CNN Training Research Articles

Internet of things image recognition system based on deep learning

At present, many exciting results have been achieved in the application of deep learning to image recognition. However, there are still many problems to be overcome before deep learning is used in practical applications such as image retrieval, image annotation, and image-text conversion. This paper studies the structure of deep learning, improves the commonly used training algorithms, and proposes two new neural network models for different application scenarios. This paper uses Support Vector Machine (SVM) as the main classifier for Internet of Things image recognition and uses the database of this paper to train SVM and CNN. At the same time, the effectiveness of the two for image recognition is tested, and the trained classifier is used for image recognition. The result surface: In the labeled data set, the rank-1 accuracy of CNN is 85.77%, which is higher than 90.28% of the SVM method. In the detection data, CNN’s rank-1 accuracy rate is 83.11%, which also exceeds SVM’s 80.22%. SVM+CNN has a rank 1 value of 84.69% for the detection data set. This shows that deep learning can map the feature representation of the image and the feature representation of the word to the same space, making the calculation of the similarity and correlation between the image and the text easier and more straightforward.

Near Lossless Transfer Learning for Spiking Neural Networks

Spiking neural networks (SNNs) significantly reduce energy consumption by replacing weight multiplications with additions. This makes SNNs suitable for energy-constrained platforms. However, due to its discrete activation, training of SNNs remains a challenge. A popular approach is to first train an equivalent CNN using traditional backpropagation, and then transfer the weights to the intended SNN. Unfortunately, this often results in significant accuracy loss, especially in deeper networks. In this paper, we propose CQ training (Clamped and Quantized training), an SNN-compatible CNN training algorithm with clamp and quantization that achieves near-zero conversion accuracy loss. Essentially, CNN training in CQ training accounts for certain SNN characteristics. Using a 7 layer VGG-* and a 21 layer VGG-19, running on the CIFAR-10 dataset, we achieved 94.16% and 93.44% accuracy in the respective equivalent SNNs. It outperforms other existing comparable works that we know of. We also demonstrate the low-precision weight compatibility for the VGG-19 structure. Without retraining, an accuracy of 93.43% and 92.82% using quantized 9-bit and 8-bit weights, respectively, was achieved. The framework was developed in PyTorch and is publicly available.

Generating Building Infrastructure Modeling Using cGAN

In today’s world to propose a building model to an investor and builder, an architect has to undergo a full process of making a full 3D model of a building and detailing each and every aspect of the building. The details include the interior and exterior of the building. But, at the end the investor and builder are interested in the front elevation design of the building and not the detailed part. Hence, for providing the important information like faces of the building, we propose to generate new synthetic versions of images from sketch, labels, and edges. The model used for developing this platform i.e., pix2pix is an approach for training CNN for image-to-image translations. Pix2pix can be demonstrated on various image translation tasks; transforming an image from one domain to another or to a corresponding output could have innumerable applications, changing art styles, building infrastructure modelling and criminal identification are just a few applications. The building infrastructure modeling platform proposed aims to be feasible and have wide applicability for the architects by providing an interactive pallet and a convenient platform.

Can a CNN trained on the Ising model detect the phase transition of the q-state Potts model?

Abstract Employing a deep convolutional neural network (deep CNN) trained on spin configurations of the 2D Ising model and the temperatures, we examine whether the deep CNN can detect the phase transition of the 2D $q$-state Potts model. To this end, we generate binarized images of spin configurations of the $q$-state Potts model ($q\ge 3$) by replacing the spin variables $\{0,1,\ldots,\lfloor q/2\rfloor-1\}$ and $\{\lfloor q/2\rfloor,\ldots,q-1\}$ with $\{0\}$ and $\{1\}$, respectively. Then, we input these images to the trained CNN to output the predicted temperatures. The binarized images of the $q$-state Potts model are entirely different from Ising spin configurations, particularly at the transition temperature. Moreover, our CNN model is not trained on the information about whether phases are ordered/disordered but is naively trained by Ising spin configurations labeled with temperatures at which they are generated. Nevertheless, the deep CNN can detect the transition point with high accuracy, regardless of the type of transition. We also find that, in the high-temperature region, the CNN outputs the temperature based on the internal energy, whereas, in the low-temperature region, the output depends on the magnetization and possibly the internal energy as well. However, in the vicinity of the transition point, the CNN may use more general factors to detect the transition point.

Automatic prediction of COVID- 19 from chest images using modified ResNet50.

Recently coronavirus 2019 (COVID-2019), discovered in Wuhan city of China in December 2019 announced as world pandemic by the World Health Organization (WHO). It has catastrophic impacts on daily lives, public health, and the global economy. The detection of coronavirus (COVID− 19) is now a critical task for medical specialists. Laboratory methods for detecting the virus such as Polymerase Chain Reaction, antigens, and antibodies have pros and cons represented in time required to obtain results, accuracy, cost and suitability of the test to phase of infection. The need for accurate, fast, and cheap auxiliary diagnostic tools has become a necessity as there are no accurate automated toolkits available. Other medical investigations such as chest X−ray and Computerized Tomography scans are imaging techniques that play an important role in the diagnosis of COVID− 19 virus. Application of advanced artificial intelligence techniques for processing radiological imaging can be helpful for the accurate detection of this virus. However, Due to the small dataset available for COVID− 19, transfer learning from pre−trained convolution neural networks, CNNs can be used as a promising solution for diagnosis of coronavirus. Transfer learning becomes an effective mechanism by transferring knowledge from generic object recognition tasks to domain-specific tasks. Hence, the main contribution of this paper is to exploit the pre−trained deep learning CNN architectures as a cornerstone to enhance and build up an automated tool for detection and diagnosis of COVID− 19 in chest X−Ray and Computerized Tomography images. The main idea is to make use of their convolutional neural network structure and its learned weights on large datasets such as ImageNet. Moreover, a modification to ResNet50 is proposed to classify the patients as COVID infected or not. This modification includes adding three new layers, named, ′Conv′, ′Batch_Normaliz′ and ′Activation_Relu′ layers. These layers are injected in the ResNet50 architecture for accurate discrimination and robust feature extraction. Extensive experiments are carried out to assess the performance of the proposed model on COVID− 19 chest X−Ray and Computerized Tomography scan images. Experimental results approve that the proposed modification, injected layers, increases the diagnosis accuracy to 97.7% for Computerized Tomography dataset and 97.1% for X−Ray dataset which is superior compared to other approaches.

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.

Analysis on Electrocardiogram Measurement Performance Using Metal Electrode Based on 1-D Convolutional Neural Network for Application of Wearable Device

Disposable electrode is mainly used when measuring ECG signals. However, disposable electrode has low re-usability and durability, making it difficult to obtain high-quality signals when used for a long term. In this paper, we evaluated the applicability of metal plates and conductive tapes as new reference electrode to replace disposable electrode. For evaluation, we implemented a 1D CNN-based training model and an ECG measurement system to collect data. The implemented neural network was trained based on the MIT-BIH arrhythmia data set, and 5 classes of ECG classification including normal signals are possible. For measurement evaluation of metal plates, conductive tapes, and disposable electrode, field tests were conducted using the trained model. The ECG class classification of the implemented CNN training model confirmed a classification accuracy of more than 96%, and compared with previously studied pattern matching algorithms and other studies, confirmed excellent detection performance. In the field test, performance evaluation of arrhythmia classification was performed using a neural network trained on ECG signals obtained with disposable electrode, metal plates, and conductive tape. The ECG signals for evaluation used in the experiment, 300 samples were measured for each class from 10 subjects. As a result of the experiment, it was confirmed that the disposable electrode was classified as 98.59%, the metal plate was classified as 95.26%, and the conductive tape was classified as 96.13%. It is inconvenient to measure ECG using disposable electrode in everyday life. However, there is no difference in ECG detection performance in the case of a metal plate and a conductive tape compared to a disposable electrode. In addition, metal plates and conductive tapes can be easily applied to existing wearables because they can be reused and long term.

Implementation of Photoplethysmography Measurement System Based on Convolution Neural Network for Personalized Exercise Intensity

With increasing interest in health, many people are exercising to lose weight, prevent disease, and improve cardiorespiratory function. For effective exercise, users should proceed with appropriate intensity depending on their physical strength. The system implemented in this paper classifies exercise intensity according to PPG signal using CNN training model for objective exercise intensity classification. The PPG signal was measured after exercise through the PPG sensor, and the training data set was constructed that based on the interval between P-peaks. The training data set is trained on the CNN model to classify the three exercise intensity according to the PPG signal. In order to analyze the accuracy of the implemented CNN training model, the performance evaluation of the classification evaluation metrics and the exercise intensity classification monitoring system was performed. First, the performance evaluation of the CNN model for classifying exercise intensity was conducted. In the performance evaluation, the classification evaluation metrics was calculated according to the training result, the recall rate representing the percentage of successful prediction among the actual correct answers, the precision representing the actual correct answer rate among the predicted data, and the F 1 score representing the harmonic average of recall and precision were confirmed. As a result of CNN training model classification evaluation metrics, it was the accuracy was 99.3%, the recall rate was 99.9%, the precision was 99.8%, and the F 1 score was 99.4%. Second, to evaluate the performance of the exercise intensity classification monitoring system, jump rope experiment was conducted with 5 subjects. The experiment measured PPG at the end of each set after low, moderate, and high intensity jump rope. The classification accuracy was analyzed by entering the measured PPG data into the CNN model 50 times each. As a result of the experiment, the accuracy of low intensity was 98%, moderate intensity was 93.6%, and high intensity was 97.6%, confirming a total accuracy of 96.4%. Some errors are thought to have occurred due to the fact that the data located at the boundary line between the exercise intensity was classified incorrectly. In future studies, we would like to conduct a study of exercise intensity monitoring system that can be applied to various exercises by measuring acceleration signals for each exercise together.

Traffic thermal infrared texture generation based on siamese semantic CycleGAN

Thermal infrared texture generation is a promising infrared imaging framework for various applications. A novel thermal infrared texture generation algorithm, based on siamese semantic CycleGAN (SS-CycleGAN), is proposed for thermal infrared systems. Different from traditional infrared simulation frameworks, SS-CycleGAN depends on no extra environmental information, such as air temperature, humidity and radiation properties of objects. In other words, visible images could be directly transformed into thermal infrared images like using style transfer algorithms, after traffic scene has been fully understood through training CNN. In this paper, style transfer is firstly introduced for generating thermal textures from color visible images. Siamese semantic loss for visible-infrared transformation is designed and introduced to generate object-oriented thermal infrared textures, while maintaining high definition. Compared to other style transfer algorithms, SS-CycleGAN could generate reasonable thermal infrared textures with clear edge details, in traffic scenes.

Design and Implementation of Missing Data Classification Technique for IoT Applications Using Artificial Intelligence

The combination of various sensors with different data methods is a common technique used to increase precision in the classification of IoT health data. However, for even the assessment outcomes, all modalities are barely available and this scarcity of evidence poses significant barriers to multimodal education. Driven by recent developments in deep education, we are providing a cross-neural network for the segmentation of the IoT Health Data Classification, which is trained on data modalities not all available during trials. In IoT Health Data Classification, we train our architecture with a cost function that is especially tailored to unbalanced classes. We are providing the device with a benchmark data set with incomplete data. Assuming that they are not present in the research process, our methodology goes beyond both the CNN training and the collection of two CNNs trained in the missing modality by utilising time data

A new CNN training approach with application to hyperspectral image classification

Three main requirements of a successful application of deep learning are the network architecture, a large enough training dataset, and a good optimization algorithm. In this paper we mainly focus on the optimization part. We propose a training algorithm for convolutional neural networks which makes use of both first and second order derivatives for training different layers. We utilize an approximate second order algorithm for the classification layer while we train the rest of the network with the conventional approach which is backpropagation with first order derivatives. We show that this approach helps us achieve a higher classification accuracy with a much smaller number of training iterations compared to training the whole network with gradient descent based algorithms. Moreover, although second order optimization is generally costlier, we show that the proposed approach is trained faster not only in terms of the number of iterations but also training duration. We also present the integration of CNNs with a probabilistic spatial model and apply this to the land cover classification problem in hyperspectral images. The results show that the algorithm allows us to achieve superior results with a simple network even with limited training data compared to existing approaches.

Automatic Detection of COVID-19 Disease by Using Transfer Learning of Light Weight Deep Learning Model

COVID-19 began to appear in China in December 2019 with cases of pneumonia. Controlling the effects of this virus type is closely related to all scientists, especially medical doctors. In this study, ShuffleNet CNN architecture, which is suitable for mobile devices, is one of the deep learning models that perform today's most successful image recognition and classification processes. Obtaining a large data set in the training process is an important problem for the diagnosis of COVID-19 disease. However, rapid diagnosis of COVID-19 X-ray or CT images plays a crucial role in the immediate decision of medical doctors. However, the lack of training data causes poor performance due to the problem of over-fitting. Also, the back propagation algorithm used in CNN training is often very slow and requires the tuning of different hyper parameters. To overcome these disadvantages, this study proposed a new approach based entirely on the machine learning algorithms, extracting CNN features from the light weight ShuffleNet architecture. In this study, a total of 300 images were used, the images diagnosed as COVID-19 and normal chest images. During the classification phase, 10-fold cross validation was applied to the database and 99.98% accuracy rate was obtained.

PLANT DISEASE DETECTION USING CONVOLUTIONAL NEURAL NETWORK

When plants and crops are suffering from pests it affects the agricultural production of the country. Usually, farmers or experts observe the plants with eye for detection and identification of disease. But this method is often time processing, expensive and inaccurate. Automatic detection using image processing techniques provide fast and accurate results. This paper cares with a replacement approach to the development of disease recognition model, supported leaf image classification, by the utilization of deep convolutional networks. Advances in computer vision present a chance to expand and enhance the practice of precise plant protection and extend the market of computer vision applications within the field of precision agriculture. a completely unique way of training and therefore the methodology used facilitate a fast and straightforward system implementation in practice. All essential steps required for implementing this disease recognition model are fully described throughout the paper, starting from gathering images to make a database, assessed by agricultural experts, a deep learning framework to perform the deep CNN training. This method paper may be a new approach in detecting plant diseases using the deep convolutional neural network trained and finetuned to suit accurately to the database of a plants leaves that was gathered independently for diverse plant diseases. The advance and novelty of the developed model dwell its simplicity healthy leaves and background images are in line with other classes, enabling the model to distinguish between diseased leaves and healthy ones or from the environment by using CNN. Plants are the source of food on earth. Infections and diseases in plants are therefore a big threat, while the foremost common diagnosis is primarily performed by examining the plant body for the presence of visual symptoms [1]. As an alternative to the traditionally time-consuming process, different research works plan to find feasible approaches towards protecting plants. In recent years, growth in technology has engendered several alternatives to traditional arduous methods [2]. Deep learning techniques are very successful in image classification problems.

A Novel Method for CNN Training Using Existing Color Datasets for Classifying Hand Postures in Bayer Images

Deep learning has revolutionized the computer vision by achieving human level performance on classification and detection of the objects in the images. These algorithms are often trained using supervised learning methods which require a labeled training data set. Since the labeling requires human intervention, the human interpretable tri-color images are needed for this purpose even though the image sensors capture subsampled images in Bayer format. For inference also, the Bayer format images are converted to tri-color images which require sophisticated image processing pipeline. This conversion seems unnecessary given the proven ability of the deep networks to learn in multiple different scenarios. Moreover, the image processing pipeline can be simplified by directly using the Bayer images for computer vision tasks. Availability of the dataset for training the deep networks remains the challenge for building such systems; so, a methodology is needed to use the color images for training DNNs which can classify the Bayer images for a given application. The existing tri-color images from training datasets can be subsampled in pseudo-Bayer pattern for training such deep neural networks. This paper presents the methodology and results of one such method adopted for hand pose recognition application using SqueezeNet with encouraging performance. Using this methodology, we achieved 100% accuracy in nine out of ten hand postures with pseudo-Bayer images and in six out of ten hand postures with real Bayer images.

Shadow Removal by a Lightness-Guided Network With Training on Unpaired Data

Shadow removal can significantly improve the image visual quality and has many applications in computer vision. Deep learning methods based on CNNs have become the most effective approach for shadow removal by training on either paired data, where both the shadow and underlying shadow-free versions of an image are known, or unpaired data, where shadow and shadow-free training images are totally different with no correspondence. In practice, CNN training on unpaired data is more preferred given the easiness of training data collection. In this paper, we present a new Lightness-Guided Shadow Removal Network (LG-ShadowNet) for shadow removal by training on unpaired data. In this method, we first train a CNN module to compensate for the lightness and then train a second CNN module with the guidance of lightness information from the first CNN module for final shadow removal. We also introduce a loss function to further utilise the colour prior of existing data. Extensive experiments on widely used ISTD, adjusted ISTD and USR datasets demonstrate that the proposed method outperforms the state-of-the-art methods with training on unpaired data.

Joint Face Retrieval System Based On a New Quadruplet Network in Videos of Multi-Camera

At present, a large number of off-line videos is stored in the server of surveillance network. In order to retrieve the target face in these massive videos frames, the face retrieval system is designed. A new Quadruplet Network is constructed by changing the RELU structure of CNN network and training the new Quadruplet Network to acquire the depth features. Join with the online fugitive face picture that launched online to initiate the wanted, with the help of the depth feature contrast to launch the Content-Based Image Retrieval (CBIR). The new Quadruplet Network converges faster than familiar networks such as Alexnet, Googlenet, VGGNet and ResNet. Because of the shared weight design of the network, the retrieval has a high precision, recall and the retrieval rate. Image depth features can be shared quickly online between the cameras. The experimental results show that the proposed method is effective, with an accuracy of 98.74% and a precision of 99.54%, and a frame rate of 28 FPS.

Transfer learning-based and originally-designed CNNs for robotic pick and place operation

The authors have developed a CNN and SVM design and training application for defect detection, and the effectiveness and the usefulness have been proved through several design, training and classif...

ReHy: A ReRAM-based Digital/Analog Hybrid PIM Architecture for Accelerating CNN Training

ReHy: A ReRAM-based Digital/Analog Hybrid PIM Architecture for Accelerating CNN Training

Image-Based TF Colorization With CNN for Direct Volume Rendering

In the direct volume rendering (DVR), it often takes a long time for a novice to manipulate the transfer function (TF) and analyze the volume data. To lessen the difficulty in volume rendering, several researchers have developed deep learning techniques. However, the existing techniques are not easy to apply directly to existing DVR pipelines. In this study, we propose an image-based TF colorization with CNN to automatically generate a direct volume rendering image (DVRI) similar to a target image. Our system includes CNN model training, TF labeling, image-based TF generation, and volume rendering by matching the target image. We introduce a technique for training CNN and labeling the TF with images similar to the input volume dataset. Moreover, we extract the primary colors from the target image according to the labels classified with the CNN model. We render the volume data with the TF to produce the DVRI reproducing the prominent colors in the target image.

Aging Assessment of Oil–Paper Insulation Based on Visional Recognition of the Dimensional Expanded Raman Spectra

Detecting insulation performance through testing the Raman spectroscopy of transformer oil is noninvasive and shows great promise in on-site diagnosis. The information, stored in spectra, that may reflect aging states of insulation aims to be effectively unfolded and provides decision support as well. In this article, a transformation method based on intrinsic product arithmetic is used to transform the 1-D Raman spectral data into 2-D image data. With the help of CNN based on the dimensionality expanding method, diagnosis of oil–paper insulation with different aging states based on the Raman spectroscopy was realized. The network-based oil–paper insulation aging diagnosis model has an accuracy of 95% in ten aging categories of 60 test samples and 100% in four aging categories. The result shows the trained CNN was able to identify the differences of transformer oil samples at each aging stage through visualized Raman spectral image. This method provides a more accurate determination method for diagnosing oil-immersed transformers’ aging conditions; it also develops an available method for a more detailed evaluation of insulation aging levels.