Trained CNN Model Research Articles

Energy-constrained production optimization for the in-situ conversion process of oil shale based on deep learning algorithms

The in-situ conversion process (ICP) involves complicated thermal-reactive-compositional flow processes where the production parameters are closely interdependent. Therefore, manual optimization of the production parameters of ICP through numerical simulation is arduous and time-consuming. This paper introduces a computational framework that integrates deep learning techniques and particle swarm optimization (PSO) to automatically optimize the values of production parameters of ICP, and thus locating the highest energy efficiency and the optimum energy usage. This approach utilizes a 3D CNN model to predict key metrics. The prediction process considers the detailed 3D heterogeneity of oil shale, resulting in remarkably high prediction accuracy, as evidenced by determination coefficients (R2) above 0.99. Subsequently, the trained CNN model is integrated to the PSO algorithm to automatically fine-tune the production parameters to optimize the energy efficiency of ICP.The optimization results yield three significant findings. First, as the energy consumption limit increases, the optimal number of heaters, well spacing, and heating temperature exhibit an upward trend, but stabilized beyond the threshold of 7 × 105 kWh. Secondly, the optimal input energy (7 × 105 kWh) is found for the given ICP model. Lastly, the analysis reveals that variations in initial reservoir pressure or bottomhole pressure have limited impact on cumulative oil production and energy usage.

Feedback Analyser through Facial Expressions

Abstract—Feedback is an essential part of human connection and communication since it may used to express expression, feelings, and truth. This research paper works on the complex function that facial signals play in the interpretation and communication of information by offering a thorough overview and analysis of feedback through facial expressions. We investigate the brain systems that underlie facial feedback perception and response in humans, as well as the mechanisms behind face expressions, cultural variances, and their universality. Deep neural networks facial expression feedback analyzers are cur- rently in the starting stages of development, but they have the strength to solve a variety of issues with present human- computer interface systems .Feedback detection, for instance, is used to build more immersive entertainment experiences, more natural and educational systems, and more in customer service experiences, everywhere feedback is used. One of the main advantages of feedback analyzers using face expression by deep neural networks is that they can provide real time feedback. This is in way of natural feedback collection methods, such as surveys and interviews, which can be time consuming and expensive to administer. Real time feedback analyzer allows system developers to make rapid changes to their system in order to improve and develop the user experience. we have leveraged a custom trained CNN model, to accurately classify the seven different emotions of an human face that are availed in the dataset. The model has achieved an impressive accuracy of 83%. This model can be used in various applications. Index Terms—Convolutional Neural Networks, facial expres- sion, emotion, feedback, Resnet.

Convolutional neural networks for signal identification and extraction and seabed classification

Sound underwater source (SUS) charges can be used for seabed characterization experiments. Many SUS were deployed during experiments in the New England MudPatch during 2017 and 2022. The goal of this work is to automatically detect and extract SUS signals in the dataset and then perform seabed classification on the extracted SUS signals. A binary classifier CNN is trained on simulated SUS charges and measured ambient noise to detect if a SUS signal is present in one minute pressure waveforms. The trained CNN model is then used to identify and extract the SUS signals from the measured data on the 52-channel PROTEUS L-array deployed by Applied Research Laboratories, University of Texas at Austin. This automated extraction process expedites the identification and time alignment of hundreds of SUS signals. The extracted signals will then be input to a ResNet-18 network trained on synthetic signals to perform seabed classification using a catalog of 34 seabeds. Lessons learned from the automated identification and extraction process will be presented as well as a statistical analysis of the seabed classification results. [Work supported by the Office of Naval Research, Grant N00014-22-12402.]

Integrated convolutional neural networks for joint super-resolution and classification of radar images

Deep learning techniques have been widely used for two-dimensional (2D) and three-dimensional (3D) computer vision problems, such as object detection, super-resolution (SR) and classification to name a few. Radar images suffer from poor resolution as compared to optical images, hence developing a high-accuracy model to solve computer vision problems, such as a classifier, is a challenge. This is because of the lack of high-frequency details in the input images which makes it difficult for the classifier model to generate accurate predictions. Ways of addressing this challenge include training the learning model with a large dataset or using a more complicated model, such as deeper layer architecture. However, employing such approaches might result in the overfitting of the model, where the model might not generalize well on previously unseen data. Also, generating a large dataset for training the model is a challenging task, especially in the case of radar images. An alternate solution for achieving high accuracy in radar classification problems is provided in this paper wherein a CNN-enabled super-resolution (SR) model is integrated with the classifier model. The SR model is trained to generate high-resolution (HR) millimeter-wave (mmW) images from any input low-resolution (LR) mmW images. These resolved images from the SR model will be used by the classifier model to classify the input images into appropriate classes, consisting of threat and non-threat objects. The training data for the dual CNN layers are generated using a numerical model of a near-field coded-aperture computational imaging (CI) system. This trained dual CNN model is tested with simulated data generated from the CI numerical model wherein a high classification accuracy of 95% and a fast inference time of 0.193 s are obtained, making it suitable for real-time automated threat classification applications. For fair comparison, the developed CNN model is also validated with experimentally generated reconstruction data, in which case, a classification accuracy of 94% is obtained.

Segmentation of hard exudate lesions in color fundus image using two-stage CNN-based methods

The presence of hard exudate (EX) lesions is an early clinical symptom of Diabetic Retinopathy (DR); its accurate segmentation is essential for diagnosis and treatment. Automatic segmentation of EX lesions is challenging because they have a wide range of sizes, shapes, and brightness, and can be confused with soft exudate lesions. In this paper, we present and assess an efficient segmentation method for interactively segmenting EX lesions in fundus images from DR patients. Our approach consisted of two main stages: (1) the first stage generates an automatic segmentation using a CNN-based method. (2) If the obtained segmentation is suboptimal, the second stage is used to refine each inaccurately segmented region using a CNN-based interactive segmentation method. To train and evaluate method performance, we used two public datasets (IDRiD and DDR) and two local datasets from two medical centers (EHos and EWT) containing a wide range of EX lesions. Several state-of-the art CNN-based models and the proposed method were trained and evaluated on IDRiD, DDR and EHos datasets using the area under the precision–recall curve (AUPR), the dice similarity coefficient (Dice) and the intersection over union (IoU). In addition, EWT dataset was used to evaluate the methods without retraining the CNN models. The evaluation results using IDRiD dataset showed that the proposed method achieved AUPR, mean Dice and mean IoU scores of 0.893, 76.6% and 62.4%, respectively, which are comparable to the state-of-the-art method on EX segmentation. When evaluated on the DDR dataset with large portion of small EX lesions, the proposed method achieved AUPR, mean Dice and mean IoU scores of 0.693, 60.8% and 45.5%, yielding state-of-the-art performance. Furthermore, with the main evaluation metric being AUPR, the evaluation results on two datasets, EHos and ETW, with scores of 0.689 and 0.667 demonstrated superiority of the proposed method over the other segmentation methods. The proposed two-stage CNN-based segmentation method is an useful approach for accurately segmenting EX lesions in color fundus images, enabling lesion quantification to aid experts in treating patients with DR. The source code, the trained CNN models and the data are publicly available at: https://github.com/dvquang2000/Interactive_HardExudates_Segmentation.

Deep learning-based identification of sinoatrial node-like pacemaker cells from SHOX2/HCN4 double-positive cells differentiated from human iPS cells.

Cardiomyocytes derived from human iPS cells (hiPSCs) include cells showing SAN- and non-SAN-type spontaneous APs. To examine whether the deep learning technology could identify hiPSC-derived SAN-like cells showing SAN-type-APs by their shape. We acquired phase-contrast images for hiPSC-derived SHOX2/HCN4 double-positive SAN-like and non-SAN-like cells and made a VGG16-based CNN model to classify an input image as SAN-like or non-SAN-like cell, compared to human discriminability. All parameter values such as accuracy, recall, specificity, and precision obtained from the trained CNN model were higher than those of human classification. Deep learning technology could identify hiPSC-derived SAN-like cells with considerable accuracy.

Automatically Labeling Offensive Formations in American Football Film Using Deep Learning

Web services for storing, annotating, and sharing sports videos by high-school athletic teams have become more prevalent in recent years. However, most services lack the ability for coaches to automatically tag film, leading to many hours of manual annotation. For American football videos, coaches need to label formations, plays, and field positions in order to extract insights and create strategic game plans. This paper presents an end-to-end machine learning pipeline for automatically labeling American football offensive formations in videos. The pipeline includes pre-processing of videos, image classification, and a novel inference approach. The study compares a custom CNN model with pre-trained image classifier models using transfer learning. Specifically, CNN-based architectures (MobileNet, Inception, EfficientNet, etc.) and a transformer-based Vision Transformer (ViT) are compared. All models are trained on ~1400 images with the three most popular formation labels extracted from video clips of high-school football team games. The results show that several models, including the custom CNN model, achieved greater than 90% classification accuracy on the test dataset. The inference is performed by sampling multiple frames from a video clip, passing them through a trained image classifier, and taking a majority vote on the classification results to determine the final outcome. Our study found that using a sampling rate of 0.5 seconds, starting at 1 second, and taking five frames yields the highest inference accuracy of 95.4% using a trained customized CNN model. This system can assist all levels of football coaches in the analysis of game footage and formation identification.

Weights analysis for CNN models in MRI classification

AbstractBackgroundIn deep neural network learnable parameters like weights and bias of convolution kernel, normalization layer or fully connected layers (FCLs) are the one responsible for giving prediction score. Though one parameter or parameters in a layer has not decisive role, however when whole network prediction is to be done, each one has role to generate prediction value. Hence, they are important for the network performance.MethodWe modeled a CNN for classification and trained it using 50% of total MRI scans for 3 classes. After training the network for 50 epochs, the network reaches convergences with 100% training accuracy and around 70% validation accuracy. Then, the trained models’ parameters i.e., weights and bias of FCL layers were analyzed class wise for any correlation with its parent class.ResultThe weights of final FCL are plotted as in Figure 1. The correlation matrix for weights on sample MRIs with trained weights are shown in Tables 1 and 2. Here, two activation functions were used for final test results were ReLU and Leaky ReLU with final test accuracy around 67.3% and 70.9% respectively.ConclusionWe attempted to study the weights pattern in FCL layer for a trained CNN model along with its correlation value for each class.[Acknowledgement]This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. NRF‐2021R1I1A3050703). And this research was supported by the BrainKorea21Four Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (4299990114316).

Breast tumor detection by 1D-convolutional neural network based on ultra-wide-band microwave technology

Ultra-wide band (UWB) microwave is a promising technology for non-invasive detection of breast tumors due to its low cost and the absence of ionizing radiation. However, breast heterogeneity is still the main technical barrier of microwave imaging at present. It is challenging to identify whether there are tumor responses from the detected microwave signals because they contain many high-magnitude skin and fibroglandular scattering signals. To this end, a 1D-convolution neural network (CNN) is proposed to detect the presence of breast tumors directly from raw time-domain microwave signals, avoiding complex microwave image reconstruction and time-consuming feature engineering. The designed network has 10 weight layers; it is composed of a stack of convolution-rectified linear unit-batch normalization layers, and every two stacks are followed by a maxpooling layer. At the end of the network are two fully connected layers to complete the detection task. With the proposed network, the detection accuracies on two datasets simulated by an electromagnetic solver reach 98.20% and 95.97%. The accuracy of experimental verification based on the practical UWB detection system remains good, reaching 94.66%. Finally, through the t-distributed stochastic neighbor embedding technique, the results of feature visualization prove the effectiveness of the trained CNN model. The evaluation results indicate the superior performance of the proposed approach for UWB microwave-based breast tumor detection.

An end to end trained hybrid CNN model for multi-object tracking

An end to end trained hybrid CNN model for multi-object tracking

MR Image Based Brain Tumor Classification with Deep Learning Neural Networks

The unique combination of Artificial Intelligence and Machine Learning, which helps the computer to imitate the ways and behaviour of human beings can be termed as deep learning. The field of deep learning is an emerging field that has gained a lot of interest toward past years. The Deep Learning have proven already to solve the complex problem using the powerful machine learning tools. One of the best deep learning algorithm is used to classify the brain tumor data set in this paper. The deep learning architecture is able to classify the brain tumor into 4 categories of images. The first being no tumor, the second being pituitary tumor, the third is meningioma and the last one classified as glioma. As we are well aware, the training datasets for the medical imaging scenario are very few. This is a challenging task to apply the deep learning that is obtained from a trained CNN model to dig up the small data set to attain the result. A pre trained CNN model is used here to solve the problem. The obtained results are good over all Performance is measured.

CT2US: Cross-modal transfer learning for kidney segmentation in ultrasound images with synthesized data

Accurate segmentation of kidney in ultrasound images is a vital procedure in clinical diagnosis and interventional operation. In recent years, deep learning technology has demonstrated promising prospects in medical image analysis. However, due to the inherent problems of ultrasound images, data with annotations are scarce and arduous to acquire, hampering the application of data-hungry deep learning methods. In this paper, we propose cross-modal transfer learning from computerized tomography (CT) to ultrasound (US) by leveraging annotated data in the CT modality. In particular, we adopt cycle generative adversarial network (CycleGAN) to synthesize US images from CT data and construct a transition dataset to mitigate the immense domain discrepancy between US and CT. Mainstream convolutional neural networks such as U-Net, U-Res, PSPNet, and DeepLab v3+ are pretrained on the transition dataset and then transferred to real US images. We first trained CNN models on a data set composed of 50 ultrasound images and validated them on a validation set composed of 30 ultrasound images. In addition, we selected 82 ultrasound images from another hospital to construct a cross-site data set to verify the generalization performance of the models. The experimental results show that with our proposed transfer learning strategy, the segmentation accuracy in dice similarity coefficient (DSC) reaches 0.853 for U-Net, 0.850 for U-Res, 0.826 for PSPNet and 0.827 for DeepLab v3+ on the cross-site test set. Compared with training from scratch, the accuracy improvement was 0.127, 0.097, 0.105 and 0.036 respectively. Our transfer learning strategy effectively improves the accuracy and generalization ability of ultrasound image segmentation model with limited training data.

Strawberry Fungal Leaf Scorch Disease Identification in Real-Time Strawberry Field Using Deep Learning Architectures

Plant health is the basis of agricultural development. Plant diseases are a major factor for crop losses in agriculture. Plant diseases are difficult to diagnose correctly, and the manual disease diagnosis process is time consuming. For this reason, it is highly desirable to automatically identify the diseases in strawberry plants to prevent loss of crop quality. Deep learning (DL) has recently gained popularity in image classification and identification due to its high accuracy and fast learning. In this research, deep learning models were used to identify the leaf scorch disease in strawberry plants. Four convolutional neural networks (SqueezeNet, EfficientNet-B3, VGG-16 and AlexNet) CNN models were trained and tested for the classification of healthy and leaf scorch disease infected plants. The performance accuracy of EfficientNet-B3 and VGG-16 was higher for the initial and severe stage of leaf scorch disease identification as compared to AlexNet and SqueezeNet. It was also observed that the severe disease (leaf scorch) stage was correctly classified more often than the initial stage of the disease. All the trained CNN models were integrated with a machine vision system for real-time image acquisition under two different lighting situations (natural and controlled) and identification of leaf scorch disease in strawberry plants. The field experiment results with controlled lightening arrangements, showed that the model EfficientNet-B3 achieved the highest classification accuracy, with 0.80 and 0.86 for initial and severe disease stages, respectively, in real-time. AlexNet achieved slightly lower validation accuracy (0.72, 0.79) in comparison with VGGNet and EfficientNet-B3. Experimental results stated that trained CNN models could be used in conjunction with variable rate agrochemical spraying systems, which will help farmers to reduce agrochemical use, crop input costs and environmental contamination.

FLOPs-efficient filter pruning via transfer scale for neural network acceleration

Model pruning is a useful technique to reduce the computational cost of convolutional neural networks. In this paper, we first propose a simple but effective filter level pruning criterion, which assesses the importance of a filter by exploring the transfer scale (TS) of its feature maps in the next layer. The principle is that for a trained CNN model, an important filter should have strong connections with the next layer, otherwise the transfer scale of its feature map will be low and hence removing it will have little influence on the network. Besides, we observe that filters from the computationally-intensive layers are more sensitive to pruning, which makes it difficult to further compress the floating-point operations (FLOPs) of the model without reducing accuracy. To solve this problem, we propose a FLOPs-efficient group Lasso approach for TS to guide the network to use fewer filters in the computationally-intensive layers, which leads to better FLOPs compression performance after pruning. We refer to the proposed method as FETS. Compared with the state-of-the-art methods, our FETS achieves similar or better accuracy, but with significantly larger FLOPs compression ratio. In particular, with VGG-16, ResNet-56 and DenseNet-40 on CIFAR-10, we achieve similar or better accuracies than other methods, with only 48%, 64% and 58% of the FLOPs. With ResNet-50 on ImageNet, we also achieve a relative FLOPs reduction of 30%.

Lung Disease Detection Using CNN

Covid Pneumonia is a life-threatening bacterial disease in humans that affects one or both lungs and is caused by the bacteria Streptococcus pneumonia. Also known as Covid-19, this is a respiratory illness that was first discovered in Wuhan, China. Expert radiotherapists must evaluate chest X-rays used to diagnose pneumonia. As a result, establishing an autonomous system for detecting pneumonia would be advantageous for treating the condition quickly, especially in distant places. The statistical results show that using pre trained CNN models and supervised classifier algorithms to analyse chest X-ray pictures, specifically to diagnose Pneumonia, can be highly advantageous. By constructing certain convolution neural network designs, we are developing a classifier model that accurately predicts if a person has covid or pneumonia.

Combined Improved Dirichlet Models and Deep Learning Models for Road Extraction from Remote Sensing Images

Combining Dirichlet Mixture Models (DMM) with deep learning models for road extraction is an attractive study topic. Benefiting from DMM, the manually labeling work is alleviated. However, DMM suffers from high computational complexity due to pixel by pixel computations. Also, traditional constant parameter settings of DMM may not be suitable for different target images. To address the above problems, we propose an improved DMM which embeds superpixel strategy and sparse representation into DMM. In our road extraction framework, we first use improved DMM to filter out most backgrounds. Then, a trained deep CNN model is used for further precise road area recognition. To further promote the processing speed, we also apply the superpixel scanning strategy for CNN models. We tested our method on a Shaoshan dataset and proved that our method not only can achieve better results than other compared state-of-the-art image segmentation methods, but the processing speed and accuracy of DMM are also improved.

Exploring the Impact of Data Augmentation on Deep Learning Models for Plant Disease Detection in PlantVillage Dataset

Plant diseases are a vital part of ensuring the safety of agricultural crops and the protection of food security. With the development of deep learning techniques, it has been able to automate the detection of these diseases. Unfortunately, one of the main challenges that researchers face when it comes to training models for this field is the lack of labeled training data. The study analyzed the effects of the data augmentation on the deep learning models' performance when it came to identifying plant diseases in the PlantVillage dataset. Here focused on two common ailments affecting tomato and potato plants. To address this issue, we trained CNN models on the PlantVillage dataset using different augmentation techniques, such as cropping, rotation, scaling, and flipping. The results of the study were evaluated using various metrics, including accuracy, recall, area under the curve, and F1 score. Data augmentation can improve a deep learning model's performance when it comes to detecting plant diseases. Our study revealed that CNN models, which were trained using the same methods, achieved higher accuracy and recall metrics when compared to the original models. Furthermore, our findings show that cropping techniques, which are commonly used in augmentation, were more effective at improving the models' performance. Even though the training data is limited, data augmentation can still help improve deep learning models' performance when it comes to detecting plant diseases. The use of such techniques could significantly enhance the models' accuracy and speed in identifying plant diseases, which are important for global food security and the agricultural industry.

Facial expression recognition with trade-offs between data augmentation and deep learning features

A novel facial expression recognition system has been proposed in this paper. The objective of this paper is to recognize the types of expressions in the human face region. The implementation of the proposed system has been divided into four components. In the first component, a region of interest as face detection has been performed from the captured input image. For extracting more distinctive and discriminant features, in the second component, a deep learning-based convolutional neural network architecture has been proposed to perform feature learning tasks for classification purposes to recognize the types of expressions. To enhance the performance of the proposed system, in the third component, some novel data augmentation techniques have been applied to the facial image to enrich the learning parameters of the proposed CNN model. In the fourth component, a trade-off between data augmentation and deep learning features have been performed for fine-tuning the trained CNN model. Extensive experimental results have been demonstrated using three benchmark databases: KDEF (seven expression classes), GENKI-4k (two expression classes), and CK+ (seven expression classes). The performance of the proposed system respect for each database has been well presented and described and finally, these performances have been compared with the existing state-of-the-art methods. The comparison with competing methods shows the superiority of the proposed system.

Computer aided diagnosis of obesity based on thermal imaging using various convolutional neural networks

ObjectivesThe study aims at the following: i) To construct a custom Deep Learning Network for classifying the thermal images of abdomen, forearm and shank regions into obese and normal cases ii) To compare the performance of the proposed CNN with some of the state-of-the-art pre-trained CNN and Machine Learning models in detection of obesity. MethodsFifty healthy subjects along with fifty other age cum sex matched obese subjects were included in the study. The mean skin surface temperature was measured in abdomen, shank and forearm region for normal and obese subjects. After data augmentation, the images are fed to the proposed CNN and pre-trained networks for training, validation and classification of normal and obese thermograms. ResultsAmong the ROI studied, the abdomen region exhibited a high temperature difference of 4.703 % between the normal and obese compared to other regions. The proposed custom network-2 provided an overall accuracy of 92 %, area under the curve (AUC) value of 0.948 whereas the pre-trained model VGG16 net produced an accuracy of 79 % and AUC value of 0.90 for discrimination into obese and normal thermograms. ConclusionsHence, the deep learning system based on custom CNN provided a reliable classification performance to identify the occurrence of obesity in test subjects. The experimental analysis showed that custom CNN network-2 provided a commendable degree of accuracy in classification of normal and obese subjects from the thermal images. Thus, the trained Custom-2 CNN model can be used for computer aided screening of test subjects for obesity detection.

Deep multi-label learning for image distortion identification

Image Distortion Identification is important for image processing system enhancement, image distortion correction and image quality assessment. Although images may suffer various number of distortions while going through different systems, most of the previous researches of image distortion identification were focus on identifying single distortion in image. In this paper, we proposed a CNN-based multi-label learning model (called MLLNet) to identify distortions for different scenarios, including images having no distortion, single distortion and multiple distortions. Concretely, we transform the multi-label classification for image distortion identification to a number of multi-class classifications and use a deep multi-task CNN model to train all associated classifiers simultaneously. For unseen image, we use the trained CNN model to predict a number of classifications at same time and fuse them to final multi-label classification. The extensive experiments demonstrate that the propose algorithm can achieve good performance on several databases. Moreover, the network architecture of the CNN model can make flexible adjustment according to the different requirements.