Complexity Of Convolutional Neural Networks Research Articles

Automated Segmentation and Classification of Aerial Forest Imagery

Monitoring the health and safety of forests has become a rising problem with the advent of global wildfires, rampant logging, and reforestation efforts. This paper proposes a model for the automatic segmentation and classification of aerial forest imagery. The model is based on U-net architecture and relies on dice coefficients, binary cross-entropy, and accuracy as loss functions. While models without autoencoder-based structures can only reach a dice coefficient of 45%, the proposed model can achieve a dice coefficient of 79.85%. In addition, for barren adn dense forestry image classification, the proposed model can achieve 82.51%. This paper demonstrates how complex convolutional neural networks can be applied to aerial forest images to help preserve and save the forest environment.

Robust Force Estimation for Magnetorheological Damper Based on Complex Value Convolutional Neural Network

AbstractRecent developments in semi-active control technologies enhance the possibility of an effective response reduction during a wide range of dynamic loading conditions. Most semi-active control schemes employ magnetorheological dampers (MRDs) as actuators. These devices exhibit nonlinear and hysterical behavior that complicates reactive force estimation to compensate for disturbances. In this paper, we present a novel robust schema to estimate MRD forces using a complex value convolutional neural network (CV-CNN) to overcome these problems. CV-CNN utilizes random complex value convolutional filters as parameters to reduce the measured noise by combining the training stage and the max-by-magnitude operation. Furthermore, CV-CNN is a hysteresis-model-free strategy that overcomes the parameterization in nonlinear systems. The proposed CV-CNN only requires displacement and voltage measurements for force estimation. Different metrics are used to compare results between the CV-CNN, genetic algorithm (GA), particle swarm optimization (PSO), and shallow neural network (SNN). Experimental results show the potential of the proposed CV- CNN for practical applications due to its simplicity and robustness. The CV-CNN computational time is less than that of GA and PSO. In the training stage, the CV-CNN uses 0.7% of GA's time and 1.4% of PSO. Although SNN uses 5.5% of the time consumed by the CV-CNN, the latter performs the force estimation for MRD better; its mean square error is 78.3% lower than the GA's and PSO's, and 71.4% lower than SNN's.

MI-EEG classification using Shannon complex wavelet and convolutional neural networks

Many classification methods by machine learning and convolutional neural networks (CNN) have been proposed to recognize MI-EEG recently. However, the indescribable properties and individual differences of the MI-EEG signals cause low classification accuracy. In this study, a new MI-EEG classification method was designed to improve classification accuracy by combining Shannon complex wavelets and convolutional neural networks. First, the original MI-EEG was preprocessed using EEGLAB by channel selection and bandpass filtering. Second, the Shannon complex wavelet was used as the time–frequency transform strategy to calculate the time–frequency​ matrix. Finally, an improved Resnet was used to classify the time–frequency​ matrix to complete the MI-EEG identification. BCI competition IV dataset 2b as a public motor imagination dataset was tested to prove the validation of this proposed method. The classification accuracy and kappa value were adopted to prove the superiority of the proposed method by comparing it with the state-of-the-art classification methods. Experimental results showed that the classification accuracy and kappa values are 0.852 and 0.704, respectively, and they are the highest in the state-of-the-art. The parameter influence of wavelet wavelength and interception time on classification accuracy was discussed and optimized. This method can effectively improve classification accuracy and has a wide range of applications in MI-EEG classification.

RoofSplit: An edge computing framework with heterogeneous nodes collaboration considering optimal CNN model splitting

Intelligent applications based on AI have put much challenge on the edge of wireless networks, considering the heterogeneous characteristics and insufficient resources of the edge nodes. In this paper, we propose an edge computing framework named ’RoofSplit’ with the integration of communication, caching and computing ability. With this framework, deep learning models such as convolutional neural network (CNN) can be executed at the network edge with high efficiency. We first utilize Roofline theory and cut the one special complex CNN to several optimal sub-models. And then deploy the sub models on different edge nodes and process them in parallel. We also build a hardware testbed with the GPU and caching optimization to adapt to the RoofSplit framework. Finally, we make the experiments on the testbed. The results show that, compared with non-split scheme, RoofSplit can increase throughput by 4%–15% and reduce memory usage by 40%–60%. By using Redis caching, database is deemed as message queue in internal communication. The delay of processing one learning model is 40%–70% lower than that of Cloud-Only scheme, and more than 10% lower than that of Edge–Cloud collaboration scheme, which also reflects good real-time performance of the RoofSplit framework.

Multisignal Modulation Classification Using Sliding Window Detection and Complex Convolutional Network in Frequency Domain

With the development of the Internet of Things (IoT), the IoT devices are increasing day by day, resulting in increasingly scarce spectrum resources. At the same time, many IoT devices are facing inevitable malicious attacks. The cognitive Radio-enabled IoT (CR-IoT) is proposed as an effective method for spectrum resource allocation and risk monitoring in the IoT. The signal detection and modulation recognition are the key technologies for CR-IoT, addressing the problem of multisignal detection and automatic modulation classification (AMC) is one of the prerequisites for realizing secure dynamic spectrum access. Based on sliding window and deep learning (DL), this study proposes a multisignal frequency domain detection and recognition method. The frequency spectrum of the time-domain overlapping signal is obtained through the fast Fourier transform (FFT), and the frequency spectrum is segmented based on the signal energy detection method. Finally a complex convolutional neural network (CNN) is constructed for the identification of signal spectrum information. The proposed method can recognize 264 time-domain aliasing and frequency-closed signals with an accuracy of 97.3% under the influence of −2 dB corresponding to the noise of the calibration signal. In addition, the proposed method eliminates the influence of bandwidth, which can effectively detect and recognize the signal types of each component in the frequency band. This method has wide applicability and provides an effective scheme for the IoT cognitive technology.

Development of Light-Weight Convolutional Neural Network Model to Diagnose Tuberculosis

Tuberculosis (TB) is a contagious chest infection. World Health Organization has introduced different TB control programs in various countries. For the diagnosis of TB, the doctors mostly recommend chest X-ray (CXR) because it is more cost-effective and less time-consuming than existing sputum tests and Tuberculosis Skin Tests (TST). As per the research, deep learning models are best for TB diagnosis, by using CXR rather than normal eye-sight-based traditional method. Since doctor’s eye-sight or his experience is prone to human error, therefore, to solve this problem many Convolutional Neural Network (CNN) based models are introduced. Some of these models have high computational costs, and better accuracy making them heavy model. Whereas, others have less computational costs and lower accuracy making them light-weight models. Such models are further modified by the researchers to be more appropriate for better TB diagnosis, termed as Transfer Learning (TL) technique. However, TL leads to complex CNN structure and high computational cost. The proposed model named as Light TBNET(L-TBNET), attempts to provide less computational costs and higher accuracy simultaneously, as compared to other models such as, ShuffleNet, ResNet-50, MobileNet v2, Inception, and DenseNet. Moreover, the proposed does not include TL technique. This is accomplished by combining standard convolutional layers as well as depth-wise separable convolutional layers resulting in a hybrid model. The accuracy of the proposed model is 96% with lesser computational cost. In this way, the model contributes in providing a light-weight CNN model with higher accuracy.

Unsupervised Domain Adaptive Person Re-Identification Method Based on Transformer

Person re-identification (ReID) is the problem of cross-camera target retrieval. The extraction of robust and discriminant features is the key factor in realizing the correct correlation of targets. A model based on convolutional neural networks (CNNs) can extract more robust image features. Still, it completes the extraction of images from local information to global information by continuously accumulating convolution layers. As a complex CNN, a vision transformer (ViT) captures global information from the beginning to extract more powerful features. This paper proposes an unsupervised domain adaptive person re-identification model (ViTReID) based on the vision transformer, taking the ViT model trained on ImageNet as the pre-training weight and a transformer encoder as the feature extraction network, which makes up for some defects of the CNN model. At the same time, the combined loss function of cross-entropy and triplet loss function combined with the center loss function is used to optimize the network; the person’s head is evaluated and trained as a local feature combined with the global feature of the whole body, focusing on the head, to enhance the head feature information. The experimental results show that ViTReID exceeds the baseline method (SSG) by 14% (Market1501 → MSMT17) in mean average precision (mAP). In MSMT17 → Market1501, ViTReID is 1.2% higher in rank-1 (R1) accuracy than a state-of-the-art method (SPCL); in PersonX → MSMT17, the mAP is 3.1% higher than that of the MMT-dbscan method, and in PersonX → Market1501, the mAP is 1.5% higher than that of the MMT-dbscan method.

Few-Shot Object Detection in Remote Sensing Image Interpretation: Opportunities and Challenges

Recent years have witnessed rapid development and remarkable achievements on deep learning object detection in remote sensing (RS) images. The growing improvement of the accuracy is inseparable from the increasingly complex deep convolutional neural network and the huge amount of sample data. However, the under-fitting neural network will damage the detection performance facing the difficulty of sample acquisition. Thus, it evolves into few-shot object detection (FSOD). In this article, we first briefly introduce the object detection task and its algorithms, to better understand the basic detection frameworks followed by FSOD. Then, FSOD design methods in RS images for three important aspects, such as sample, model, and learning strategy, are respectively discussed. In addition, some valuable research results of FSOD in computer vision field are also included. We advocate a wide research technique route, and some advice about feature enhancement and multi-modal fusion, semantics extraction and cross-domain mapping, fine-tune and meta-learning strategies, and so on, are provided. Based on our stated research route, a novel few-shot detector that focuses on contextual information is proposed. At the end of the paper, we summarize accuracy performance on experimental datasets to illustrate the achievements and shortcomings of the stated algorithms, and highlight the future opportunities and challenges of FSOD in RS image interpretation, in the hope of providing insights into future research.

Field validation of deep learning based Point-of-Care device for early detection of oral malignant and potentially malignant disorders

Early detection of oral cancer in low-resource settings necessitates a Point-of-Care screening tool that empowers Frontline-Health-Workers (FHW). This study was conducted to validate the accuracy of Convolutional-Neural-Network (CNN) enabled m(mobile)-Health device deployed with FHWs for delineation of suspicious oral lesions (malignant/potentially-malignant disorders). The effectiveness of the device was tested in tertiary-care hospitals and low-resource settings in India. The subjects were screened independently, either by FHWs alone or along with specialists. All the subjects were also remotely evaluated by oral cancer specialist/s. The program screened 5025 subjects (Images: 32,128) with 95% (n = 4728) having telediagnosis. Among the 16% (n = 752) assessed by onsite specialists, 20% (n = 102) underwent biopsy. Simple and complex CNN were integrated into the mobile phone and cloud respectively. The onsite specialist diagnosis showed a high sensitivity (94%), when compared to histology, while telediagnosis showed high accuracy in comparison with onsite specialists (sensitivity: 95%; specificity: 84%). FHWs, however, when compared with telediagnosis, identified suspicious lesions with less sensitivity (60%). Phone integrated, CNN (MobileNet) accurately delineated lesions (n = 1416; sensitivity: 82%) and Cloud-based CNN (VGG19) had higher accuracy (sensitivity: 87%) with tele-diagnosis as reference standard. The results of the study suggest that an automated mHealth-enabled, dual-image system is a useful triaging tool and empowers FHWs for oral cancer screening in low-resource settings.

Low-complexity trainable denoising filter structure for image recognition applications

Even in modern imaging equipment, different types of noise make their way into the images especially under bad lighting conditions and rough communication channels. While applying denoising filters, the preservation of the edge in images has always been an essential albeit a difficult task. This is particularly important for accuracy of further image processing tasks, such as recognition. We described a denoising approach to process noisy images specific for the task of classification with minimal impact on image resolution using low-complexity operations. Specifically, support vector machine-based machine learning classifiers have been proposed to dynamically select an appropriate filtering kernel while processing individual pixels in the input image. The selection is based on the image features in a local window. The provided experimental results on standard dataset prove that the proposed scheme performs as well as contemporary complex convolutional neural network (CNN)-based approaches for the task of denoising images before classification in the presence of a variety of mixed-type noise. Moreover, the proposed filtering approach can run at much higher speeds than their CNN-based counterparts while delivering similar or better performance. The proposed framework has been made publicly available as open-source code to facilitate further work by researchers and practitioners interested in this area.

Real Time Object Detection Using Deep Learning

Abstract: Visually impaired people have difficulty moving safely and independently, which interferes with normal indoor and outdoor work and social activities. Similarly, they have a hard time identifying the basics of the environment. This paper presents a model for detecting the brightness and key colors of real-time images using the RGB method with an external camera and identifying basic objects and face recognition from human datasets.[2]. Object detection is a department of pc imaginative and prescient that appears for times of lexical entities in photographs and videos. The gadget makes use of the ESP-32 Cam's digital digicam to continuously seize severa frames, which can be sooner or later converted to audio segments. In this project, we use the You Only Look Once V3 (YOLO v3)algorithm, which runs thru a version of a really complex Convolutional Neural Network structure with OpenCV. Then with the aid of using the usage of Google Text to Speech, we convert the photo to textual content and afterwards textual content - to - speech for the visually impaired individual. Thus, the Visually Impaired individual receives the place of the gadgets withinside the digital digicam's view through audio. Distance calculation is aided with the aid of using an ultrasonic sensor. By the usage of The amassed consequences show that the proposed prototype is a hit in presenting visually impaired customers with the cappotential to realise surprising settings the usage of a user-pleasant machine that integrates this unique item detection Model

Complex "zero-shot" super-resolution reconstruction algorithm for THz imaging.

To deal with a terahertz (THz) super-resolution (SR) algorithm based on a convolutional neural network (CNN) without standard training datasets, a complex "zero-shot" SR (CZSSR) reconstruction algorithm is proposed according to the internal image statistics with a five-layer complex CNN model. Instead of relying on pre-training, the proposed method is of sound self-adaptability. Compared with real ZSSR, the peak SNR of CZSSR rose by about 0.94 dB, MSE decreased by 0.042, and SSIM increased by about 40% for the SR result of the measured data. The results show that the CZSSR method can solve the low-resolution problem of a THz imaging system and the shortage of datasets in THz SR based on CNN. Therefore, this research is of great significance for application in the fields of medical imaging and non-destructive detection.

A slimmer and deeper approach to deep network structures for low‐level vision tasks

AbstractDeep network design is a fundamental challenge. A right trade‐off between depth and complexity of convolutional neural networks is of significant importance to applications in low‐level vision tasks. Wider feature maps could be beneficial to performance and generality but would increase computational complexity. In this paper, we rethink the balance between width of the feature maps and depth of the network especially for image restoration tasks including deblurring, dehazing, super‐resolution, and denoising. We explore a new approach to network structure by encouraging more depth to deal with restoration requirements while decreasing the width of some feature maps. Such a slimmer and deeper approach can enhance the performance while maintaining the same level of computational costs. We have experimentally evaluated the performances of the proposed approach on four image restoration tasks and obtained state‐of‐the‐art results on quantitative measures and qualitative assessments, demonstrating the effectiveness of the approach.

Part-Based Obstacle Detection Using a Multiple Output Neural Network.

Detecting the objects surrounding a moving vehicle is essential for autonomous driving and for any kind of advanced driving assistance system; such a system can also be used for analyzing the surrounding traffic as the vehicle moves. The most popular techniques for object detection are based on image processing; in recent years, they have become increasingly focused on artificial intelligence. Systems using monocular vision are increasingly popular for driving assistance, as they do not require complex calibration and setup. The lack of three-dimensional data is compensated for by the efficient and accurate classification of the input image pixels. The detected objects are usually identified as cuboids in the 3D space, or as rectangles in the image space. Recently, instance segmentation techniques have been developed that are able to identify the freeform set of pixels that form an individual object, using complex convolutional neural networks (CNNs). This paper presents an alternative to these instance segmentation networks, combining much simpler semantic segmentation networks with light, geometrical post-processing techniques, to achieve instance segmentation results. The semantic segmentation network produces four semantic labels that identify the quarters of the individual objects: top left, top right, bottom left, and bottom right. These pixels are grouped into connected regions, based on their proximity and their position with respect to the whole object. Each quarter is used to generate a complete object hypothesis, which is then scored according to object pixel fitness. The individual homogeneous regions extracted from the labeled pixels are then assigned to the best-fitted rectangles, leading to complete and freeform identification of the pixels of individual objects. The accuracy is similar to instance segmentation-based methods but with reduced complexity in terms of trainable parameters, which leads to a reduced demand for computational resources.

Leveraging Guided Backpropagation to Select Convolutional Neural Networks for Plant Classification.

The development of state-of-the-art convolutional neural networks (CNN) has allowed researchers to perform plant classification tasks previously thought impossible and rely on human judgment. Researchers often develop complex CNN models to achieve better performances, introducing over-parameterization and forcing the model to overfit on a training dataset. The most popular process for evaluating overfitting in a deep learning model is using accuracy and loss curves. Train and loss curves may help understand the performance of a model but do not provide guidance on how the model could be modified to attain better performance. In this article, we analyzed the relation between the features learned by a model and its capacity and showed that a model with higher representational capacity might learn many subtle features that may negatively affect its performance. Next, we showed that the shallow layers of a deep learning model learn more diverse features than the ones learned by the deeper layers. Finally, we propose SSIM cut curve, a new way to select the depth of a CNN model by using the pairwise similarity matrix between the visualization of the features learned at different depths by using Guided Backpropagation. We showed that our proposed method could potentially pave a new way to select a better CNN model.

Quantum-inspired complex convolutional neural networks

Quantum-inspired neural network is one of the interesting researches at the junction of the two fields of quantum computing and deep learning. Several models of quantum-inspired neurons with real parameters have been proposed, which are mainly used for three-layer feedforward neural networks. In this work, we improve the quantum-inspired neurons by exploiting the complex-valued weights which have richer representational capacity and better non-linearity. We then extend the method of implementing the quantum-inspired neurons to the convolutional operations, and naturally draw the models of quantum-inspired convolutional neural networks (QICNNs) capable of processing high-dimensional data. Five specific structures of QICNNs are discussed which are different in the way of implementing the convolutional and fully connected layers. The performance of classification accuracy of the five QICNNs are tested on the MNIST and CIFAR-10 datasets. The results show that the QICNNs can perform better in classification accuracy on MNIST dataset than the classical CNN. More learning tasks that our QICNN can outperform the classical counterparts will be found.

Differentiating Malignant from Benign Pigmented or Non-Pigmented Skin Tumours-A Pilot Study on 3D Hyperspectral Imaging of Complex Skin Surfaces and Convolutional Neural Networks.

Several optical imaging techniques have been developed to ease the burden of skin cancer disease on our health care system. Hyperspectral images can be used to identify biological tissues by their diffuse reflected spectra. In this second part of a three-phase pilot study, we used a novel hand-held SICSURFIS Spectral Imager with an adaptable field of view and target-wise selectable wavelength channels to provide detailed spectral and spatial data for lesions on complex surfaces. The hyperspectral images (33 wavelengths, 477–891 nm) provided photometric data through individually controlled illumination modules, enabling convolutional networks to utilise spectral, spatial, and skin-surface models for the analyses. In total, 42 lesions were studied: 7 melanomas, 13 pigmented and 7 intradermal nevi, 10 basal cell carcinomas, and 5 squamous cell carcinomas. All lesions were excised for histological analyses. A pixel-wise analysis provided map-like images and classified pigmented lesions with a sensitivity of 87% and a specificity of 93%, and 79% and 91%, respectively, for non-pigmented lesions. A majority voting analysis, which provided the most probable lesion diagnosis, diagnosed 41 of 42 lesions correctly. This pilot study indicates that our non-invasive hyperspectral imaging system, which involves shape and depth data analysed by convolutional neural networks, is feasible for differentiating between malignant and benign pigmented and non-pigmented skin tumours, even on complex skin surfaces.

Rolling bearing fault diagnosis algorithm using overlapping group sparse-deep complex convolutional neural network

Rolling bearing fault diagnosis algorithm using overlapping group sparse-deep complex convolutional neural network

Image feature extraction and galaxy classification: a novel and efficient approach with automated machine learning

ABSTRACT In this work, we explore the possibility of applying machine learning methods designed for 1D problems to the task of galaxy image classification. The algorithms used for image classification typically rely on multiple costly steps, such as the point spread function deconvolution and the training and application of complex Convolutional Neural Networks of thousands or even millions of parameters. In our approach, we extract features from the galaxy images by analysing the elliptical isophotes in their light distribution and collect the information in a sequence. The sequences obtained with this method present definite features allowing a direct distinction between galaxy types. Then, we train and classify the sequences with machine learning algorithms, designed through the platform Modulos AutoML. As a demonstration of this method, we use the second public release of the Dark Energy Survey (DES DR2). We show that we are able to successfully distinguish between early-type and late-type galaxies, for images with signal-to-noise ratio greater than 300. This yields an accuracy of $86{{\ \rm per\ cent}}$ for the early-type galaxies and $93{{\ \rm per\ cent}}$ for the late-type galaxies, which is on par with most contemporary automated image classification approaches. The data dimensionality reduction of our novel method implies a significant lowering in computational cost of classification. In the perspective of future data sets obtained with e.g. Euclid and the Vera Rubin Observatory, this work represents a path towards using a well-tested and widely used platform from industry in efficiently tackling galaxy classification problems at the peta-byte scale.

A method of multi-channel SAR moving target detection and imaging based on complex CNN

Aiming at the problem of the detecting difficulty and imaging the moving target submerged in clutter, a multi-channel Synthetic Aperture Radar (SAR) moving target detection and imaging method based on complex Convolutional Neural Network (CNN) is proposed. First, using Range Doppler algorithm (RDA) to image SAR data to obtain images with clutter, noise and defocused moving targets. Then, the SAR image is used as the input of complex CNN. A SAR image with the clutter and noise eliminated, and the moving target in focus is obtained after processing by neural network. The simulation experiment results show this method is effectiveness.