for wafer surface defect detection and has broad application prospects. The dataset containing surface defect data on wafers is currently publicly available on GitHub (https://github.com/ztao3243/Wafer-Datas.git).

The impact of feature extraction techniques on the performance of text data classification models

<p class="MsoNormal" style="text-align: justify;">Sentiment analysis is a crucial discipline that focuses on the interpretation of feelings and points of view in textual data. Our study aims to assess the impact of different feature extraction methods on the accuracy of opinion research models. Techniques such as bag-of-words (BoW), term frequency-inverse document frequency (TF-IDF), Word2Vec, global vectors (GloVe) and bidirectional encoder representations from transformers (BERT) were used with three machine learning algorithms and three deep learning networks as classifiers. The IMDB movie review dataset was used for evaluation. The results showed that combining BERT with LSTM, CNN and RNN improved performance, achieving an accuracy rate of 94%, precision of 94.14%, recall of 93.27% and an F1 score of 89.33%. These results highlight the significant contribution of ERTB to model performance, outperforming other feature extraction techniques in text classification. The study concludes that the fusion of BERT and LSTM significantly improves model accuracy for opinion retrieval, recommending BERT as the main feature extraction method for optimizing performance in NLP tasks.</p>

An algorithm for power transmission line fault detection based on improved YOLOv4 model

In response to the escalating demand for real-time and accurate fault detection in power transmission lines, this paper undertook an optimization of the existing YOLOv4 network. This involved the substitution of the main feature extraction network within the original YOLOv4 model with a lighter EfficientNet network. Additionally, the inclusion of Grouped Convolution modules in the feature pyramid structure replaced conventional convolution operations. The resulting model not only reduced model parameters but also effectively ensured detection accuracy. Moreover, in enhancing the model's reliability, data augmentation techniques were employed to bolster the robustness of the power transmission line fault detection algorithm. This optimization further utilized the DIoU loss function to stabilize target box regression. Comparative experiments demonstrated the improved YOLOv4 model's superior performance in terms of loss function optimization while significantly enhancing detection speed under equivalent configurations. The parameter capacity was reduced by 81%, totaling merely 43.65 million, while the frame rate surged by 85% to achieve 24 frames per second. These experimental findings validate the effectiveness of the algorithm.

Heart disease prediction: Improved quantum convolutional neural network and enhanced features

Currently, heart disease is the leading cause of death in world. Since cardiac sickness requires knowledge and detailed information, it is challenging to anticipate. Healthcare systems have been using Internet of Things (IoT) technologies for gathering sensor data for diagnosing the heart disease and prognosis in recent years. Researchers have focused a lot of emphasis on diagnosing heart disease, although the outcomes are not always reliable. This article proposes the automated heart disease prediction model with three main stages, including preprocessing, feature extraction, and prediction. The input data undergoes an improved Z-score normalization as the preprocessing step. The appropriate features needed to train the prediction model are retrieved from the preprocessed data during feature extraction. The features extracted include improved entropy, statistical features, and information gain features. Depends on the features extracted, prediction is determined by the Improved Quantum CNN (IQCNN). The results of the IQCNN are compared to earlier systems for a various metrics. The proposed IQCNN model has achieved better accuracy of 0.91 at a 70% learning rate when evaluated over conventional methods like Bi-LSTM, CNN, QCNN, DNN, NN, RNN, MDCNN and E-KNN for better performance in the prediction of heart disease.

Deep ensemble model-based moving object detection and classification using SAR images

In recent decades, image processing and computer vision models have played a vital role in moving object detection on the synthetic aperture radar (SAR) images. Capturing of moving objects in the SAR images is a difficult task. In this study, a new automated model for detecting moving objects is proposed using SAR images. The proposed model has four main steps, namely, preprocessing, segmentation, feature extraction, and classification. Initially, the input SAR image is pre-processed using a histogram equalization technique. Then, the weighted Otsu-based segmentation algorithm is applied for segmenting the object regions from the pre-processed images. When using the weighted Otsu, the segmented grayscale images are not only clear but also retain the detailed features of grayscale images. Next, feature extraction is carried out by gray-level co-occurrence matrix (GLCM), median binary patterns (MBPs), and additive harmonic mean estimated local Gabor binary pattern (AHME-LGBP). The final step is classification using deep ensemble models, where the objects are classified by employing the ensemble deep learning technique, combining the models like the bidirectional long short-term memory (Bi-LSTM), recurrent neural network (RNN), and improved deep belief network (IDBN), which is trained with the features extracted previously. The combined models increase the accuracy of the results significantly. Furthermore, ensemble modeling reduces the variance and modeling method bias, which decreases the chances of overfitting. Compared to a single contributing model, ensemble models perform better and make better predictions. Additionally, an ensemble lessens the spread or dispersion of the model performance and prediction accuracy. Finally, the performance of the proposed model is related to the conventional models with respect to different measures. In the mean-case scenario, the proposed ensemble model has a minimum error value of 0.032, which is better related to other models. In both median- and best-case scenario studies, the ensemble model has a lower error value of 0.029 and 0.015.

Research on weed identification in soybean fields based on the lightweight segmentation model DCSAnet.

Weeds can compete with crops for sunlight, water, space and various nutrients, which can affect the growth of crops.In recent years, people have started to use self-driving agricultural equipment, robots, etc. for weeding work and use of drones for weed identification and spraying of weeds with herbicides, and the effectiveness of these mobile weeding devices is largely limited by the superiority of weed detection capability. To improve the weed detection capability of mobile weed control devices, this paper proposes a lightweight weed segmentation network model DCSAnet that can be better applied to mobile weed control devices. The whole network model uses an encoder-decoder structure and the DCA module as the main feature extraction module. The main body of the DCA module is based on the reverse residual structure of MobileNetV3, effectively combines asymmetric convolution and depthwise separable convolution, and uses a channel shuffle strategy to increase the randomness of feature extraction. In the decoding stage, feature fusion utilizes the high-dimensional feature map to guide the aggregation of low-dimensional feature maps to reduce feature loss during fusion and increase the accuracy of the model. To validate the performance of this network model on the weed segmentation task, we collected a soybean field weed dataset containing a large number of weeds and crops and used this dataset to conduct an experimental study of DCSAnet. The results showed that our proposed DCSAnet achieves an MIoU of 85.95% with a model parameter number of 0.57 M and the highest segmentation accuracy in comparison with other lightweight networks, which demonstrates the effectiveness of the model for the weed segmentation task.

A novel feature enhancement and semantic segmentation scheme for identifying low-contrast ocean oil spills

The oil spill accidents on the sea surface pose a severe threat to the marine environment and human health. This paper proposes a novel Semantic Segmentation Network (SSN) for processing oil spill images so that low-contrast oil spills on the sea surface can be accurately identified. After the detection accuracy and real-time performance of the current SSNs are compared, the basic network architecture of DeeplabV3+ based target detection is analyzed. The standard convolution is replaced by the Omni-dimensional Dynamic Convolution (ODConv) in the Ghost Module Depth-Wise separable Convolution (DWConv) to further enhance the feature extraction ability of the network. Furthermore, a new DeeplabV3+ based network with ODGhostNetV2 is constructed as the main feature extraction module, and an Adaptive Triplet Attention (ATA) module is deployed in the encoder and decoder at the same time. This not only improves the richness of semantic features but also increases the following receptive fields of the network model. ATA integrates the Adaptively Spatial Feature Fusion (ASFF) module to optimize the weight assignment problem in the feature map fusion process. The ablation experiments are conducted to verify the proposed network which show high accuracy and good real-time performance for the oil spill detection.

A swin transformer based bird nest detection approach with unmanned aerial vehicle images for power distribution and pole towers

AbstractThe authors propose a novel object detection algorithm for identifying bird nests in medium voltage power line aerial images, which is crucial for ensuring the safe operation of the power grid. The algorithm utilises an improved Swin Transformer as the main feature extraction network of Fast R‐CNN, further enhanced with a channel attention and modified binary self‐attention mechanism to improve the feature representation ability. The proposed algorithm is evaluated on a newly constructed image dataset of medium voltage transmission lines containing bird nests, which are annotated and classified. Experimental results show that the proposed algorithm achieves satisfied accuracy and robustness in recognising bird nests compared to traditional algorithms.

Current Trends in Feature Extraction and Classification Methodologies of Biomedical Signals.

Biomedical signal and image processing is the study of the dynamic behavior of various bio-signals, which benefits academics and research. Signal processing is used to assess the behavior of analogue and digital signals for the assessment, reconfiguration, improved efficiency, extraction of features, and reorganization of patterns. This paper unveils hidden characteristic information about input signals using feature extraction methods. The main feature extraction methods used in signal processing are based on studying time, frequency, and frequency domain. Feature exaction methods are used for data reduction, comparison, and reducing dimensions, producing the original signal with sufficient accuracy with a structure of an efficient and robust pattern for the classifier system. Therefore, an attempt has been made to study the various feature extraction methods, feature transformation methods, classifiers, and datasets for biomedical signals.

Deep learning based concrete compressive strength prediction model with hybrid meta-heuristic approach

In concrete design, the compressive strength of the concrete is the critical parameter that defines the quality of concrete. Determination of the compressive strength of concrete by laboratory testing is the conventional way, however, with the help of advancements in technology, the development of an exact predicting technique for the compressive strength of concrete will reduce time as well as cost to a greater extent. There are various types of concrete, as concrete is the most used material and has wide applications in the construction sector and hence huge article is being done to increase the performance of concrete. In this paper, the compressive strength of High-Performance Concrete and Fiber Reinforced High Strength Self-Compacting Concrete is predicted by next three main stages, namely (a) pre-processing, (b) feature extraction, and (c) prediction of concrete compressive strength. Initially, in the data pre-processing phase, duplicate data samples are removed. Subsequently, various features like statistical features, symmetrical uncertainty-based features, higher order statistical features, ReliefF feature, and are derived from the pre-processed data. In the prediction phase of concrete compressive strength, a hybrid algorithm is designed by mingling the concepts of Deep Belief Network and Long Short-Term Memory. Moreover, The Rock Hyraxes Updated Sparrow Algorithm is used to fine-tune the weights of DBN to improve the prediction accuracy of the suggested work. The suggested hybrid optimization method is the hybridization of the standard Sparrow Search Algorithm and Rock Hyraxes Swarm Optimization, respectively. Eventually, the output from the optimized hybrid method will reveal the compression energy of concrete. Ultimately, the efficiency of the suggested method is verified by the comparative estimation. The MAPE of the suggested method is 235.32, which is 30.6%, 21.64%, 20.4%, and 11.4% superior than the CLPSO, ABC, SSA + Hybrid Classifier and RHSO + Hybrid Classifier, respectively.

Copy-Move Forgery Detection and Localization Using Deep Learning

Forgery detection is one of the challenging subjects in computer vision. Forgery is performed using image manipulation with editor tools. Image manipulation tries to change the concept of the image but preserves the integrity of the texture and structure of the image as much as possible. Images are used as evidence in some applications, so if the images are manipulated, they will not be reliable. The copy-move forgery is one of the simplest image manipulation methods. This method removes or inserts information into the image with the least clue by copying a part of the image and pasting it into other places of the same image. Recently, traditional (block-based and keypoint-based) and deep learning methods have been proposed to detect forgery images. Traditional methods include two main steps, feature extraction, and feature matching. Unlike the traditional methods, the deep learning method performs the forgery detection automatically by extracting hierarchical features directly from the data. This paper presents a deep learning method for forgery detection at both image and pixel levels. In this method, we used a pre-trained deep model with a global average pooling (GAP) layer instead of default fully connected layers to detect forgery. The GAP layer creates a good dependency between the feature maps and the classes. In pixel forgery detection, a visualization technique called heatmap activation is used in forgery images. This technique identifies parts of the image that are candidates for forgery. Then, the best candidate is selected and the location of the forgery is determined. The proposed method is performed on the CoMoFod and MICC datasets. The extensive experiments showed the satisfactory performance of the proposed method.

Texture features extraction technology using Grey Level Co-Occurance Matrix for the KNN classification of citrus disease

The citrus disease is a problem affecting the decrease of agricultural commodity yields. One way to determine disease in citrus is through the leaves. Leaves, as a place for photosynthesis, with disease will cause stunted plant growth. Therefore, the fruit can experience a quality decrease. This study aims to classify citrus diseases based on leaf images by applying extraction technology of GLCM (Gray Level Co-occurrence Matrix) using KNN (K-Nearest Neighbor). Citrus disease classification has four main stages, namely preprocessing, segmentation, feature extraction, and classification. The preprocessing stage converts the LAB color space. Segmentation stage uses Otsu Thresholding. Texture features extraction uses GLCM. Classification uses KNN. KNN classification uses several distances, namely Chi-Square, City Block (Manhattan), Correlation, Cosine, Euclidean, and Hassanat. Comparisons are made based on the normalization of the dataset and the KNN distance used. The dataset without normalization gets the best results with Hassanat distance KNN (k = 29) with an accuracy of 91.86% and the dataset with normalization gets the best results at Euclidean distance (k = 7) with an accuracy of 98.84%. This research was expected to find out the accuracy of the method mentioned above in the classification of citrus diseases.

Rapid detection of Yunnan Xiaomila based on lightweight YOLOv7 algorithm.

Real-time fruit detection is a prerequisite for using the Xiaomila pepper harvesting robot in the harvesting process. To reduce the computational cost of the model and improve its accuracy in detecting dense distributions and occluded Xiaomila objects, this paper adopts YOLOv7-tiny as the transfer learning model for the field detection of Xiaomila, collects images of immature and mature Xiaomila fruits under different lighting conditions, and proposes an effective model called YOLOv7-PD. Firstly, the main feature extraction network is fused with deformable convolution by replacing the traditional convolution module in the YOLOv7-tiny main network and the ELAN module with deformable convolution, which reduces network parameters while improving the detection accuracy of multi-scale Xiaomila targets. Secondly, the SE (Squeeze-and-Excitation) attention mechanism is introduced into the reconstructed main feature extraction network to improve its ability to extract key features of Xiaomila in complex environments, realizing multi-scale Xiaomila fruit detection. The effectiveness of the proposed method is verified through ablation experiments under different lighting conditions and model comparison experiments. The experimental results indicate that YOLOv7-PD achieves higher detection performance than other single-stage detection models. Through these improvements, YOLOv7-PD achieves a mAP (mean Average Precision) of 90.3%, which is 2.2%, 3.6%, and 5.5% higher than that of the original YOLOv7-tiny, YOLOv5s, and Mobilenetv3 models, respectively, the model size is reduced from 12.7 MB to 12.1 MB, and the model's unit time computation is reduced from 13.1 GFlops to 10.3 GFlops. The results shows that compared to existing models, this model is more effective in detecting Xiaomila fruits in images, and the computational complexity of the model is smaller.

An End-to-End Multi-Scale Conditional Generative Adversarial Network for Image Deblurring

<p>For image deblurring, multi-scale approaches have been widely used as deep learning methods recently. In this paper, a novel multi-scale conditional generative adversarial network (CGAN) is proposed to make full use of image features, which outperforms most state-of-the-art methods. We define a generator network and a discriminator network. First of all, we use the multi-scale residual modules proposed in this paper as main feature extraction blocks, and add skip connections to extract multi-scale image features at a finer granularity in the generator network. Secondly, we construct PatchGAN as the discriminator network to enhance the local feature extraction capability. In addition, we combine the adversarial loss based on Wasserstein GAN with gradient penalty (WGAN-GP) theory with the content loss defined by perceptual loss as the total loss function, which is conducive to improving the consistency between the generated images and the ground-truth sharp images in content. The experimental results show that the method in this paper outperforms the state-of-the-art methods in visualization and quantitative results.</p> <p> </p>

Graph Convolution Point Cloud Super-Resolution Network Based on Mixed Attention Mechanism

In recent years, point cloud super-resolution technology has emerged as a solution to generate a denser set of points from sparse and low-quality point clouds. Traditional point cloud super-resolution methods are often optimized based on constraints such as local surface smoothing; thus, these methods are difficult to be used for complex structures. To address this problem, we proposed a novel graph convolutional point cloud super-resolution network based on a mixed attention mechanism (GCN-MA). This network consisted of two main parts, i.e., feature extraction and point upsampling. For feature extraction, we designed an improved dense connection module that integrated an attention mechanism and graph convolution, enabling the network to make good use of both global and local features of the point cloud for the super-resolution task. For point upsampling, we adopted channel attention to suppress low-frequency information that had little impact on the up-sampling results. The experimental results demonstrated that the proposed method significantly improved the point cloud super-resolution performance of the network compared to other corresponding methods.

Deep Learning for Visual SLAM: The State-of-the-Art and Future Trends

Visual Simultaneous Localization and Mapping (VSLAM) has been a hot topic of research since the 1990s, first based on traditional computer vision and recognition techniques and later on deep learning models. Although the implementation of VSLAM methods is far from perfect and complete, recent research in deep learning has yielded promising results for applications such as autonomous driving and navigation, service robots, virtual and augmented reality, and pose estimation. The pipeline of traditional VSLAM methods based on classical image processing algorithms consists of six main steps, including initialization (data acquisition), feature extraction, feature matching, pose estimation, map construction, and loop closure. Since 2017, deep learning has changed this approach from individual steps to implementation as a whole. Currently, three ways are developing with varying degrees of integration of deep learning into traditional VSLAM systems: (1) adding auxiliary modules based on deep learning, (2) replacing the original modules of traditional VSLAM with deep learning modules, and (3) replacing the traditional VSLAM system with end-to-end deep neural networks. The first way is the most elaborate and includes multiple algorithms. The other two are in the early stages of development due to complex requirements and criteria. The available datasets with multi-modal data are also of interest. The discussed challenges, advantages, and disadvantages underlie future VSLAM trends, guiding subsequent directions of research.

Building Occupancy Detection and Localization Using CCTV Camera and Deep Learning

Occupancy information plays a key role in analyzing and improving building energy performance. The advances of Internet of Things (IoT) technologies have engendered a shift in measuring building occupancy with IoT sensors, in which cameras in closed-circuit television (CCTV) systems can provide richer measurements. However, existing camera-based occupancy detection approaches cannot function well when scanning videos with a number of occupants and determining occupants’ locations. This article aims to develop a novel deep-learning-based approach for better building occupancy detection based on CCTV cameras. To do so, this research proposes a deep-learning model to detect the number of occupants and determine their locations in videos. This model consists of two main modules, namely, feature extraction and three-stage occupancy detection. The first module presents a deep convolutional neural network to perform residual and multibranch convolutional calculation to extract shallow and semantic features, and constructs feature pyramids through a bidirectional feature network. The second module performs a three-stage detection procedure with three sequential and homogeneous detectors which have increasing Intersection over Union (IoU) thresholds. Empirical experiments evaluate the detection performance of the approach with CCTV videos from a university building. Experimental results show that the approach achieves the superior detection performance when compared with baseline models.

Multi-modality deep forest for hand motion recognition via fusing sEMG and acceleration signals.

Bio-signal based hand motion recognition plays a critical role in the tasks of human-machine interaction, such as the natural control of multifunctional prostheses. Although a large number of classification technologies have been taken to improve the motion recognition accuracy, it is still a challenge to achieve acceptable performance for multiple modality input. This study proposes a multi-modality deep forest (MMDF) framework to identify hand motions, in which surface electromyographic signals (sEMG) and acceleration signals (ACC) are fused at the input level. The proposed MMDF framework constitutes of three main stages, sEMG and ACC feature extraction, feature dimension reduction, and a cascade structure deep forest for classification. A public database "Ninapro DB7" is used to evaluate the performance of the proposed framework, and the experimental results show that it can achieve a significantly higher accuracy than that of competitors. Besides, our experimental results also show that MMDF outperforms other traditional classifiers with the input of the single modality of sEMG signals. In sum, this study verifies that ACC signals can be an excellent supplementary for sEMG, and MMDF is a plausible solution to fuse mulit-modality bio-signals for human motion recognition.

Multi-modal fusion of deep transfer learning based COVID-19 diagnosis and classification using chest x-ray images.

COVID-19 pandemic has a significant impact on the global health and daily lives of people living over the globe. Several initial tests are based on the detecting of the genetic material of the coronavirus, and they have a minimum detection rate with a time-consuming process. To overcome this issue, radiological images are recommended where chest X-rays (CXRs) are employed in the diagnostic process. This article introduces a new Multi-modal fusion of deep transfer learning (MMF-DTL) technique to classify COVID-19. The proposed MMF-DTL model involves three main processes, namely pre-processing, feature extraction, and classification. The MMF-DTL model uses three DL models namely VGG16, Inception v3, and ResNet 50 for feature extraction. Since a single modality would not be adequate to attain an effective detection rate, the integration of three approaches by the use of decision-based multimodal fusion increases the detection rate. So, a fusion of three DL models takes place to further improve the detection rate. Finally, a softmax classifier is employed for test images to a set of six different. A wide range of experimental result analyses is carried out on the Chest-X-Ray dataset. The proposed fusion model is found to be an effective tool for COVID-19 diagnosis using radiological images with the average sensy of 92.96%, specy of 98.54%, precn of 93.60%, accuy of 98.80%, Fscore of 93.26% and kappa of 91.86%.

Wi-Fi fingerprinting-based floor detection using adaptive scaling and weighted autoencoder extreme learning machine

In practical applications, accurate floor determination in multi-building/floor environments is particularly useful and plays an increasingly crucial role in the performance of location-based services. An accurate and robust building and floor detection can reduce the location search space and ameliorate the positioning and wayfinding accuracy. As an efficient solution, this paper proposes a floor identification method that exploits statistical properties of wireless access point propagated signals to exponent received signal strength (RSS) in the radio map. Then, using single-layer extreme learning machine-weighted autoencoder (ELM-WAE) main feature extraction and dimensional reduction is implemented. Finally, ELM based classifier is trained over a new feature space to determine floor level. For the efficiency evaluation of our proposed model, we utilized three different datasets captured in the real scenarios. The evaluation result shows that the proposed model can achieve state-of-art performance and improve the accuracy of floor detection compared with multiple recent techniques. In this way, the floor level can be identified with 97.30%, 95.32%, and 96.39% on UJIIndoorLoc, Tampere, and UTSIndoorLoc datasets, respectively.