Output Of Convolutional Neural Network Research Articles

Comparative study: AutoDPR-SEM for enhancing CNN reliability in SRAM-based FPGAs through autonomous reconfiguration

Convolutional neural networks (CNNs) are widely adopted in safety-critical systems, including space applications and autonomous vehicles. Field-programmable gate arrays (FPGAs) based on SRAM are preferred for accelerating CNN computations due to their unique characteristics. However, the configuration memory of FPGAs is susceptible to single event effects (SEEs), which can corrupt computations and lead to misclassification of CNN outputs. In this study, we investigated the impact of SEEs on SRAM-based FPGAs with Two-Photon Absorption (TPA) laser fault injections through a comparative analysis of two popular CNN acceleration architectures: streaming architecture (SA) and single compute engine (SCE). Experimental results show that SA-based CNNs require more hardware resources but exhibit superior resilience against single event upsets (SEUs). Without any Radiation Hardened by Design (RHBD) protection, SCE has an error rate approximately twice as high as SA. To mitigate errors, the Xilinx IP core - Soft Error Mitigation (SEM) is used for error detection and correction, leading to error rate reductions of up to 50 % in both architectures. Importantly, we propose the AutoDPR-SEM (Autonomous Dynamic Partial Reconfiguration for Soft Error Mitigation) approach, which automatically reconfigures the SEM IP core when it remains idle due to uncorrectable errors. AutoDPR-SEM significantly improves CNN error rates, reducing errors by approximately 17.8 times in SCE and 14.8 times in SA. We also applied software level simulation to validate the TPA experiment, showing similar trends of the testing results across all models. In conclusion, the study confirms the feasibility of AutoDPR-SEM in both architectures, showcasing its potential to improve CNN error rates in safety-critical systems.

Real-time early warning and the prediction of air pollutants for sustainable development in smart cities

Air quality forecasting is vital for sustainable development in smart cities and is essential to protect public health through early warning against air pollutants. This article presents a novel ensemble deep learning method able to capture spatiotemporal features on short-term time intervals and extract high data abstraction levels. This method uses a hierarchical framework with a new activation function called HT-Tanh to improve the prediction of air pollutant emissions in real time. As such, a residual neural network (ResNet) is integrated with a convolutional neural network (CNN) to deeply extract the temporal and spatial features from pollutant and meteorological data. The model network encodes the fully connected layer to fine-tune the CNN output and decodes the temporal prediction relations based on gate recurrent unit (GRU) to obtain more accurate results. The reliability of the proposed method is verified through a case study from Shantou City, China. The model performance is compared with the recently developed and traditional deep learning models. Results reveal that the proposed method has good precision in the multiscale air quality predictions. Compared with CNN-GRU, convolutional long short-term memory, and CNN models, the forecasting results of the proposed method are significantly improved. The presented method in this article can support the early warning systems in multiple regions for air pollution modeling studies.

Super-resolution of diffusion-weighted images using space-customized learning model.

Diffusion-weighted imaging (DWI) is a noninvasive method used for investigating the microstructural properties of the brain. However, a tradeoff exists between resolution and scanning time in clinical practice. Super-resolution has been employed to enhance spatial resolution in natural images, but its application on high-dimensional and non-Euclidean DWI remains challenging. This study aimed to develop an end-to-end deep learning network for enhancing the spatial resolution of DWI through post-processing. We proposed a space-customized deep learning approach that leveraged convolutional neural networks (CNNs) for the grid structural domain (x-space) and graph CNNs (GCNNs) for the diffusion gradient domain (q-space). Moreover, we represented the output of CNN as a graph using correlations defined by a Gaussian kernel in q-space to bridge the gap between CNN and GCNN feature formats. Our model was evaluated on the Human Connectome Project, demonstrating the effective improvement of DWI quality using our proposed method. Extended experiments also highlighted its advantages in downstream tasks. The hybrid convolutional neural network exhibited distinct advantages in enhancing the spatial resolution of DWI scans for the feature learning of heterogeneous spatial data.

Two-stage prediction method for capacity aging trajectories of lithium-ion batteries based on Siamese-convolutional neural network

Accurate battery lifespan prediction can guarantee battery safety in applications. Considering the complex nonlinear nature of battery degradation, it is more challenging to predict battery aging trajectories utilizing less data. A two-stage prediction method for battery capacity aging trajectories is proposed. The first stage is battery lifespan prediction based on the Siamese-Convolutional neural network (Siamese-CNN). Taking the predicted lifespan as its prior information, the second stage is predicting capacity aging trajectories with the Convolutional neural network (CNN). The method only requires partial voltage-current data from 30 cycles, of which the similarity in different batteries is obtained by the Siamese-CNN. The lifespan-unknown batteries in the testing set are fed into the Siamese-CNN to match the lifespan-known training set battery with the highest similarity. Then, CNN is employed to predict the capacity aging trajectories under different profiles. The ratio of each cell's predicted lifespan to the fixed length of the CNN output sequence is regarded as the interval for uniform sampling to keep the series consistent. The prediction accuracy of the method is validated by an open dataset. It is demonstrated that the battery lifespan and capacity aging trajectory prediction achieve a mean absolute percentage error (MAPE) of 2.44% and 1.28%, respectively.

A CNN-based denoising method trained with images acquired with electron density phantoms for thin-sliced coronary artery calcium scans.

This work proposed a convolutional neural network (CNN)-based method trained with images acquired with electron density phantoms to reduce quantum noise for coronary artery calcium (CAC) scans reconstructed with slice thickness less than 3mm. A DenseNet model was used to estimate quantum noise for CAC scans reconstructed with slice thickness of 0.5, 1.0 and 1.5mm. Training data was acquired using electron density phantoms in three different sizes. The label images of the CNN model were real noise maps, while the input images of the CNN model were pseudo noise maps. Image denoising was conducted by subtracting the CNN output images from thin-sliced CAC scans. The efficacy of the proposed method was verified through both phantom study and patient study. By means of phantom study, the proposed method was proven effective in reducing quantum noise in CAC scans reconstructed with 1.5-mm slice thickness without causing significant texture change or variation in HU values. With regard to patient study, calcifications were more clear on the denoised CAC scans reconstructed with slice thickness of 0.5, 1.0 and 1.5mm than on 3-mm slice images, while over-smooth changes were not observed in the denoised CAC scans reconstructed with 1.5-mm slice thickness. Our results demonstrated that the electron density phantoms can be used to generate training data for the proposed CNN-based denoising method to reduce quantum noise for CAC scans reconstructed with 1.5-mm slice thickness. Because anthropomorphic phantom is not a necessity, our method could make image denoising more practical in routine clinical practice.

BEYOND TRADITIONAL TOOLS: EXPLORING CONVOLUTIONAL NEURAL NETWORKS AS INNOVATIVE PROGNOSTIC MODELS IN PANCREATIC DUCTAL ADENOCARCINOMA.

Pancreatic ductal adenocarcinoma (PDAC) is a highly aggressive and lethal form of cancer with limited prognostic accuracy using traditional factors. This has led to the exploration of innovative prognostic models, including convolutional neural networks (CNNs), in PDAC. CNNs, a type of artificial intelligence algorithm, have shown promise in various medical applications, including image analysis and pattern recognition. Their ability to extract complex features from medical images makes them suitable for improving prognostication in PDAC. However, implementing CNNs in clinical practice poses challenges, such as data availability and interpretability. Future research should focus on multi-center studies, integrating multiple data modalities, and combining CNN outputs with biomarker panels. Collaborative efforts and patient autonomy should be considered to ensure the ethical implementation of CNN-based prognostic models. Further validation and optimisation of CNN-based models are necessary to enhance their reliability and clinical utility in PDAC prognostication. •Pancreatic ductal adenocarcinoma (PDAC) is an aggressive cancer with limited prognostic accuracy through traditional methods. •Convolutional neural networks (CNNs) are being explored for prognostic models in PDAC. •They can extract complex features from images, aiding PDAC prognostication. •Further validation and optimization of CNN-based models are needed for better reliability and clinical utility in PDAC.

Generative adversarial deep learning in images using Nash equilibrium game theory

A generative adversarial learning (GAL) algorithm is presented to overcome the manipulations that take place in adversarial data and to result in a secured convolutional neural network (CNN). The main objective of the generative algorithm is to make some changes to initial data with positive and negative class labels in testing, hence the CNN results in misclassified data. An adversarial algorithm is used to manipulate the input data that represents the boundaries of learner’s decision-making process. The algorithm generates adversarial modifications to the test dataset using a multiplayer stochastic game approach, without learning how to manipulate the data during training. Then the manipulated data is passed through a CNN for evaluation. The multi-player game consists of an interaction between adversaries which generates manipulations and retrains the model by the learner. The Nash equilibrium game theory (NEGT) is applied to Canadian Institute for Advance Research (CIFAR) dataset. This was done to produce a secure CNN output that is more robust to adversarial data manipulations. The experimental results show that proposed NEGT-GAL achieved a grater mean value of 7.92 and takes less wall clock time of 25,243 sec. Therefore, the proposed NEGT-GAL outperforms the compared existing methods and achieves greater performance.

Automated classification of hidradenitis suppurativa disease severity by convolutional neural network analyses using calibrated clinical images.

The assessment of hidradenitis suppurativa (HS) severity requires detailed, and error-prone lesion counts. This proof-of-concept study aimed to automatically classify HS disease severity using machine learning of clinical smartphone images. 777 ambient-light and size-controlled images were used to build a class-balanced synthetic dataset (n = 7675). Convolutional neural networks (CNN) were used for automated severity classification (scale 0-3), and to assess disease-dynamics. International Hidradenitis Suppurativa Severity Score System (IHS4) served as reference. A U-NET algorithm was implemented for automated localization of diseased skin. CNNs were able to distinguish no/mild from moderate/severe disease with an overall prediction accuracy of 78% [receiver operating curve (AUC) 0.85]. Correct IHS4 classification was achieved with an overall accuracy of 72% (AUC 0.84-0.89). In addition, disease dynamics using IHS4 numerical values aligned with CNN outputs (NRMSE 0.262). The UNET algorithm localized lesions with a pixel accuracy of 88.1% and test loss of 0.42. Limitations in assessing tattooed and hairy skin. Limited number of patients with dark skin colour and Hurley I. CNNs were able to distinguish no/mild from moderate/severe disease, classify disease severity over time, and automatically identify diseased skin areas and the skin phototype. This study breaks new grounds for fast, reliable, reproducible and easy-to-use HS severity assessments using clinical images.

New method of using a convolutional neural network for 2D intraprostatic tumor segmentation from PET images

PurposeA new method of using a convolutional neural network (CNN) to perform automatic tumor segmentation from two-dimensional transaxial slices of positron emission tomography (PET) images of high-risk primary prostate cancer patients is introduced.MethodsWe compare three different methods including (1) usual image segmentation with a CNN whose continuous output is converted to binary labels with a constant threshold, (2) our new technique of choosing separate thresholds for each image PET slice with a CNN to label the pixels directly from the PET slices, and (3) the combination of the two former methods based on using the second CNN to choose the optimal thresholds to convert the output of the first CNN. The CNNs are trained and tested multiple times by using a data set of 864 slices from the PET images of 78 prostate cancer patients.ResultsAccording to our results, the Dice scores computed from the predictions of the second method are statistically higher than those of the typical image segmentation (p-value<0.002).ConclusionThe new method of choosing unique thresholds to convert the pixels of the PET slices directly into binary tumor masks is not only faster and more computationally efficient but also yields better results.

Poster Session: Eye-tracking while learning greebles.

There has been a lot of attention on so-called 'adversarial images' that fool machine learning models (MLM). Manipulations to an image or including an unexpected object in a scene can severely disrupt object labeling and parsing by state-of-the-art MLM such as convolutional neural networks (CNN). In contrast, human observers may not even notice these changes, highlighting a significant gap between the robustness of human vision and CNNs. One well-studied class of novel objects are 'Greebles,' introduced by Gauthier and Tarr (1997; Gauthier et al., 1998). A large number of 'families', 'genders' and individuals can be created by systematically varying the arrangements and shapes of Greeble components. We report on eye-tracking (SMI iRed 250) participants as they are taught to recognize Greebles following the procedure described by Gauthier, Tarr and colleagues. Greeble expertise is assessed using a naming task and a verification task. Once an individual became a Greeble 'expert' we assessed their tolerance for rotations of the Greebles around the y-axis. The pattern of eye movements when learning and being tested on the Greebles are compared to the output of CNNs trained to recognize Greebles.

SeriesSleepNet: an EEG time series model with partial data augmentation for automatic sleep stage scoring.

Introduction: We propose an automatic sleep stage scoring model, referred to as SeriesSleepNet, based on convolutional neural network (CNN) and bidirectional long short-term memory (bi-LSTM) with partial data augmentation. We used single-channel raw electroencephalography signals for automatic sleep stage scoring. Methods: Our framework was focused on time series information, so we applied partial data augmentation to learn the connected time information in small series. In specific, the CNN module learns the time information of one epoch (intra-epoch) whereas the bi-LSTM trains the sequential information between the adjacent epochs (inter-epoch). Note that the input of the bi-LSTM is the augmented CNN output. Moreover, the proposed loss function was used to fine-tune the model by providing additional weights. To validate the proposed framework, we conducted two experiments using the Sleep-EDF and SHHS datasets. Results and Discussion: The results achieved an overall accuracy of 0.87 and 0.84 and overall F1-score of 0.80 and 0.78 and kappa value of 0.81 and 0.78 for five-class classification, respectively. We showed that the SeriesSleepNet was superior to the baselines based on each component in the proposed framework. Our architecture also outperformed the state-of-the-art methods with overall F1-score, accuracy, and kappa value. Our framework could provide information on sleep disorders or quality of sleep to automatically classify sleep stages with high performance.

Development of a Human-Robot Interaction System for Industrial Applications

The use of robots is increasing day by day. In this study, it was aimed to develop manufacturing-assistant robot software for small production plants involving non-mass production. The main purpose of this study is to eliminate the difficulty of recruiting an expert robot operator thanks to the developed software and to facilitate the use of robots for non-experts. The developed software consists of three parts: the convolutional neural network (CNN), process selection-trajectory generation, and trajectory regulation modules. Before the operations in these modules are executed, operators record the desired process and the trajectory of the process in the video by hand gestures and index finger. Then recorded video is separated into images. The separated images are classified by the CNN module and the positions of landmarks (joint and fingernail of index finger) were calculated by the same module and using the images. Eight different pre-trained CNN structures were tested in the CNN module, and the best result Xception structure (test loss = 0.0051) was used. The desired process was determined and the trajectory of the process was created with the CNN output data. The connection of the generated trajectory with the object was detected by the trajectory regulation module, and unnecessary trajectory parts were cleaned. Regulated trajectory and desired tasks such as welding or sealing were simulated via an industrial robot in a simulation environment. As a result, an industrial robot could be programmed by non-expert operators for companies whose production line is not standard by using the developed software.

An optimized hybrid deep learning model based on word embeddings and statistical features for extractive summarization

Extractive summarization has recently gained significant attention as a classification problem at the sentence level. Most current summarization methods rely on only one way of representing sentences in a document (i.e., extracted features, word embeddings, BERT embeddings). However, classification performance and summary generation quality will be improved if we combine two ways of representing sentences. This paper presents a novel extractive text summarization method based on word embeddings and statistical features of a single document. Each sentence is encoded using a Convolutional Neural Network (CNN) and a Feed-Forward Neural Network (FFNN) based on word embeddings and statistical features. CNN and FFNN outputs are concatenated to classify the sentence using a Multilayer Perceptron (MLP). In addition, hybrid model parameters are optimized by the KerasTuner optimization technique to determine the most efficient hybrid model. The proposed method was evaluated on the standard Newsroom dataset. Experiments show that the proposed method effectively captures the document’s semantic and statistical information and outperforms deep learning, machine learning, and state-of-the-art approaches with scores of 78.64, 74.05, and 72.08 for ROUGE-1 ROUGE-2, and ROUGE-L, respectively.

Topological Structure and Semantic Information Transfer Network for Cross-Scene Hyperspectral Image Classification.

Domain adaptation techniques have been widely applied to the problem of cross-scene hyperspectral image (HSI) classification. Most existing methods use convolutional neural networks (CNNs) to extract statistical features from data and often neglect the potential topological structure information between different land cover classes. CNN-based approaches generally only model the local spatial relationships of the samples, which largely limits their ability to capture the nonlocal topological relationship that would better represent the underlying data structure of HSI. In order to make up for the above shortcomings, a Topological structure and Semantic information Transfer network (TSTnet) is developed. The method employs the graph structure to characterize topological relationships and the graph convolutional network (GCN) that is good at processing for cross-scene HSI classification. In the proposed TSTnet, graph optimal transmission (GOT) is used to align topological relationships to assist distribution alignment between the source domain and the target domain based on the maximum mean difference (MMD). Furthermore, subgraphs from the source domain and the target domain are dynamically constructed based on CNN features to take advantage of the discriminative capacity of CNN models that, in turn, improve the robustness of classification. In addition, to better characterize the correlation between distribution alignment and topological relationship alignment, a consistency constraint is enforced to integrate the output of CNN and GCN. Experimental results on three cross-scene HSI datasets demonstrate that the proposed TSTnet performs significantly better than some state-of-the-art domain-adaptive approaches. The codes will be available from the website: https://github.com/YuxiangZhang-BIT/IEEE_TNNLS_TSTnet.

Reducing scan time in 177 Lu planar scintigraphy using convolutional neural network: A Monte Carlo simulation study.

The aim of this study was to reduce scan time in 177 Lu planar scintigraphy through the use of convolutional neural network (CNN) to facilitate personalized dosimetry for 177 Lu-based peptide receptor radionuclide therapy. The CNN model used in this work was based on DenseNet, and the training and testing datasets were generated from Monte Carlo simulation. The CNN input images (IMGinput ) consisted of 177 Lu planar scintigraphy that contained 10-90% of the total photon counts, while the corresponding full-count images (IMG100% ) were used as the CNN label images. Two-sample t-test was conducted to compare the difference in pixel intensities within region of interest between IMG100% and CNN output images (IMGoutput ). No difference was found in IMGoutput for rods with diameters ranging from 13 to 33mm in the Derenzo phantom with a target-to-background ratio of 20:1, while statistically significant differences were found in IMGoutput for the 10-mm diameter rods when IMGinput containing 10% to 60% of the total photon counts were denoised. Statistically significant differences were found in IMGoutput for both right and left kidneys in the NCAT phantom when IMGinput containing 10% of the total photon counts were denoised. No statistically significant differences were found in IMGoutput for any other source organs in the NCAT phantom. Our results showed that the proposed method can reduce scan time by up to 70% for objects larger than 13mm, making it a useful tool for personalized dosimetry in 177 Lu-based peptide receptor radionuclide therapy in clinical practice.

Simulation of unsteady flow around bluff bodies using knowledge-enhanced convolutional neural network

The unsteady flow with massive separation poses challenges to accurately and efficiently simulate wind effects on civil structures, especially in the search for optimal aerodynamic shapes during the preliminary design stage. To this end, an encoding-transformation-decoding convolutional neural network (CNN) architecture is developed in this study to take the pixelated geometry of a bluff body as the input and instantly output the corresponding unsteady aerodynamic flow (represented by a sequence of consecutive flow snapshots). First, a unique (shared) encoding part is used to extract the low-resolution features from the structural shapes and associated initial flow conditions. Then, the transformation part consecutively maps the encoded features (originated from encoder) to a prediction of next flow snapshot in high-level feature space while the separate decoding part unravels the abstract features (from corresponding transformer) to pixelated velocity and pressure fields. Prior domain knowledge is leveraged at various stages of machine learning pipeline to enhance the learning efficiency and accuracy of the proposed convolutional encoder-transformer-decoder. Specifically, the qualitative knowledge that the flow fields vary on a very small length scale within the boundary layer is used to empirically design a nonuniform grid for CNN image inputs and outputs. In addition, the equation-free quantitative knowledge learned from previous flow snapshot (i.e., high-level features from the transformer and filters in the decoder) is employed in the learning of current flow snapshot as the CNN initial conditions based on transfer learning technique. Furthermore, the equation-based quantitative knowledge of both mass and momentum conservation laws is utilized to augment the CNN loss function. Numerical examples suggest good performance of the proposed knowledge-enhanced CNN for unsteady flow simulation of bridge decks. Effects of these three types of knowledge on the performance of the convolutional encoder-transformer-decoder are also examined to highlight their high efficacy to facilitate the training efficiency and simulation accuracy.

Multiaxial fatigue life prediction for various metallic materials based on the hybrid CNN‐LSTM neural network

AbstractA new algorithm optimization‐based hybrid neural network model is proposed in the present study for the multiaxial fatigue life prediction of various metallic materials. Firstly, a convolutional neural network (CNN) is applied to extract the in‐depth features from the loading sequence composed of the critical fatigue loading conditions. Meanwhile, the multiaxial historical loading information with time‐series features is retained. Then, a long short‐term memory (LSTM) network is adopted to capture the time‐series features and in‐depth features of the CNN output. Finally, a full connection layer is used to achieve dimensional transformation, which makes the fatigue life predictable. Herein, the hyperparameters of the LSTM network are automatically determined using the slime mold algorithm (SMA). The test results demonstrate that the proposed model has pleasant prediction performance and extrapolation capability, and it is suitable for the life prediction of various metallic materials under uniaxial, proportional multiaxial, nonproportional multiaxial loading conditions.

Urban Road Traffic Flow Prediction with Attention-Based Convolutional Bidirectional Long Short-Term Memory Networks

Accurate traffic prediction is critical for industry practitioners and researchers in intervening and dredging future traffic in advance to avoid traffic congestion. Considering that most prediction models fail to effectively capture the complex nonlinearity of traffic data and thus cannot obtain satisfactory prediction results, we propose a novel deep-learning architecture for traffic flow prediction, called AC-BLSTM (attention-based convolutional bidirectional long short-term memory). The proposed model captures traffic information through multilayer network architectures composed of convolutional bidirectional long short-term memory (conv-BiLSTM) network and attention mechanism. The spatiotemporal features of traffic flow are extracted by convolutional neural network (CNN) and bidirectional long short-term memory (BiLSTM) network. Then attention mechanism combines the outputs of CNN and BiLSTM to assign corresponding weights to the features extracted at different times. In addition, we employ a parallel sub-module structure to model three temporal properties of traffic flow, that is, weekly, daily, and recent dependencies. Finally, the results of these three parts are fused to predict the traffic flow values through the fully connected (FC) layers. Experiment results using a real-world urban road traffic dataset demonstrate that compared with other competing models, the proposed model has better prediction performance.

Dead pixel test using effective receptive field

Deep neural networks have been used in various fields, but their internal behavior in how they understand images is not well known. In this study, we discuss two counterintuitive properties of convolutional neural networks (CNNs). First, we evaluated the size of the receptive field of CNNs with their classification accuracy. Previous studies have attempted to increase the size of the receptive field for performance gain. However, we observed that some CNNs with a smaller receptive field can achieve higher classification accuracy. In this regard, we claim that a larger receptive field does not guarantee improved classification accuracy. Second, using the effective receptive field, we examined the contribution of each pixel to the output of CNN. Intuitively, each pixel is expected to equally contribute to the final output, but we found that there exist pixels in a partially dead state with little contribution to the output. We reveal that the reason for dead pixels lies in even stride operations with odd-sized kernels in CNN and propose a kernel padding method to remove the dead pixels. We demonstrated the vulnerability of CNNs with dead pixels when we detect a noise or small box that is on dead pixels. Our findings on dead pixels should be understood and considered in practical applications of CNN.

Justifying convolutional neural network with argumentation for explainability

Convolutional neural network (CNN) has emerged as one of the most accurate methods for sentiment analysis, but it is largely uninterpretable, while case-based reasoning (CBR) is less accurate but o ff ers interpretable outputs in the form of arguments from analogy. This paper presents an approach to combine these two methods, CNN for accuracy and CBR for explainability, using an assumption-based argumentation (ABA) framework. Our approach focuses on justifying CNN outputs using analogous sentences from CBR, while ensuring that the combined process is argumentative and hence self-explainable. To demonstrate the proposal, we construct a CNN model M 1 and a CBR model M 2 for sentiment analysis using di ff erent subsets of a dataset of which the remaining part is used for testing and comparing these input models with combined models. For an input sentence, if M 1 and M 2 predict the same sentiment, then the analogous sentence, which M 2 finds, is used to explain the sentiment. If they give conflicting sentiments, a hybrid model M 3 determines which one should be followed using a system of strict rules that takes into account how assertive M 1 and M 2 are. Another hybrid model M 4 , which is implemented by an ABA framework, improves on M 3 by considering the probability distribution of the set of all labels from M 1 , and the second (or third) most similar sentences from M 2 . M 3 and M 4 preserve the accuracy of the CNN model (specifically, 88.32% and 88.28% in comparison with 87.59% accuracy of the CNN). They justify 69.95% and 74.53% of CNN outputs, respectively.