Cross-entropy Research Articles

Convolutional Neural Network Driven Computer Vision Based Facial Emotion Detection and Recognition

Computer vision is a sub branch of artificial intelligence (AI) that enables computers and systems to derive substitutive information from digital images and Video. Artificial intelligence plays a significant role in the area of security and surveillance, image processing and machine learning. In computer vision and image processing object detection algorisms are used to detect objects from certain classes of images or video. There is a scope identification of human face emotion Facial emotion recognition is done using computer vision algorism whether the person’s emotion is Happy, sad, fear, disgust, neutral and so on. Object detection algorism are used in deep learning used to classify the detected the regions. Facial emotion recognition is an emerging research area for improving human and computer interaction. It plays a crucial role in security, social communication commercial enterprise and law enforcement. In this research project CNN is used for training the data and predicting seven emotions such as anger, happy, sad, disgust, fear neutral and surprise. In this paper the experiment will be conduct using convolutional neural network as classifier, since it is multi class classification relu, softmax (activation function), categorical cross entropy(loss function) dropout max pooling conducted. The researcher tried to train the model by 80/20, 70/30, 90/10 train test split. However 70/30 train test split out performs over the other. The performance of the model is measured by using the epoch 10 and dropout 0.3. Totally the model is performed 93.8% in the training accuracy and it 75% for the testing.

SENTIMENT ANALYSIS USING LONG TERM MEMORY (LSTM) BOOK CASE STUDY: UNSOLICITED ADVICE FOR MURDERERS BY VERA WONG'S

Reading books is one of the most effective ways to reduce stress. In today's digital era, access to finding and buying books is getting easier, so reader reviews are important in choosing books that match interests. However, with a large number of reviews, Natural Language Processing (NLP) with the Long Short-Term Memory (LSTM) method is used to help analyze positive and negative sentiments from many book reviews. This sentiment analysis is useful for readers to evaluate the quality of books, as well as for authors and sellers to find out the opinions of readers and improve the quality of their work. In this study, the book review dataset "Vera Wong's Unsolicited Advice for Murderers" from the Goodreads website is used, which is then divided into training data and validation data with a ratio of 75%: 25%. The Long Short-Term Memory (LSTM) method is used to analyze the sentiment of the reviews. The model architecture built consists of Embedding Layer, LSTM Layer with 128 neuron units, 3 Dense Layer with ReLU activation function, 3 Dropout Layer, and Fully Connected Layer with and Sigmoid activation function, Binary Cross Entropy loss function, and RMSprop optimizer. The model training process was conducted with 30 epochs. The evaluation results show that the model achieved an accuracy of 90%, indicating the model performs relatively well in correctly classifying positive sentiments.

DualAttNet: Synergistic fusion of image-level and fine-grained disease attention for multi-label lesion detection in chest X-rays

Chest radiographs are the most commonly performed radiological examinations for lesion detection. Recent advances in deep learning have led to encouraging results in various thoracic disease detection tasks. Particularly, the architecture with feature pyramid network performs the ability to recognise targets with different sizes. However, such networks are difficult to focus on lesion regions in chest X-rays due to their high resemblance in vision. In this paper, we propose a dual attention supervised module for multi-label lesion detection in chest radiographs, named DualAttNet. It efficiently fuses global and local lesion classification information based on an image-level attention block and a fine-grained disease attention algorithm. A binary cross entropy loss function is used to calculate the difference between the attention map and ground truth at image level. The generated gradient flow is leveraged to refine pyramid representations and highlight lesion-related features. We evaluate the proposed model on VinDr-CXR, ChestX-ray8 and COVID-19 datasets. The experimental results show that DualAttNet surpasses baselines by 0.6% to 2.7% mAP and 1.4% to 4.7% AP50 with different detection architectures. The code for our work and more technical details can be found at https://github.com/xq141839/DualAttNet.

Automatic segmentation framework of X-Ray tomography data for multi-phase rock using Swin Transformer approach

A thorough understanding of the impact of the 3D meso-structure on damage and failure patterns is essential for revealing the failure conditions of composite rock materials such as coal, concrete, marble, and others. This paper presents a 3D XCT dataset of coal rock with 1372 slices (each slice contains 1720 × 1771 pixels in x × y direction). The 3D XCT datasets were obtained by MicroXMT-400 using the 225/320kv Nikon Metris custom bay. The raw datasets were processed by an automatic semantic segmentation method based on the Swin Transformer (Swin-T) architecture, which aims to overcome the issue of large errors and low efficiency for traditional methods. The hybrid loss function proposed can also effectively mitigate the influence of large volume features in the training process by incorporating modulation terms into the cross entropy loss, thereby enhancing the accuracy of segmentation for small volume features. This dataset will be available to the related researchers for further finite element analysis or microstructural statistical analysis, involving complex physical and mechanical behaviors at different scales.

A Novel Channel Estimation Framework in MIMO Using Serial Cascaded Multiscale Autoencoder and Attention LSTM with Hybrid Heuristic Algorithm.

In wireless communication, multiple signals are utilized to receive and send information in the form of signals simultaneously. These signals consume little power and are usually inexpensive, with a high data rate during data transmission. An Multi Input Multi Output (MIMO) system uses numerous antennas to enhance the functionality of the system. Moreover, system intricacy and power utilization are difficult and highly complicated tasks to achieve in an Analog to Digital Converter (ADC) at the receiver side. An infinite number of MIMO channels are used in wireless networks to improve efficiency with Cross Entropy Optimization (CEO). ADC is a serious issue because the data of the accepted signal are completely lost. ADC is used in the MIMO channels to overcome the above issues, but it is very hard to implement and design. So, an efficient way to enhance the estimation of channels in the MIMO system is proposed in this paper with the utilization of the heuristic-based optimization technique. The main task of the implemented channel prediction framework is to predict the channel coefficient of the MIMO system at the transmitter side based on the receiver side error ratio, which is obtained from feedback information using a Hybrid Serial Cascaded Network (HSCN). Then, this multi-scaled cascaded autoencoder is combined with Long Short Term Memory (LSTM) with an attention mechanism. The parameters in the developed Hybrid Serial Cascaded Multi-scale Autoencoder and Attention LSTM are optimized using the developed Hybrid Revised Position-based Wild Horse and Energy Valley Optimizer (RP-WHEVO) algorithm for minimizing the "Root Mean Square Error (RMSE), Bit Error Rate (BER) and Mean Square Error (MSE)" of the estimated channel. Various experiments were carried out to analyze the accomplishment of the developed MIMO model. It was visible from the tests that the developed model enhanced the convergence rate and prediction performance along with a reduction in the computational costs.

NIMG-26. DIFFERENTIATION OF IDH-WILDTYPE GLIOBLASTOMA AND TUMEFACTIVE MULTIPLE SCLEROSIS USING PRE-OPERATIVE BRAIN MRI AND DEEP LEARNING: RIGOROUS AND UNBIASED EXPERIMENTAL DESIGN AND ANALYTICAL APPROACH

Abstract Clinical management of IDH wildtype glioblastoma (GBM) and tumefactive multiple sclerosis (tMS) is drastically different. GBM requires maximal safe resection followed by chemoradiation, while tMS outcome is worsened by surgery and radiotherapy. Misdiagnosis of brain lesions may expose patients to unnecessary anxiety, surgery, or radiotherapy. Noninvasive methods to help with accurate diagnosis of tumor and non-tumor etiologies are needed. tMS subjects were identified by the Mayo Clinic Center for Multiple Sclerosis and Autoimmune Neurology (CMSAN). GBM subjects were matched to tMS by age at diagnosis, index MRI date, and sex. Inclusion criteria included one cm minimal lesion size and pre-operative post-contrast T1 and T2 images available for analysis. To identify potential experimental design bias, MRI technology and clinical characteristics were compared across groups. A 3D-DenseNet121 was used to develop a classification model using prespecified parameters: 600 epochs, batch size 16, learning rate 10-3, cross entropy loss, and AdamW optimizer. The stopping rule was defined as three sequential epoch cross entropy loss < 0.02. Five-fold cross validation (CV) was utilized to obtain an unbiased estimate of predictive performance (overfitting) and sensitivity of the 3D-DenseNet121 algorithm was evaluated. 224 subjects (114 GBM, 110 tMS) were analyzed: 40% male, median age 46. After one round of CV, the average unbiased estimate of sensitivity and specificity for predicting GBM from the final classification model was 0.86 and 0.73, respectively. To assess the reproducibility of the deep learning algorithm (3D-DenseNet121), the exact same CV folds were rerun resulting in average sensitivity and specificity of 0.82 and 0.75, respectively. Obtaining unbiased estimates of model performance using internal validation is critical before evaluating a final model on an external dataset. Understanding sensitivity of a deep learning algorithm is important for model validity. Differentiation of GBM and tMS appears reasonable with MRI; external validation is in progress.

Deep learning model to identify potential occurrence of reduced LVEF in patients with implantable loop recorders

Abstract Background Early detection of left ventricular systolic dysfunction (LVSD) using minimally invasive electrocardiographic assessment in an ambulatory setting may be important considering the increasing economic burden of HF. Recently artificial intelligence (AI) algorithms have been reported to classify reduced ejection fraction using snapshot 12-lead ECG measurements, however the ability of AI to identify LVSD using only a single lead ambulatory ECG is unknown. Purpose We aimed to develop a convolution neural network (CNN) based deep learning algorithm using single lead ECG to predict ejection fraction less than 40% in ambulatory patients implanted with implantable loop recorders (ILRs). Methods We linked ILR patients with LVEF measurements from a de-identified database of aggregated electronic health record (EHR) data during the period of 2007-2021 to a manufacturer’s device database with ambulatory single lead ECGS. The routine ECG transmissions from the ILR devices were paired with LVEF measured within a week. The data was pre-processed to obtain the time-frequency components of ECG using wavelet transform. The wavelet coefficients were then used as input to train and validate a 2D-CNN model. An independent test dataset consisting of data from patient not included in training/validation dataset was used to assess the performance of CNN model to classify LVEF< = 40% using the metrics area under the curve (AUC), sensitivity and specificity. Results A total of 35,741 unique LVEF-ECG pairs collected from 2,249 ILR patients were used to train the model; an independent validation dataset consisted of 6,721 unique LVEF ECG pairs from 750 patients and independent test dataset consisted of 6,611 unique LVEF ECG pairs from 750 patients. In our CNN model, cross entropy was used as loss function and adaptive moment estimation (ADAM) was used as the optimizer. The initial learning rate was selected as 0.0001, epochs as 20 and mini batch size as 64. The threshold to delineate reduced ejection fraction (LVEF< = 40%) was set as 0.1 (Figure 1). The model yielded accuracy, sensitivity, specificity, AUC of 75%,70%, 76% and 0.8 respectively in the independent test set (Figure 2). Conclusion A deep learning algorithm applied to ambulatory single lead ECGs acquired by ILR can detect LVEF< = 40%. This continuous ambulatory AI-ECG monitoring may allow to identify longitudinal changes in LVEF post ILR implant via remote monitoring.Figure 1:Model output vs Actual EF rangeFigure 2:ROC

A novel cross-entropy framework for pattern design quality evaluation based on interval-valued intuitionistic fuzzy multi-attribute decision making

Paper packaging is favored by manufacturers in various industries due to its advantages such as economy and environmental friendliness, and the application of paper packaging in various types of products is also becoming increasingly popular. However, with the growth of the national economy and the increasing purchasing power of the people, consumers not only have higher requirements for the functionality of products, but also have higher requirements for the visual effect of product packaging. Based on the potential consumption trend of “looking for things”, enterprises need to further strengthen the artistic design effect when using paper packaging, promoting the unity of practicality and aesthetics of paper packaging. Pattern design art, as a component of graphic design art, is mainly based on pattern modeling to convey artistic concepts and aesthetic temperament. Integrating pattern design art into the process of paper packaging design can effectively enhance the artistic value of paper packaging, thereby adding color to products. The pattern design quality evaluation is a classical multi-attribute decision making (MADM). In this paper, the MADM methodology is investigated based on cross-entropy (CE) method under interval-valued intuitionistic fuzzy sets (IVIFSs) for pattern design quality evaluation. First of all, this paper extends the CE method under the IVIFSs to propose the interval-valued intuitionistic fuzzy number cross-entropy (IVIFN-CE) method, it enlarges the application range of the CE method. Secondly, a new MADM model for pattern design quality evaluation based on IVIFN-CE algorithm is proposed. In this algorithm, attribute weight is calculated by CRITIC method which is a well-known objective method. At the end of this study, we noticed to a comparison between the proposed IVIFN-CE approach with other existing methods to verify the effectiveness of the algorithm.

Nonlinear Feature Selection Neural Network via Structured Sparse Regularization.

Feature selection is an important and effective data preprocessing method, which can remove the noise and redundant features while retaining the relevant and discriminative features in high-dimensional data. In real-world applications, the relationships between data samples and their labels are usually nonlinear. However, most of the existing feature selection models focus on learning a linear transformation matrix, which cannot capture such a nonlinear structure in practice and will degrade the performance of downstream tasks. To address the issue, we propose a novel nonlinear feature selection method to select those most relevant and discriminative features in high-dimensional dataset. Specifically, our method learns the nonlinear structure of high-dimensional data by a neural network with cross entropy loss function, and then using the structured sparsity norm such as l2,p -norm to regularize the weights matrix connecting the input layer and the first hidden layer of the neural network model to learn weight of each feature. Therefore, a structural sparse weights matrix is obtained by conducting nonlinear learning based on a neural network with structured sparsity regularization. Then, we use the gradient descent method to achieve the optimal solution of the proposed model. Evaluating the experimental results on several synthetic datasets and real-world datasets shows the effectiveness and superiority of the proposed nonlinear feature selection model.

CA-YOLOX: Deep Learning-Guided Road Intersection Location From High-Resolution Remote Sensing Images

The location of road intersection from high resolution remote sensing (HRRS) images can be automatically obtained by deep learning. This has become one of the current data sources in urban smart transportation. However, limited by the small size, diverse types, complex distribution, and missing sample labels of road intersections in actual scenarios, it is difficult to accurately represent key features of road intersection by deep neural network (DNN) model. A new coordinate attention (CA) module-YOLOX (CA-YOLOX) method for accurately locating road intersections from HRRS images is presented. First, the spatial pyramid pooling (SPP) module is introduced into the backbone convolution network between the Darknet-53’ last feature layer and feature pyramid networks (FPN) structure. Second, the CA module is embedded into the feature fusion structure in FPN to focus more on the spatial shape distribution and texture features of road intersections. Third, we use focal loss to replace the traditional binary cross entropy (BCE) loss in the confidence loss to improve the iteration speed of the CA-YOLOX network. Finally, an extensive empirical experiment on Potsdam, IKONOS datasets, and ablation study is then implemented and tested. The results show that the presented CA-YOLOX method can promote the location accuracy of road intersection from HRRS images compared to the traditional You only look once (YOLO) model.

A Business Intelligence Expert System for Predicting Market Price in Stock Trading using Data Analysis: Deep Learning Model With Feature Selection Mechanism

Because of the availability of data and reasonable processing capability, business intelligence methods are rapidly being used in finance, such as managing assets, trading using algorithms, credit financing, and blockchain-based financing. Machine learning (ML) algorithms use enormous amounts of data to automatically understand and enhance predictability and performance via knowledge and data without being programmed by someone. Due to the stock data’s dynamic, high-noise, non-parametric, non-linear, and chaotic qualities, the stock market prediction has been a challenge and has received much interest from scholars over the last decade. Some studies seek a method for accurately predicting stock prices; however, due to the high correlation between stock prices, stock market analysis is more complex. So, this paper proposes an improved stock price prediction (SPP) model using a novel optimal parameter tuned with cross entropy included bidirectional long short-term memory (OPCBLSTM) with efficient feature extraction and selection schemes. It starts with missing values imputation, and data standardization on the collected dataset. From the preprocessed dataset, the features are extracted using modified rectifier linear unit activation based residual network (MRResNet50). Then the optimal features are selected using the improved whale optimization algorithm (IWOA). Finally, the SPP is done using the OPCBLSTM. The experimental results proved that the proposed method achieves more high-level outcomes than the traditional methods.

Detecting Persuasion Attempts on Social Networks: Unearthing the Potential of Loss Functions and Text Pre-Processing in Imbalanced Data Settings

The rise of social networks and the increasing amount of time people spend on them have created a perfect place for the dissemination of false narratives, propaganda, and manipulated content. In order to prevent the spread of disinformation, content moderation is needed. However, manual moderation is unfeasible due to the large amount of daily posts. This paper studies the impact of using different loss functions on a multi-label classification problem with an imbalanced dataset, consisting of 20 persuasion techniques and only 950 samples, provided by SemEval’s 2021 Task 6. We used machine learning models, such as Naive Bayes and Decision Trees, and a custom deep learning architecture, based on DistilBERT and Convolutional Layers. Overall, the machine learning models achieved far worse results than the deep learning model, using Binary Cross Entropy, which we considered our baseline deep learning model. To address the class imbalance problem, we trained our model using different loss functions, such as Focal Loss and Asymmetric Loss. The latter providing the best results, particularly for the least represented classes.

Do we really need dice? The hidden region-size biases of segmentation losses

Most segmentation losses are arguably variants of the Cross-Entropy (CE) or Dice losses. On the surface, these two categories of losses (i.e., distribution based vs. geometry based) seem unrelated, and there is no clear consensus as to which category is a better choice, with varying performances for each across different benchmarks and applications. Furthermore, it is widely argued within the medical-imaging community that Dice and CE are complementary, which has motivated the use of compound CE-Dice losses. In this work, we provide a theoretical analysis, which shows that CE and Dice share a much deeper connection than previously thought. First, we show that, from a constrained-optimization perspective, they both decompose into two components, i.e., a similar ground-truth matching term, which pushes the predicted foreground regions towards the ground-truth, and a region-size penalty term imposing different biases on the size (or proportion) of the predicted regions. Then, we provide bound relationships and an information-theoretic analysis, which uncover hidden region-size biases: Dice has an intrinsic bias towards specific extremely imbalanced solutions, whereas CE implicitly encourages the ground-truth region proportions. Our theoretical results explain the wide experimental evidence in the medical-imaging literature, whereby Dice losses bring improvements for imbalanced segmentation. It also explains why CE dominates natural-image problems with diverse class proportions, in which case Dice might have difficulty adapting to different region-size distributions. Based on our theoretical analysis, we propose a principled and simple solution, which enables to control explicitly the region-size bias. The proposed method integrates CE with explicit terms based on L1 or the KL divergence, which encourage segmenting region proportions to match target class proportions, thereby mitigating class imbalance but without losing generality. Comprehensive experiments and ablation studies over different losses and applications validate our theoretical analysis, as well as the effectiveness of explicit and simple region-size terms. The code is available at https://github.com/by-liu/SegLossBias .

Diagnosing brain tumors based on Edge-Impulse platform

Brain tumors pose a serious threat to human health, and timely diagnosis is crucial for effective treatment. Conventional methods for diagnosing brain tumors, such as manual diagnosis, are often inaccurate and comparatively expensive. In this paper, a new method of diagnosis that can replace classic methods in the future is proposed. Ideas of machine learning is used to train a model so that it can classify images of brains into 2 categories: with tumors and with no tumors. Process of training is based on Edge Impulse, an AI platform on which model can be created conveniently. Technology used in the training is Convolutional Neural Network, which uses convolution as a mathematical operation instead of traditional matrix multiplication. It comprises convolutional layers, pooling layers and fully connected layers. Convolutional layers and pooling layers are used to operate feature maps, transforming them into a 1-dimensional matrix, followed by being multiplied by matrices of weights in fully connected layers. Cross entropy loss function is utilised to assess accuracy and modify weights. Transfer learning is also used to improve effectiveness and efficiency of training. Additionally, experiments are taken to determine the best image size suiting the model. This is accomplished by changing sizes of input image while keeping other parameters constant. Results of the experiments indicate that the model has a fantastic performance on brain tumor diagnosis.

Suppressing Sidelobes in Metasurface-Based Antennas Using a Cross-Entropy Method Variant and Full Wave Electromagnetic Simulations

Managing sidelobe levels (SLLs) in metasurface-driven beam-steering antennas poses a significant challenge due to intrinsic factors leading to grating lobes. Our proposed method employs an equivalent model to efficiently optimize large periodic metasurfaces. This model predicts complete metasurface performance, accounting for mutual coupling between patches. We introduce an evolutionary optimization algorithm based on the cross-entropy (CE) method to enhance PGM-based beam-steering antennas and suppress sidelobes. Two strategies are employed: the first is to optimize the patch dimensions for a sidelobe-free pattern, and the second is to maintain the PGM dimensions while optimizing the feed array amplitudes. Both strategies effectively suppress sidelobes, offering insights into the CE method’s applicability and effectiveness for CPU-intensive electromagnetic optimization challenges. The proposed CE method variant retains its simplicity while improving monitoring capabilities, addressing this limitation. Smaller generations yield better improvements per evaluation. The uniqueness of the proposed optimization strategy lies in its utilization of an equivalent 1D metasurface model for optimization that not only considers the mutual coupling between identical unit cells along the y-direction within a complete metasurface but also takes into account the distinct cells along the x-direction. Moreover, the 1D metasurface model incorporates the influence of edge effects along the x-direction.

Profit Maximization for Cache-Enabled Vehicular Mobile Edge Computing Networks

In this paper, we investigate a multiuser cache-enabled vehicular mobile edge computing (MEC) network, where one edge server (ES) has some caching and computing capabilities to assist the task computing from the vehicular users. The introduce of caching into the MEC network significantly affects the system performance such as the latency, energy consumption and profit at the ES, which imposes a critical challenge on the system design and optimization. To solve this challenge, we firstly design the vehicular MEC network in a non-competitive environment by maximizing the profit of the ES with a predetermined threshold of user QoE, and jointly exploit the caching and computing resources in the network. We then model the optimization problem into a binary integer programming problem, and adopt the cross entropy (CE) method to obtain the effective offloading and caching decision with a low complexity. Simulation results are finally presented to verify that the proposed scheme can achieve the near optimal performance of the conventional branch and bound (BnB) scheme, while sharply reduce the computational complexity compared to the BnB.

Early chatter identification based on optimized VMD with multi-band information fusion and compression method in robotic milling process

Undesirable self-excited chatter has always been a typical issue restricting the improvement of robotic milling quality and efficiency. Sensitive chatter identification based on processing signals can prompt operators to adjust the machining process and prevent chatter damage. Compared with the traditional machine tool, the uncertain multiple chatter frequency bands and the band-moving of the chatter frequency in robotic milling process make it more challenging to extract chatter information. This paper proposes a novel method of chatter identification using optimized variational mode decomposition (OVMD) with multi-band information fusion and compression technology (MT). During the robotic milling process, the number of decomposed modes k and the penalty coefficient α are optimized based on the dominant component of frequency scope partition and fitness of the mode center frequency. Moreover, the mayfly optimization algorithm (MA) is employed to obtain the global optimal parameter selection. In order to conquer information collection about the uncertain multiple chatter frequency bands and the band-moving of the chatter frequency in robotic milling process, MT is presented to reduce computation and extract signal characteristics. Finally, the cross entropy of the image (CEI) is proposed as the final chatter indicator to identify the chatter occurrence. The robotic milling experiments are carried out to verify the proposed method, and the results show that it can distinguish the robotic milling condition by extracting the uncertain multiple chatter frequency bands and overcome the band-moving of the chatter frequency in robotic milling process.

Self-Adaptive Imbalanced Domain Adaptation With Deep Sparse Autoencoder

Domain adaptation aims to transfer knowledge between different domains to develop an effective hypothesis in the target domain with scarce labeled data, which is an effective method for remedying the problem of labeled data requirement in deep learning. In reality, it is unavoidable that the dataset has a large gap in the number of positive and negative instances across different categories in source and target domains, which is the imbalanced domain adaptation problem. However, since the imbalanced degree always varies greatly in different source and target domains datasets, most existing imbalanced domain adaptation models fix the imbalanced parameters which can not adapt to the change of the proportion between positive and negative instances in different domains. To address this problem, in this paper, we propose a self-adaptive imbalanced domain adaptation method via deep sparse autoencoder, which can adjust the model automatically according to the imbalanced extent for bridging the chasm of domains. More specifically, the self-adaptive imbalanced cross entropy loss is designed for emphasizing more on minority categories and compensating the bias of training loss automatically. In addition, to alleviate the deficient problem of labeled data, we further propose the unlabeled information incorporating method by minimizing the distribution discrepancy of high-level representation space between the source and target domains. Experiments on several real-world datasets demonstrate the effectiveness of our method compared to other state-of-the-art methods.

Understanding Cloud Systems’ Structure and Organization Using a Machine’s Self-Learning Approach

Abstract In this study, we introduce a self-supervised deep neural network approach to classify satellite images into independent classes of cloud systems. The driving question of the work is to understand whether our algorithm can capture cloud variability and identify distinct cloud regimes. Ultimately, we want to achieve generalization such that the algorithm can be applied to unseen data and thus help automatically extract relevant information important to atmospheric science and renewable energy applications from the ever-increasing satellite data stream. We use cloud optical depth (COD) retrieved from postprocessed high-resolution Meteosat Second Generation (MSG) satellite data as input for the network. The network’s architecture is based on the DeepCluster, version 2, and consists of a convolutional neural network and a multilayer perceptron, followed by a k-means algorithm. We explore the network’s training capabilities by analyzing the centroids and feature vectors found from progressive minimization of the cross-entropy loss function. By making use of additional MSG retrieval products based on multichannel information, we derive the optimum number of classes to determine independent cloud regimes. We test the network capabilities on COD data from 2013 and find that the trained neural network gives insights into the cloud systems’ persistence and transition probability. The generalization on the 2015 data shows good skills of our algorithm with unseen data, but results depend on the spatial scale of cloud systems. Significance Statement This study uses a self-supervised deep neural network to identify distinct cloud systems from cloud optical depth satellite images over central Europe. Satellite-retrieved products support the physical interpretation of the identified cloud classes and help optimize the number of identified classes. The trained neural network gives insights into cloud systems’ persistence and transition probability. The generalization capacity of the deep neural network with unseen data is promising but depends on the spatial scale of cloud systems.

Aerial Gaze Target Recognition Based on Head and Eye Movements

Aerial gaze target recognition is an important step in aerial eye control interaction. In order to achieve accurate aerial gaze target recognition, the VT4LM recognition algorithm was constructed in this paper. By this algorithm, the facial images containing the eyes of the flight operators were first inputted into Vision Transformer (ViT) to extract the local features, and the head posture of the flight operators was inputted into four LSTMs to extract the global features. Then, the local and global features of the flight operators were inputted into three fully connected layers, two dropout layers and one softmax classifier. Finally, the recognition result of aerial gaze target was obtained. In this paper the effectiveness of the VT4LM recognition algorithm was verified through the identification of four aerial line-of-sight gaze targets named Head-up display, Accelerator push rod, Control lever and Rudder by four flight operators during simulated flight. The experimental results showed that the accuracy of the VT4LM algorithm for aerial gaze target recognition reached 89.29% and the Cross Entropy Loss was 1.45. Compared to the other three recognition methods, the VT4LM algorithm had the highest recognition accuracy and minimum loss. When using the VT4LM algorithm to detect four simulated flight operators staring at four aerial gaze targets, the recognition accuracy was all higher than 85.00%. It could be seen that the VT4LM algorithm had a good performance in aerial gaze target recognition.