Independent Convolutional Neural Networks Research Articles

Real Time Violence Detection and Alert System

Abstract: This paper talks about a précised safety model incorporating elements of Artificial Intelligence for real-time violence detection along with Alert System. This model utilizes the advancements in technology to rapidly respond to any potential violent assault incident. When these technologies are further developed, it also increases the possibilities of how they could be used in future to safeguard schools, public area safety, personal safety, and social stability. Numerous researches and trials have been conducted to counter violence with the passage of time that includes the installations of surveillance systems for warning or alerting to violent activities. Its main objective is to get surveillance systems to automatically annotate the violence activities and issue any Warning/Alerts. For this purpose, first of all the system proceeds with the process of foregrounding a person in each frame, then relevant frames are extracted and irrelevant are ignored, after this violent pattern is identified by the trained model is detected, and end up saving these frames as images. The image is then enhanced and the corresponding details like time and location are transmitted as an alert via Telegram app. The proposed technique is essentially based on deep learning for automatic violence detection using Convolutional Neural networks (CNN). For this diagnosis, a light weight pre-trained model, MobileNetV2, is used to ensure better accuracy than an independent CNN which requires massive computation time and reduced precision.

Toward Generalized Artificial Intelligence by Assessment Aggregation With Applications to Standard and Extreme Classifications.

The article proposes a plural learning framework combining the ingredients found in a tribunal for the derivation of a more generalized artificial intelligence (GAI) when starting from a specialized set of convolutional neural networks (CNNs). This framework involves at least two different training stages called, respectively, specialization and generalization. In the specialization stage, any CNN considered in a given set learns to predict independently of other elements of the set. In the second stage called generalization, an integration network learns to predict from assessment measures fed by downstream specialized CNNs. The assessment measures considered are categorical softmax probabilities and learning to judge from these assessments relies on independent CNNs. Generalization proof of concepts is provided in terms of multimodel, multimodal, and distributed schemes. The multimodel framework is such that different CNN models operating on the same modality cooperate for decision purpose. The multimodal framework implies specializations of CNN with respect to different input modalities. The distributed framework proposed is associated with assessment exchanges: it such that the aggregation aims at determining relevant joint assessments for mapping a given input to a single or a multiple output category. The performance of these aggregation frameworks is shown to be outstanding for both standard and extreme classification issues.

Unsupervised Domain Adaptive Dose Prediction Via Cross-Attention Transformer and Target-Specific Knowledge Preservation.

Radiotherapy is one of the leading treatments for cancer. To accelerate the implementation of radiotherapy in clinic, various deep learning-based methods have been developed for automatic dose prediction. However, the effectiveness of these methods heavily relies on the availability of a substantial amount of data with labels, i.e. the dose distribution maps, which cost dosimetrists considerable time and effort to acquire. For cancers of low-incidence, such as cervical cancer, it is often a luxury to collect an adequate amount of labeled data to train a well-performing deep learning (DL) model. To mitigate this problem, in this paper, we resort to the unsupervised domain adaptation (UDA) strategy to achieve accurate dose prediction for cervical cancer (target domain) by leveraging the well-labeled high-incidence rectal cancer (source domain). Specifically, we introduce the cross-attention mechanism to learn the domain-invariant features and develop a cross-attention transformer-based encoder to align the two different cancer domains. Meanwhile, to preserve the target-specific knowledge, we employ multiple domain classifiers to enforce the network to extract more discriminative target features. In addition, we employ two independent convolutional neural network (CNN) decoders to compensate for the lack of spatial inductive bias in the pure transformer and generate accurate dose maps for both domains. Furthermore, to enhance the performance, two additional losses, i.e. a knowledge distillation loss (KDL) and a domain classification loss (DCL), are incorporated to transfer the domain-invariant features while preserving domain-specific information. Experimental results on a rectal cancer dataset and a cervical cancer dataset have demonstrated that our method achieves the best quantitative results with [Formula: see text], [Formula: see text], and HI of 1.446, 1.231, and 0.082, respectively, and outperforms other methods in terms of qualitative assessment.

Surveillance-image-based outdoor air quality monitoring

Air pollution threatens human health, necessitating effective and convenient air quality monitoring. Recently, there has been a growing interest in using camera images for air quality estimation. However, a major challenge has been nighttime detection due to the limited visibility of nighttime images. Here we present a hybrid deep learning model, capitalizing on the temporal continuity of air quality changes for estimating outdoor air quality from surveillance images. Our model, which integrates a convolutional neural network (CNN) and long short-term memory (LSTM), adeptly captures spatial-temporal image features, enabling air quality estimation at any time of day, including PM2.5 and PM10 concentrations, as well as the air quality index (AQI). Compared to independent CNN networks that solely extract spatial features, our model demonstrates superior accuracy on self-constructed datasets with R2 = 0.94 and RMSE = 5.11 μg m−3 for PM2.5, R2 = 0.92 and RMSE = 7.30 μg m−3 for PM10, and R2 = 0.94 and RMSE = 5.38 for AQI. Furthermore, our model excels in daytime air quality estimation and enhances nighttime predictions, elevating overall accuracy. Validation across diverse image datasets and comparative analyses underscore the applicability and superiority of our model, reaffirming its applicability and superiority for air quality monitoring.

PM2.5 Concentration Prediction Model: A CNN–RF Ensemble Framework

Although many machine learning methods have been widely used to predict PM2.5 concentrations, these single or hybrid methods still have some shortcomings. This study integrated the advantages of convolutional neural network (CNN) feature extraction and the regression ability of random forest (RF) to propose a novel CNN-RF ensemble framework for PM2.5 concentration modeling. The observational data from 13 monitoring stations in Kaohsiung in 2021 were selected for model training and testing. First, CNN was implemented to extract key meteorological and pollution data. Subsequently, the RF algorithm was employed to train the model with five input factors, namely the extracted features from the CNN and spatiotemporal factors, including the day of the year, the hour of the day, latitude, and longitude. Independent observations from two stations were used to evaluate the models. The findings demonstrated that the proposed CNN-RF model had better modeling capability compared with the independent CNN and RF models: the average improvements in root mean square error (RMSE) and mean absolute error (MAE) ranged from 8.10% to 11.11%, respectively. In addition, the proposed CNN-RF hybrid model has fewer excess residuals at thresholds of 10 μg/m3, 20 μg/m3, and 30 μg/m3. The results revealed that the proposed CNN-RF ensemble framework is a stable, reliable, and accurate method that can generate superior results compared with the single CNN and RF methods. The proposed method could be a valuable reference for readers and may inspire researchers to develop even more effective methods for air pollution modeling. This research has important implications for air pollution research, data analysis, model estimation, and machine learning.

Neural Born Iterative Method for Solving Inverse Scattering Problems: 2D Cases

In this article, we propose the neural Born iterative method (NeuralBIM) for solving 2-D inverse scattering problems (ISPs) by drawing on the scheme of the physics-informed supervised residual learning (PhiSRL) to emulate the computing process of the traditional Born iterative method (TBIM). NeuralBIM uses independent convolutional neural networks (CNNs) to learn the alternate update rules of two different candidate solutions regarding the residuals. Two different schemes are presented in this article, including the supervised and unsupervised learning schemes. With the dataset generated by the method of moments (MoM), supervised NeuralBIM is trained with the knowledge of the total fields and contrasts. Unsupervised NeuralBIM is guided by the physics-embedded objective function founding on the governing equations of ISPs, which results in no requirement of the total fields and contrasts for training. Numerical and experimental results further validate the efficacy of NeuralBIM.

Particle identification in particle tracking velocimetry using two-stage neural networks

<p style='text-indent:20px;'>In this paper, a new sub-pixel intensive particle image identification method used for particle tracking velocimetry (PTV), is discussed. With aiming to improve the performance of accuracy and robustness, a two-stage deep learning framework consisting of two independent convolutional neural networks (CNN) is proposed in this method. The first neural network is to segment the particle blobs from the image, and the second one is to locate the position of the segmented particles at the sub-pixel level. A synthetic dataset containing particle images and the ground-truth positions is generated for network training. The effect of different characteristic parameters (e.g., the particle conditions; the tuning of the loss function; the noise level) is evaluated on synthetic images. To further verify the generalization capabilities of the technique, we also apply the proposed network to real-world images. Both synthetic and real-world experiment results strongly demonstrate that the proposed method possesses better accuracy and robustness than other conventional methods.</p>

Distributed Neural Network System for Multimodal Sleep Stage Detection

Existing automatic sleep stage detection methods predominantly use convolutional neural network classifiers (CNNs) trained on features extracted from single-modality signals such as electroencephalograms (EEG). On the other hand, multimodal approaches propose very complexly stacked network structures with multiple CNN branches merged by a fully connected layer. It leads to very high computational and data requirements. This study proposes replacing a stacked network with a distributed neural network system for multimodal sleep stage detection. It has relatively low computational and training data requirements while providing highly competitive results. The proposed multimodal classification and decision-making system (MM-DMS) method applies a fully connected shallow neural network, arbitrating between classification outcomes given by an assembly of independent convolutional neural networks (CNNs), each using a different single-modality signal. Experiments conducted on the CAP Sleep Database data, including the EEG-, ECG-, and EMG modalities representing six stages of sleep, show that the MM-DMS significantly outperforms each single-modality CNN. The fully-connected shallow network arbitration included in the MM-DMS outperforms the traditional majority voting-, average probability-, and maximum probability decision-making methods.

Image and Texture Independent Deep Learning Noise Estimation Using Multiple Frames

In this study, a novel multiple frame based image and texture independent Convolutional Neural Network (CNN) noise estimator is introduced. Noise estimation is a crucial step for denoising algorithms, especially for ones that are called “non-blind”. The estimator works for additive Gaussian noise for varying noise levels. The noise levels studied in this work have a standard deviation equal to 5 to 25 increasing 5 by 5. Since there is no database for noisy multiple images to train and validate the network, two frames of synthetic noisy images with a variety of noise levels are created by adding Additive White Gaussian Noise (AWGN) to each clean image. The proposed method is applied on the most popular gray level images besides the color image databases such as Kodak, McMaster, BSDS500 in order to compare the results with the other works. Image databases comprise indoor and outdoor scenes that have fine details and richer texture. The estimator has an accuracy rate of 99 % for the classification and favourable results for the regression. The proposed method outperforms traditional methods in most cases. And the regression output can be used with any non-blind denoising method.

Pathologic Image Classification of Flat Urothelial Lesions Using Pathologic Criteria-Based Deep Learning

Pathologic diagnosis of flat urothelial lesions is subject to high interobserver variability. We expected that deep learning could improve the accuracy and consistency of such pathologic diagnosis, although the learning process is a black box. We therefore propose a new approach for pathologic image classification incorporating the diagnostic process of the pathologist into a deep learning method. A total of 267 H&E-stained slides of normal urothelium and urothelial lesions from 127 cases were examined. Six independent convolutional neural networks were trained to classify pathologic images according to six pathologic criteria. We then used these networks in the main training for the final diagnosis. Compared with conventional manual analysis, our method significantly improved the classification accuracy of images of flat urothelial lesions. The automated classification showed almost perfect agreement (weighted κ = 0.98) with the consensus reading. In addition, our approach provides the advantages of reliable diagnosis corresponding to histologic interpretation. We used deep learning to establish an automated subtype classifier for flat urothelial lesions that successfully combines traditional morphologic approaches and complex deep learning to achieve a learning mechanism that seems plausible to the pathologist.

Prediction of sap flow with historical environmental factors based on deep learning technology

Sap flow is an important intermediate link that reflects the continuous soil–plant-atmosphere cycle. Therefore, it is important to predict the sap flow to analyze the amount of tree transpiration for assessing of water consumption. In this paper, we propose a new sap flow assessment using environmental factors based on a convolutional neural network-gated recurrent unit (CGRU) hybrid deep learning method. The model was trained and tested with the sap flow and environmental factors from 17,568 group observations from public SAPFLUXNET dataset. These group observations measured from January 1, 2012 to December 31, 2012, with acquisition interval of 30 min for one tree of New Zealand Agathis australis. After designed the CGRU structure by integrated a convolutional neural network (CNN) and a gated recurrent unit (GRU) neural network, the input variables were selected with a correlation analysis between sap flow and environmental factors. Additionally, the number of previous conditions were introduced into the input of the model. Results showed that when the number of previous conditions set to 16, the learning rate set to 0.01 with Adam optimization algorithm, the mean squared error, mean absolute percentage error, and coefficient of determination of the CGRU model were 0.00231, 22.31 and 0.948 respectively. Comparing results showed that the CGRU-based sap flow prediction model has more accuracy than other eight models participating in the test including the independent CNN, GRU, CNN-Long short-term memory (LSTM) and five traditional machine learning based models. The least time spent on training is also the CGRU model. The CGRU-based sap flow prediction model proposed in this paper can capture complex nonlinear dependencies and yield accurate assessments of sap flow. Our model we established in this paper can be useful for research on forest stand transpiration, water consumption and the soil–plant-atmosphere cycle.

DOA estimation using two independent convolutional neural networks with residual blocks

Acoustic source localization in the spherical harmonic domain with reverberation has hitherto not been investigated extensively. When the joint direction-of-arrival (DOA) estimation is treated as a classification task, it needs a lot of different classes and large-scale dataset, which is time-consuming and less cost-effective. Thus, a novel learning-based method is proposed in this paper, which separately estimates azimuths and elevations by two independent convolution neural networks (CNNs) with residual blocks. First, a mapping matrix is adopted to transform the spherical harmonic function into the Kronecker product of two Vandermonde vectors. Based on the special structure, features dependent on azimuth and elevation, respectively, are extracted as inputs to neural networks. This can divide the joint DOA estimation into two sub-tasks completed by two independent CNNs with residual blocks, which can be realized in parallel. Furthermore, the performance of the model learned by CNNs with residual blocks is better than ordinary CNNs. Simulations are conducted both on the simulated and real speech data for evaluating the performance of the proposed method. The results show that the method has higher accuracy. Moreover, the method is time-saving and effectively reduces the computational complexity.

Human activity recognition using CNN and long term short term memory

Human activity recognition aims to work out the activities performed by someone in a picture or video. Examples of actions are running, sitting, sleeping, and standing. Complex movement patterns and harmful occurrences like falling may be a part of these activities. The suggested ConvLSTM network can be created by successively combining fully connected layers, long immediate memory (LSTM) networks, and convolutional neural networks (CNN). The acquisition system will pre calculate skeleton coordinates using human detection and pose estimation from the image/video sequence. The ConvLSTM model builds new controlled features from the raw skeleton coordinates and their distinctive geometric and kinematic properties. Raw skeleton coordinates are utilized to generate geometric and kinematic properties supported by relative joint position values, joint differences, and their angular velocities. By utilizing a multi-player trained CNN-LSTM combination, novel spatiotemporal directed features can be obtained. The classification head with completely connected layers is then utilized. The suggested model was tested using the KinectHAR dataset, which consists of 130,000 samples with 81 attribute variables and was compiled using the Kinect (v2) sensor Experimental data is used to compare the performance of independent CNN and LSTM networks.

Study on Molecular Information Intelligent Diagnosis and Treatment of Bladder Cancer on Pathological Tissue Image.

BackgroundMolecular information about bladder cancer is significant for treatment and prognosis. The immunohistochemistry (IHC) method is widely used to analyze the specific biomarkers to determine molecular subtypes. However, procedures in IHC and plenty of reagents are time and labor-consuming and expensive. This study established a computer-aid diagnosis system for predicting molecular subtypes, p53 status, and programmed death-ligand 1 (PD-L1) status of bladder cancer with pathological images.Materials and MethodsWe collected 119 muscle-invasive bladder cancer (MIBC) patients who underwent radical cystectomy from January 2016 to September 2018. All the pathological sections are scanned into digital whole slide images (WSIs), and the IHC results of adjacent sections were recorded as the label of the corresponding slide. The tumor areas are first segmented, then molecular subtypes, p53 status, and PD-L1 status of those tumor-positive areas would be identified by three independent convolutional neural networks (CNNs). We measured the performance of this system for predicting molecular subtypes, p53 status, and PD-L1 status of bladder cancer with accuracy, sensitivity, and specificity.ResultsFor the recognition of molecular subtypes, the accuracy is 0.94, the sensitivity is 1.00, and the specificity is 0.909. For PD-L1 status recognition, the accuracy is 0.897, the sensitivity is 0.875, and the specificity is 0.913. For p53 status recognition, the accuracy is 0.846, the sensitivity is 0.857, and the specificity is 0.750.ConclusionOur computer-aided diagnosis system can provide a novel and simple assistant tool to obtain the molecular subtype, PD-L1 status, and p53 status. It can reduce the workload of pathologists and the medical cost.

Image Target Recognition Based on Multiregional Features under Hybrid Attention Mechanism

The growing volume of image data calls for better real-time performance of image feature extraction algorithms. To enhance the recognition accuracy of image targets, it is significant to build a more scientific deep learning network. Multimodal cross convolution or densely connected blocks have been introduced to classic deep learning networks, aiming to promote the recognition of image targets. However, these attempts fail to satisfactorily extract detailed features from the original image. To solve the problem, this paper explores the image target recognition based on multiregional features under hybrid attention mechanism. Specifically, a convolutional neural network (CNN) was established for extracting multiregional features based on the loss function of local feature aggregation. The model consists of three independent CNN modules, which are responsible for extracting the global multiregional features and the local features of different regions. Next, the channel domain attention mechanism and spatial domain attention mechanism were embedded in the proposed CNN, such that the model can recognize targets more accurately, without increasing the computing load. Finally, the proposed network was proved effective through the training and testing on a self-developed sample set of surveillance video images.

Doa Estimation Using Two Independent Convolutional Neural Networks with Residual Blocks

Doa Estimation Using Two Independent Convolutional Neural Networks with Residual Blocks

RAGCN: Region Aggregation Graph Convolutional Network for Bone Age Assessment From X-Ray Images

Rapid and accurate measurement of bone age from hand X-ray images is a significant task for children’s maturity assessment and metabolic disorders diagnosis. With the development of deep learning technology, assessment methods based on Convolutional Neural Network (CNN) have become the mainstream. However, the existing CNN method generates the assessment results solely based on images, which ignores the clinical practice and thus weakens the evaluation performance. In this paper, an automatic bone age assessment method based on CNN and Graph Convolutional Network (GCN) is proposed. The overall method uses CNN for feature extraction and GCN for bone key regions inference, which mimics the physician’s clinical process. Specifically, the key regions of the hand bone are firstly defined according to the clinical standard. Then, independent CNN pathways are established to extract the features of different key regions. Finally, a novel Region Aggregation Graph Convolutional Network (RAGCN) is designed, which can aggregate the region features into the overall bone age representation according to the adjacency relation of the regions. In addition, RAGCN can also infer the importance of different regions in the feature aggregation process. The proposed method is validated on the RSNA dataset and the RHPE dataset. The MAE is 4.09 months on the RSNA dataset and 6.78 months on the RHPE dataset, it is competitive and superior over other state-of-the-art methods.

An MRI sequence independent convolutional neural network for synthetic head CT generation in proton therapy

A magnetic resonance imaging (MRI) sequence independent deep learning technique was developed and validated to generate synthetic computed tomography (sCT) scans for MR guided proton therapy. 47 meningioma patients previously undergoing proton therapy based on pencil beam scanning were divided into training (33), validation (6), and test (8) cohorts. T1, T2, and contrast enhanced T1 (T1CM) MRI sequences were used in combination with the planning CT (pCT) data to train a 3D U-Net architecture with ResNet-Blocks. A hyperparameter search was performed including two loss functions, two group sizes of normalisation, and depth of the network. Training outcome was compared between models trained for each individual MRI sequence and for all sequences combined. The performance was evaluated based on a metric and dosimetric analysis as well as spot difference maps. Furthermore, the influence of immobilisation masks that are not visible on MRIs was investigated. Based on the hyperparameter search, the final model was trained with fixed features per group for the group normalisation, six down-convolution steps, an input size of 128×192×192, and feature loss. For the test dataset for body/bone the mean absolute error (MAE) values were on average 79.8/216.3Houndsfield unit (HU) when trained using T1 images, 71.1/186.1HU for T2, and 82.9/236.4HU for T1CM. The structural similarity metric (SSIM) ranged from 0.95 to 0.98 for all sequences. The investigated dose parameters of the target structures agreed within 1% between original proton treatment plans and plans recalculated on sCTs. The spot difference maps had peaks at ±0.2cm and for 98% of all spots the difference was less than 1cm. A novel MRI sequence independent sCT generator was developed, which suggests that the training phase of neural networks can be disengaged from specific MRI acquisition protocols. In contrast to previous studies, the patient cohort consisted exclusively of actual proton therapy patients (i.e. “real-world data”).

Reconnection of fragmented parts of coronary arteries using local geometric features in X-ray angiography images

The segmentation of coronary arteries in X-ray images is essential for image-based guiding procedures and the diagnosis of cardiovascular disease. However, owing to the complex and thin structures of the coronary arteries, it is challenging to accurately segment arteries in X-ray images using only a single neural network model. Consequently, coronary artery images obtained by segmentation with a single model are often fragmented, with parts of the arteries missing. Sophisticated post-processing is then required to identify and reconnect the fragmented regions. In this paper, we propose a method to reconstruct the missing regions of coronary arteries using X-ray angiography images. Method: We apply an independent convolutional neural network model considering local details, as well as a local geometric prior, for reconnecting the disconnected fragments. We implemented and compared the proposed method with several convolutional neural networks with customized encoding backbones as baseline models. Results: When integrated with our method, existing models improved considerably in terms of similarity with ground truth, with a mean increase of 0.330 of the Dice similarity coefficient in local regions of disconnected arteries. The method is efficient and is able to recover missing fragments in a short number of iterations. Conclusion and Significance: Owing to the restoration of missing fragments of coronary arteries, the proposed method enables a significant enhancement of clinical impact. The method is general and can simply be integrated into other existing methods for coronary artery segmentation.

Non-melanoma skin cancer diagnosis: a comparison between dermoscopic and smartphone images by unified visual and sonification deep learning algorithms

PurposeNon-melanoma skin cancer (NMSC) is the most frequent keratinocyte-origin skin tumor. It is confirmed that dermoscopy of NMSC confers a diagnostic advantage as compared to visual face-to-face assessment. COVID-19 restrictions diagnostics by telemedicine photos, which are analogous to visual inspection, displaced part of in-person visits. This study evaluated by a dual convolutional neural network (CNN) performance metrics in dermoscopic (DI) versus smartphone-captured images (SI) and tested if artificial intelligence narrows the proclaimed gap in diagnostic accuracy.MethodsA CNN that receives a raw image and predicts malignancy, overlaid by a second independent CNN which processes a sonification (image-to-sound mapping) of the original image, were combined into a unified malignancy classifier. All images were histopathology-verified in a comparison between NMSC and benign skin lesions excised as suspected NMSCs. Study criteria outcomes were sensitivity and specificity for the unified output.ResultsImages acquired by DI (n = 132 NMSC, n = 33 benign) were compared to SI (n = 170 NMSC, n = 28 benign). DI and SI analysis metrics resulted in an area under the curve (AUC) of the receiver operator characteristic curve of 0.911 and 0.821, respectively. Accuracy was increased by DI (0.88; CI 81.9–92.4) as compared to SI (0.75; CI 68.1–80.6, p < 0.005). Sensitivity of DI was higher than SI (95.3%, CI 90.4–98.3 vs 75.3%, CI 68.1–81.6, p < 0.001), but not specificity (p = NS).ConclusionTelemedicine use of smartphone images might result in a substantial decrease in diagnostic performance as compared to dermoscopy, which needs to be considered by both healthcare providers and patients.