CNN-based Classification Model Research Articles

Diabetic Retinopathy Severity Level Detection Using Convolution Neural Network

Diabetic retinopathy is a common complication of diabetes mellitus, leading to damage and blockage of retinal blood vessels. Early and accurate detection of diabetic retinopathy severity levels is crucial for timely treatment and prevention of blindness. Diagnostic methods rely on manual examination and human interpretation, resulting in slower and less efficient treatment processes. As a branch of artificial intelligence, computer vision offers a potential solution to analyze retinal images quickly and accurately. The developed system employs image processing techniques and a CNN-based classification model to detect and classify the severity levels of diabetic retinopathy. By providing an automated and efficient approach, the system aims to assist doctors and optometrists in making informed decisions and reducing subjectivity in diagnosis. Early detection through this system can facilitate prompt treatment and improve patient outcomes. The developed system achieves promising results through experimentation and testing with various datasets, with accuracy ranging from 80% to 97%. This project's integration of artificial intelligence, machine learning, and image processing technologies demonstrates their potential in healthcare applications, particularly in diabetic retinopathy diagnosis.

Learning visual stimulus-evoked EEG manifold for neural image classification

Visual neural decoding, namely the ability to interpret external visual stimuli from patterns of brain activity, is a challenging task in neuroscience research. Recent studies have focused on characterizing patterns of activity across multiple neurons that can be described in terms of population-level features. In this study, we combine spatial, spectral, and temporal features to achieve neural manifold classification capable to characterize visual perception and to simulate the working memory activity in the human brain. We treat spatio-temporal and spectral information separately by means of custom deep learning architectures based on Riemann manifold and the two-dimensional EEG spectrogram representation. In addition, a CNN-based classification model is used to classify visual stimulus-evoked EEG signals while viewing the 11-class (i.e., all-black plus 0-9 digit images) MindBigData Visual MNIST dataset. The effectiveness of the proposed integration strategy is evaluated on the stimulus-evoked EEG signal classification task, achieving an overall accuracy of 86%, comparable to state-of-the-art benchmarks.

Deep learning-based phenotype classification of three ark shells: Anadara kagoshimensis, Tegillarca granosa, and Anadara broughtonii

The rapid and accurate classification of aquatic products is crucial for ensuring food safety, production efficiency, and economic benefits. However, traditional manual methods for classifying ark shell species based on phenotype are time-consuming and inefficient, especially during peak seasons when the demand is high and labor is scarce. This study aimed to develop a deep learning model for the automated identification and classification of commercially important three ark shells (Tegillarca granosa, Anadara broughtonii, and Anadara kagoshimensis) from images. The ark shells were collected and identified using a polymerase chain reaction method developed in a previous study, and a total of 1,400 images were categorized into three species. Three convolutional neural network (CNN) models, Visual Geometry Group Network (VGGnet), Inception-Residual Network (ResNet), and SqueezeNet, were then applied to two different classification sets, Set-1 (four bivalve species) and Set-2 (three ark shell species). Our results showed that SqueezeNet demonstrated the highest accuracy during the training phase for both classification sets, whereas Inception-ResNet exhibited superior accuracy during the validation phase. Similar results were obtained after developing a third classification set (Set-3) to classify six categories by combining Set-1 and Set-2. Overall, the developed CNN-based classification model exhibited a performance comparable or superior to that presented in previous studies and can provide a theoretical basis for bivalve classification, thereby contributing to improved food safety, production efficiency, and economic benefits in the aquatic products industry.

Hyperspectral image classification method based on narrowing semantic gap convolutional neural network

ABSTRACT In recent years, numerous hyperspectral image classification methods based on convolutional neural network (CNN) have been proposed in recent years. However, the majority of CNN-based classification models fail to take into account the semantic gap between shallow and deep features, and instead, just sum up or splice these two types of features, which leads to subpar feature fusion. To address the above problems, the paper proposes a hyperspectral image classification method based on Narrowing Semantic Gap Convolutional Neural Network (NSGCNN). We developed the feature fusion sub-network, which consists of Feature Semantic Enhancement Module (FSEM), Spatial Detail Supplement Module (SDCM), and Cross-layer Feature Fusion Module (CFFM). FSEM extracts rich global semantic information based on multi-scale dilated convolution. Reverse attention mechanism is introduced in the SDCM to complement deep-layer features with fine spatial features like edges and textures. The fusing of hierarchical features is then achieved using CFFM. Additionally, we can regulate the model’s parameter count by regulating the number of output feature maps. Experimental results on three benchmark hyperspectral datasets demonstrate that NSGCNN outperforms other state-of-the-art methods in terms of classification performance and significantly outperforms them in terms of model complexity.

Classification of Bovine Cumulus-Oocyte Complexes with Convolutional Neural Networks

Aim: Determining oocyte quality is crucial for successful fertilization and embryonic development, and there is a serious correlation between live birth rates and oocyte quality. Parameters such as the regular/irregular formation of the cumulus cell layer around the oocyte, the number of cumulus cell layers and the homogeneity of the appearance of the ooplasm are used to determine the quality of the oocytes to be used in in vitro fertilization (IVF) and intracytoplasmic sperm injection (ICSI) methods. Material and Methods: In this study, classification processes have been carried out using convolutional neural networks (CNN), a deep learning method, on the images of the cumulus-oocyte complex selected based on the theoretical knowledge and professional experience of embryologists. A convolutional neural network with a depth of 4 is used. In each depth level, one convolution, one ReLU and one max-pooling layer are included. The designed network architecture is trained using the Adam optimization algorithm. The cumulus-oocyte complexes (n=400) used in the study were obtained by using the oocyte aspiration method from the ovaries of the bovine slaughtered at the slaughterhouse. Results: The CNN-based classification model developed in this study showed promising results in classifying three-class image data in terms of cumulus-oocyte complex classification. The classification model achieved high accuracy, precision, and sensitivity values on the test dataset. Conclusion: Continuous research and optimization of the model can further improve its performance and benefit the field of cumulus-oocyte complexes classification and oocyte quality assessment.

VMD and CNN-Based Classification Model for Infrasound Signal

VMD and CNN-Based Classification Model for Infrasound Signal

Automated atrial fibrillation classification based on denoising stacked autoencoder and optimized deep network

This study proposed an efficient ECG preprocessing technique, and the preprocessed ECG data was subsequently used to classify atrial fibrillation (AFib) using end-to-end deep neural networks. This is significant since early identification of AFib can help to prevent mortality. With this in mind, the two-fold study is proposed, in which three denoising autoencoders for ECG signal pre-processing were evaluated and compared. Denoising autoencoders (DAEs) utilising convolutional neural networks (CNNs) as backbones outperformed the other two DAE models. As a result, the preprocessed signal is then used to classify atrial fibrillation using the three classification models. The combination of CNN-based DAE and CNN-based classification model yielded the best results, with 99.20 percent accuracy, 99.50 percent specificity, 99.50 percent sensitivity, and 99.00 percent true positive rate. The average accuracy of the algorithms we investigated was 96.26 percent, and our technique was 3.2 percent more accurate than the other algorithms we compared. Furthermore, with a 24-h ECG signal processing time of only 1.3 s, the proposed model is computationally inexpensive for real-time applications. To determine its robustness, the proposed framework is tested on previously unseen datasets with varied proportions of arrhythmias, producing a 99.10 percent recall rate and a 98.50 percent accuracy.

An improved 3-D attention CNN with hybrid loss and feature fusion for pulmonary nodule classification

Background and ObjectiveLung cancer has the highest cancer-related mortality worldwide, and lung nodule usually presents with no symptom. Low-dose computed tomography (LDCT) was an important tool for lung cancer detection and diagnosis. It provided a complete three-dimensional (3-D) chest image with a high resolution.Recently, convolutional neural network (CNN) had flourished and been proven the CNN-based computer-aided diagnosis (CADx) system could extract the features and help radiologists to make a preliminary diagnosis. Therefore, a 3-D ResNeXt-based CADx system was proposed to assist radiologists for diagnosis in this study. MethodsThe proposed CADx system consists of image preprocessing and a 3-D CNN-based classification model for pulmonary nodule classification. First, the image preprocessing was executed to generate the normalized volumn of interest (VOI) only including nodule information and a few surrounding tissues. Then, the extracted VOI was forwarded to the 3-D nodule classification model. In the classification model, the RestNext was employed as the backbone and the attention scheme was embedded to focus on the important features. Moreover, a multi-level feature fusion network incorporating feature information of different scales was used to enhance the prediction accuracy of small malignant nodules. Finally, a hybrid loss based on channel optimization which make the network learn more detailed information was empolyed to replace a binary cross-entropy (BCE) loss. ResultsIn this research, there were a total of 880 low-dose CT images including 440 benign and 440 malignant nodules from the American National Lung Screening Trial (NLST) for system evaluation. The results showed that our system could achieve the accuracy of 85.3%, the sensitivity of 86.8%, the specificity of 83.9%, and the area-under-curve (AUC) value was 0.9042. It was confirmed that the designed system had a good diagnostic ability. ConclusionIn this study, a CADx composed of the image preprocessing and a 3-D nodule classification model with attention scheme, feature fusion, and hybrid loss was proposed for pulmonary nodule classification in LDCT. The results indicated that the proposed CADx system had potential for achieving high performance in classifying lung nodules as benign and malignant.

Classification of Tea Leaves Based on Fluorescence Imaging and Convolutional Neural Networks.

The development of the smartphone and computer vision technique provides customers with a convenient approach to identify tea species, as well as qualities. However, the prediction model may not behave robustly due to changes in illumination conditions. Fluorescence imaging can induce the fluorescence signal from typical components, and thus may improve the prediction accuracy. In this paper, a tea classification method based on fluorescence imaging and convolutional neural networks (CNN) is proposed. Ultra-violet (UV) LEDs with a central wavelength of 370 nm were utilized to induce the fluorescence of tea samples so that the fluorescence images could be captured. Five kinds of tea were included and pre-processed. Two CNN-based classification models, e.g., the VGG16 and ResNet-34, were utilized for model training. Images captured under the conventional fluorescent lamp were also tested for comparison. The results show that the accuracy of the classification model based on fluorescence images is better than those based on the white-light illumination images, and the performance of the VGG16 model is better than the ResNet-34 model in our case. The classification accuracy of fluorescence images reached 97.5%, which proves that the LED-induced fluorescence imaging technique is promising to use in our daily life.

Deep learning-based classification of damage-induced acoustic emission signals in UHPC

Acoustic emission (AE) is advantageous in studying and monitoring the damage state of UHPC, and its performance can be further improved if various damage-induced AE signals are classified accurately. The deep learning method is promising to enhance the classification accuracy, but it is limited by the absence of the corresponding AE dataset. In this study, fiber pullout tests and direct tension tests (DTTs) were performed to explore the time–frequency features of various damage-induced AE signals in UHPC. The AE sources can be classified into five categories: matrix cracking, fiber debonding, fiber sliding, fiber scraping, and matrix spalling. A corresponding AE dataset was established after manual pre-labeling, pseudo-labeling with transfer learning, and mix-up data augmenting. Several lightweight convolutional neural networks (CNNs) were trained from scratch based on the dataset, with modified time–frequency spectrograms as data input. The performances of CNN-based classification models are superior to traditional shallow-learning methods. The popular lightweight CNN structure of ResNet18 achieves the highest overall accuracy of 93.94% among adopted CNN structures such as GoogleNet, ResNet18, EfficientNet-b0, and MobileNetV2. The AE signal classification results of DTTs show that fiber-pullout damage and matrix cracking occur intensively at the strain-hardening phase and increase with fiber content. That suggests the increasing fiber bridging and multi-crack propagation in the uniaxial tension of UHPC with higher fiber content.

DFU_SPNet: A stacked parallel convolution layers based CNN to improve Diabetic Foot Ulcer classification

Diabetic Foot Ulcer (DFU) is a complication of diabetes that causes lower limb amputation. In this work, a unique stacked parallel convolution layers-based network (DFU_SPNet) is proposed to perform DFU vs. normal skin classification. The main objective of this work is to design an effective CNN-based classification model, along with proper fine-tuning of optimizer settings. DFU_SPNet consists of 3 blocks of parallel convolution layers with multiple kernel sizes, for local and global feature abstractions. The proposed DFU_SPNet, trained using SGD (with momentum) optimizer with 1e−2 learning rate on the DFUNet dataset, outperformed the current state-of-the-art results with an AUC of 0.974.

Assessment of deep convolutional neural network models for mandibular fracture detection in panoramic radiographs

The aim of this study was to develop automated models for the identification and detection of mandibular fractures in panoramic radiographs using convolutional neural network (CNN) algorithms. A total of 1710 panoramic radiograph images from the years 2016 to 2020, including 855 images containing mandibular fractures, were obtained retrospectively from the regional trauma centre. CNN-based classification models, DenseNet-169 and ResNet-50, were fabricated to identify fractures in the radiographic images. The CNN-based object detection models Faster R-CNN and YOLOv5 were trained to automate the placement of the bounding boxes to detect fractures in the radiographic images. The performance of the models was evaluated on a hold-out test set and also by comparison with residents in oral and maxillofacial surgery and oral and maxillofacial surgeons (experts) on a 100-image subset. The binary classification performance of the models achieved promising results with an area under the receiver operating characteristics curve (AUC), sensitivity, and specificity of 100%. The detection performance of the models achieved an AUC of approximately 90%. When compared with the accuracy of clinician observers, the identification performance of the models outperformed even an expert-level classification. In conclusion, CNN-based models identified mandibular fractures above expert-level performance. It is expected that these models will be used as an aid to improve clinician performance, with aided resident performance approximating that of expert level.

Open-set gas recognition: A case-study based on an electronic nose dataset

Electronic nose (E-Nose) has been widely used in detection and classification of gases. The learning models of traditional E-Noses are generally limited in closed-set environment: the training and test samples share the same label spaces. However, a more challenging and realistic scenario of E-Noses is open-set learning, where the test samples contains classes unseen during the model training. This study investigated the possibility of open-set learning models for the recognition and classification of gases based on a public electronic nose datasets. The dataset includes the response of a 72-channels MOS sensor array on 10 gaseous substances. The original data was preprocessed by two methods: one is to manually extract features from the response curve of each sample, and the other is to down-sample the original sample into a matrix. Then multilayer perceptron (MLP) and convolution neural network (CNN) were used to extract the feature vectors of the data processed by the two processing methods respectively. The performance of four different open-set recognition models, including softmax threshold (ST), OpenMax, extreme value machine (EVM) and class anchor clustering (CAC), was compared based on the feature vectors obtained from two neural networks. To understand the effect of sensor drift on the models, we also validated the models on a commonly used sensor drift dataset. The results demonstrated that for the open-set detection task, the CNN-based CAC (CAC-CNN) outperformed the other methods. For the closed-set recognition task, the CNN-based classification model achieved higher accuracy. On sensor drift dataset, the performance of open-set recognition models has decreased a lot, and it seems that drift has a large negative impact on the open-set gas recognition.

Early-Stage Brown Spot Disease Recognition in Paddy Using Image Processing and Deep Learning Techniques

India is an agricultural country. Paddy is the main crop here on which the livelihood of millions of people depends. Brown spot disease caused by fungus is the most predominant infection that appears as oval and round lesions on the paddy leaves. If not addressed on time, it might result in serious crop loss. Pesticide use for plant disease treatment should be limited because it raises costs and pollutes the environment. Usage of pesticide and crop loss both can be minimized if we recognize the disease in a timely manner. Our aim is to develop a simple, fast, and effective deep learning structure for early-stage brown spot disease detection by utilizing infection severity estimation using image processing techniques. The suggested approach consists of two phases. In the first phase, the brown spot infected leaf image dataset is partitioned into two sets named as early-stage brown spot and developed stage brown spot. This partition is done on the basis of calculated infection severity. Infection severity is computed as a ratio of infected pixel count to total leaf pixel count. Total leaf pixel counts are determined by segmenting the leaf region from the background image using Otsu's thresholding technique. Infected pixel counts are determined by segmenting infected regions from leaf regions using Triangle thresholding segmentation. In the second phase, a fully connected CNN architecture is built for automatic feature extraction and classification. The CNN-based classification model is trained and validated using early-stage brown spot, developed stage brown spot, and healthy leaves images of rice plants. Early-stage brown spot and developed stage brown spot images used in training and validation are the same images that are obtained in phase 1. The experimental analysis shows that the proposed fully connected CNN-based early-stage brown spot disease recognition model is an effective approach. The classification accuracy of the suggested model is found to be 99.20%. The result of the suggested method is compared with those existing CNN-based disease recognition and classification methods that have used leaf images to recognize the diseases. It is observed that the performance of our method is significantly better than compared methods.

Performance of deep convolutional neural network for classification and detection of oral potentially malignant disorders in photographic images

Oral potentially malignant disorders (OPMDs) are a group of conditions that can transform into oral cancer. The purpose of this study was to evaluate convolutional neural network (CNN) algorithms to classify and detect OPMDs in oral photographs. In this study, 600 oral photograph images were collected retrospectively and grouped into 300 images of OPMDs and 300 images of normal oral mucosa. CNN-based classification models were created using DenseNet-121 and ResNet-50. The detection models were created using Faster R-CNN and YOLOv4. The image data were randomly selected and assigned as training, validating, and testing data. The testing data were evaluated to compare the performance of the CNN models with the diagnosis results produced by oral and maxillofacial surgeons. DenseNet-121 and ResNet-50 were found to produce high efficiency in diagnosis of OPMDs, with an area under the receiver operating characteristic curve (AUC) of 95%. Faster R-CNN yielded the highest detection performance, with an AUC of 74.34%. For the CNN-based classification model, the sensitivity and specificity were 100% and 90%, respectively. For the oral and maxillofacial surgeons, these values were 91.73% and 92.27%, respectively. In conclusion, the DenseNet-121, ResNet-50 and Faster R-CNN models have potential for the classification and detection of OPMDs in oral photographs.

Using convolutional neural network for intelligent SAM inspection of flip chips

The flip-chip (FC) technology is widely used in microelectronic packaging. As the dimensions and pitch of the chip solder bumps are getting smaller, defect inspection becomes increasingly challenging. In this paper, scanning acoustic microscopy (SAM) was used for the non-destructive inspection of FC packages. However, the detection accuracy may be limited by a low resolution of the SAM image. Therefore, the image super-resolution (SR) method was employed to enhance the SAM image quality. The modified very-deep SR (VDSR) algorithm, based on a convolutional neural network (CNN), was investigated to reconstruct high resolution acoustic micrographs. In addition, a CNN-based classification model was designed for the classification of the solder joints. The results obtained in the present study demonstrate that the improved VDSR combined with the CNN classification model provides excellent performance for improving the SAM detection accuracy of the inspection of flip chip packages.

Spatial-Channel Attention-Based Class Activation Mapping for Interpreting CNN-Based Image Classification Models

Convolutional neural network (CNN) has been applied widely in various fields. However, it is always hindered by the unexplainable characteristics. Users cannot know why a CNN-based model produces certain recognition results, which is a vulnerability of CNN from the security perspective. To alleviate this problem, in this study, the three existing feature visualization methods of CNN are analyzed in detail firstly, and a unified visualization framework for interpreting the recognition results of CNN is presented. Here, class activation weight (CAW) is considered as the most important factor in the framework. Then, the different types of CAWs are further analyzed, and it is concluded that a linear correlation exists between them. Finally, on this basis, a spatial-channel attention-based class activation mapping (SCA-CAM) method is proposed. This method uses different types of CAWs as attention weights and combines spatial and channel attentions to generate class activation maps, which is capable of using richer features for interpreting the results of CNN. Experiments on four different networks are conducted. The results verify the linear correlation between different CAWs. In addition, compared with the existing methods, the proposed method SCA-CAM can effectively improve the visualization effect of the class activation map with higher flexibility on network structure.

Error-Tolerant Deep Learning for Remote Sensing Image Scene Classification.

Due to its various application potentials, the remote sensing image scene classification (RSSC) has attracted a broad range of interests. While the deep convolutional neural network (CNN) has recently achieved tremendous success in RSSC, its superior performances highly depend on a large number of accurately labeled samples which require lots of time and manpower to generate for a large-scale remote sensing image scene dataset. In contrast, it is not only relatively easy to collect coarse and noisy labels but also inevitable to introduce label noise when collecting large-scale annotated data in the remote sensing scenario. Therefore, it is of great practical importance to robustly learn a superior CNN-based classification model from the remote sensing image scene dataset containing non-negligible or even significant error labels. To this end, this article proposes a new RSSC-oriented error-tolerant deep learning (RSSC-ETDL) approach to mitigate the adverse effect of incorrect labels of the remote sensing image scene dataset. In our proposed RSSC-ETDL method, learning multiview CNNs and correcting error labels are alternatively conducted in an iterative manner. It is noted that to make the alternative scheme work effectively, we propose a novel adaptive multifeature collaborative representation classifier (AMF-CRC) that benefits from adaptively combining multiple features of CNNs to correct the labels of uncertain samples. To quantitatively evaluate the performance of error-tolerant methods in the remote sensing domain, we construct remote sensing image scene datasets with: 1) simulated noisy labels by corrupting the open datasets with varying error rates and 2) real noisy labels by deploying the greedy annotation strategies that are practically used to accelerate the process of annotating remote sensing image scene datasets. Extensive experiments on these datasets demonstrate that our proposed RSSC-ETDL approach outperforms the state-of-the-art approaches.

Neural network prediction of sound quality via domain Knowledge-Based data augmentation and Bayesian approach with small data sets

This study proposes a novel deep learning methodology to evaluate the interior noise in vehicles on mechanical and affective levels by employing small data sets. A convolutional neural network (CNN) model is constructed from the frequency-rpm spectrograms of vehicle noises to predict the mechanical attributes of the noise. The noises are classified based on the number of engine cylinders (3, 4, 6, and 8). Owing to the high variability in spectrograms, mathematical expressions for the engine order lines are derived to augment the training data. With respect to the affective attributes, three classes (i.e., ‘sporty,’ ‘powerful,’ and ‘luxurious’) are selected for noise characterization. The spectrograms are used to design another CNN-based classification model that predicts the affective attributes from the perspective of experts. Although expert knowledge is employed for data labeling, a quarter of the data remains unlabeled, owing to inherent subjectivity. The model is trained on the labeled data and is validated by comparing the predicted class probabilities for the unlabeled data and their distribution in the RGB color space. K-fold cross-validation is used to evaluate the reliability of the model. A regression model is built based on Bayesian inference to evaluate the affective attributes of the noise from the perspective of end users. Given four singular values from the matrix of sound quality metrics, the model predicts the mean jury ratings for noise. The classification models exhibit generalization performances of 98.2% and 91.6%, respectively, and the regression model exhibits a mean squared error of 2.57×10-3, thereby demonstrating the applicability of the proposed approach to vehicle noise analysis.

Multilevel Wavelet-SRNet for SAR Target Recognition

Speckle noise is an important factor affecting the accuracy of synthetic aperture radar (SAR) target recognition. Traditional speckle reduction methods based on transform domain and spatial filtering usually require professional experience to set the threshold, which will also affect the recognition accuracy. This letter proposes a multilevel wavelet speckle reduction network (Wavelet-SRNet) for noisy SAR images target recognition. First, the method designs the wavelet soft threshold denoising method as a trainable neural network module in the convolutional neural network (CNN) framework. Then, a two-level wavelet denoising branch is constructed and fused with the original noisy image. Finally, we cascade a CNN-based classification model on the above structure to form an SAR image target recognition network whose denoising threshold can be automatically learned. Experiments on the moving and stationary target acquisition and recognition database show that the classification accuracy of the proposed method for target recognition in noisy SAR images is better than the compared state-of-the-art methods. Also, the method achieved high test accuracy in the noise augmentation experiment.