Convolution Kernel Size Research Articles

Sea-Surface Small Floating Target Recurrence Plots FAC Classification Based on CNN

In this paper, we propose the False-Alarm-Rate-Controllable (FAC) classification of sea clutter Recurrence Plots (RPs) based on Convolutional Neural Networks (CNN), which is shortened as RPs-CNN. Sea clutter data is a non-linear and recursive time series, and RPs provide qualitative analysis of non-linear and recursive dynamic systems. Thus, we construct the RPs datasets to extract the recursive feature of sea clutter. In the RPs datasets, RPs parameters, embedding delay τ and embedding dimension m, are obtained by the average mutual information (AMI) and false nearest neighbor (FNN) algorithms, respectively. In addition, in order to extract the local features of RPs, CNN is applied. CNN makes full use of local features of the datasets, has the advantages of translation invariance and strong generalization ability, and shares the available weights to simplify network structure. We implement a proper LeNet-5 CNN training on the constructed RPs datasets, and verity it by IPIX measured datasets. The experimental results demonstrate that the CNN successfully classifies the RPs of targets and clutter. Moreover, the feasibility of RPs-CNN is verified by seven aspects, i.e., Accuracy, Precision, False Alarm, Miss Rate, Recall, F1-measure and Kappa. In addition, six parameters that may affect the classification performance are also analyzed, including time-series length, embedding delay, embedding dimension, convolutional kernel size, kernel depth, and optimization function. The results indicate that the proposed RPs-CNN method can reach a 92.05% F1-measure and 87.19% Kappa which performs better than other classification methods. Meanwhile, the false alarm rate (FAR) of the RPs-CNN is testified by the FAC classification. Experimental results demonstrate that the proposed RPs-CNN significantly improves the detection probability over other classification detectors in low FAR cases.

Application of CNN and Long Short-Term Memory Network in Water Quality Predicting

Water resources are an indispensable precious resource for human survival and development. Water quality prediction plays a vital role in protecting and enhancing water resources. Changes in water quality are influenced by many factors, both long-term and short-term. Therefore, according to water quality changes’ periodic and nonlinear characteristics, this paper considered dissolved oxygen as the research object and constructed a neural network model combining convolutional neural network (CNN) and long short-term memory network (LSTM) to predict dissolved oxygen index in water quality. Firstly, we preprocessed the water quality data set obtained from the water quality monitoring platform. Secondly, we used a CNN network to extract local features from the preprocessed water quality data and transferred time series with better expressive power than the original water quality information to the LSTM layer for prediction. We choose optimal parameters by setting the number of neurons in the LSTM network and the size and number of convolution kernels in the CNN network. Finally, LSTM and the proposed model were used to evaluate the water quality data. Experiments showed that the proposed model is more accurate than the conventional LSTM in the prediction effect of peak fitting. Compared with the conventional LSTM model, its root mean square error, Pearson correlation coefficient, mean absolute error and mean square error were respectively optimized by 5.99%, 2.80%, 2.24%, and 11.63%.

Multigranularity Feature Fusion Convolutional Neural Network for Seismic Data Denoising

Seismic data denoising is an important part of seismic data processing and has attracted much attention in recent years. With the rapid development of neural networks, convolutional neural network (CNN)-based denoising methods have been widely studied and used in seismic data denoising due to their unique convolutional layer and weight sharing characteristics. However, the existing CNN-based seismic data denoising methods mainly use fixed-size convolution kernels in certain layer, which forces the different kernels to extract features from areas of the same size and fails to extract features from various granularities. To overcome this limitation, we make full use of the local similarity of seismic sections, use different sizes of convolution kernels to parallelly extract features from different granularities and propose a multigranularity feature fusion-convolutional neural network (MFFCNN) method to remove random noise from seismic data. This method uses convolution kernels with different sizes to extract features from various granularities and uses feature fusion structures to fuse the extracted features. Experimental results show that the MFFCNN proposed in this paper can better deal with details and texture information than the compared methods.

HCRB-MSAN: Horizontally Connected Residual Blocks-Based Multiscale Attention Network for Semantic Segmentation of Buildings in HSR Remote Sensing Images

Accurate and efficient semantic segmentation of buildings in high spatial resolution (HSR) remote sensing images is the basis for applications such as fine urban management, high-precision mapping, land resource utilization investigation, and human settlement suitability evaluation. The current building extraction methods based on deep learning can obtain high-level abstract features of images. However, due to the limitation of convolution kernel size and the vanishing gradient, the extraction of some buildings is inaccurate, and some small-volume buildings are missing as the network deepens. In this regard, we design a horizontally connected residual blocks-based multi-scale attention network (HCRB-MSAN) to achieve high-quality extraction of buildings in HSR remote sensing image. In this network, we subdivide each residual block by channel grouping and feature horizontal connection to consider the difference and saliency of feature information between channels, and then combine the output features with multi-scale attention module to consider the contextual semantic relationship of different regions and integrate multi-level local and global information of buildings. A stepwise up-sampling mechanism is designed in the decoding process to finally achieve precise semantic segmentation of buildings. We conduct experiments on two public datasets and compare the proposed method with state-of-the-art semantic segmentation methods. The experiments show that our method could achieve better building extraction results in HSR remote sensing image, which proves the effectiveness of our proposed method.

Classification of Diabetic Retinopathy Severity Based on GCA Attention Mechanism

Diabetic retinopathy (DR) is one of the major complications caused by diabetes and can lead to severe vision loss or even complete blindness if not diagnosed and treated in a timely manner. In this paper, a new feature map global channel attention mechanism (GCA) is proposed to solve the problem of the early detection of DR. In the GCA module, an adaptive one-dimensional convolution kernel size algorithm based on the dimension of the feature map is proposed and a deep convolutional neural network model for DR color medical image severity diagnosis named GCA-EfficientNet (GENet) is designed. The training process uses transfer learning techniques with a cosine annealing learning rate adjustment strategy. The image regions of interest of GENet are visualized using a heat map. The final accuracy, precision, sensitivity and specificity of the DR dataset of the Kaggle competition reached 0.956, 0.956, 0.956, and 0.989, respectively. A large number of experiment results show that GENet based on the GCA attention mechanism can more effectively extract lesion features and classify the severity of DR.

ECG signal classification based on deep CNN and BiLSTM

BackgroundCurrently, cardiovascular disease has become a major disease endangering human health, and the number of such patients is growing. Electrocardiogram (ECG) is an important basis for {medical doctors to diagnose the cardiovascular disease, which can truly reflect the health of the heart. In this context, the contradiction between the lack of medical resources and the surge in the number of patients has become increasingly prominent. The use of computer-aided diagnosis of cardiovascular disease has become particularly important, so the study of ECG automatic classification method has a strong practical significance.MethodsThis article proposes a new method for automatic identification and classification of ECG.We have developed a dense heart rhythm network that combines a 24-layer Deep Convolutional Neural Network (DCNN) and Bidirectional Long Short-Term Memory (BiLSTM) to deeply mine the hierarchical and time-sensitive features of ECG data. Three different sizes of convolution kernels (32, 64 and 128) are used to mine the detailed features of the ECG signal, and the original ECG is filtered using a combination of wavelet transform and median filtering to eliminate the influence of noise on the signal. A new loss function is proposed to control the fluctuation of loss during the training process, and convergence mapping of the tan function in the range of 0–1 is employed to better reflect the model training loss and correct the optimization direction in time.ResultsWe applied the dataset provided by the 2017 PhysioNet/CINC challenge for verification. The experiment adopted ten-fold cross validation,and obtained an accuracy rate of 89.3% and an F1 score of 0.891.ConclusionsThis article proposes its own method in the aspects of ECG data preprocessing, feature extraction and loss function design. Compared with the existing methods, this method improves the accuracy of automatic ECG classification and is helpful for clinical diagnosis and self-monitoring of atrial fibrillation.

Nonlinear Hyperparameter Optimization of a Neural Network in Image Processing for Micromachines.

Deep neural networks are widely used in the field of image processing for micromachines, such as in 3D shape detection in microelectronic high-speed dispensing and object detection in microrobots. It is already known that hyperparameters and their interactions impact neural network model performance. Taking advantage of the mathematical correlations between hyperparameters and the corresponding deep learning model to adjust hyperparameters intelligently is the key to obtaining an optimal solution from a deep neural network model. Leveraging these correlations is also significant for unlocking the “black box” of deep learning by revealing the mechanism of its mathematical principle. However, there is no complete system for studying the combination of mathematical derivation and experimental verification methods to quantify the impacts of hyperparameters on the performances of deep learning models. Therefore, in this paper, the authors analyzed the mathematical relationships among four hyperparameters: the learning rate, batch size, dropout rate, and convolution kernel size. A generalized multiparameter mathematical correlation model was also established, which showed that the interaction between these hyperparameters played an important role in the neural network’s performance. Different experiments were verified by running convolutional neural network algorithms to validate the proposal on the MNIST dataset. Notably, this research can help establish a universal multiparameter mathematical correlation model to guide the deep learning parameter adjustment process.

Deep-Learning-Based Remaining Useful Life Prediction Based on a Multi-Scale Dilated Convolution Network

Remaining useful life (RUL) prediction of key components is an important influencing factor in making accurate maintenance decisions for mechanical systems. With the rapid development of deep learning (DL) techniques, the research on RUL prediction based on the data-driven model is increasingly widespread. Compared with the conventional convolution neural networks (CNNs), the multi-scale CNNs can extract different-scale feature information, which exhibits a better performance in the RUL prediction. However, the existing multi-scale CNNs employ multiple convolution kernels with different sizes to construct the network framework. There are two main shortcomings of this approach: (1) the convolution operation based on multiple size convolution kernels requires enormous computation and has a low operational efficiency, which severely restricts its application in practical engineering. (2) The convolutional layer with a large size convolution kernel needs a mass of weight parameters, leading to a dramatic increase in the network training time and making it prone to overfitting in the case of small datasets. To address the above issues, a multi-scale dilated convolution network (MsDCN) is proposed for RUL prediction in this article. The MsDCN adopts a new multi-scale dilation convolution fusion unit (MsDCFU), in which the multi-scale network framework is composed of convolution operations with different dilated factors. This effectively expands the range of receptive field (RF) for the convolution kernel without an additional computational burden. Moreover, the MsDCFU employs the depthwise separable convolution (DSC) to further improve the operational efficiency of the prognostics model. Finally, the proposed method was validated with the accelerated degradation test data of rolling element bearings (REBs). The experimental results demonstrate that the proposed MSDCN has a higher RUL prediction accuracy compared to some typical CNNs and better operational efficiency than the existing multi-scale CNNs based on different convolution kernel sizes.

A surface cracks detection method for nuclear fuel pellets using an improved fully convolutional network

ABSTRACT Surface cracks, one of the main quality defects in nuclear fuel pellets, are usually subtle and hidden in complex backgrounds, making them difficult to detect accurately and threatening reactor safety. In order to achieve accurate cracks detection, a fuel pellet crack detection (FPCD) network is proposed based on the fully convolutional network with the VGG16 network structure as the encoder’s backbone (FCN-VGG16). The new architecture modifies the combination mode of convolution layers and pooling layers, and optimizes the size and number of convolution kernels. And a large weight is added for crack samples in the loss function to solve the class imbalance problem. Subsequently, a self-acquired dataset containing 2240 pellet surface crack images is used for training and performance evaluation of networks. Experiments show that the proposed method achieved an F1-score of 88.89% and an intersection over union (IoU) of 80.30%, obtaining better detection results for complex background crack images.

A Hybrid GA–PSO–CNN Model for Ultra-Short-Term Wind Power Forecasting

Accurate and timely wind power forecasting is essential for achieving large-scale wind power grid integration and ensuring the safe and stable operation of the power system. For overcoming the inaccuracy of wind power forecasting caused by randomness and volatility, this study proposes a hybrid convolutional neural network (CNN) model (GA–PSO–CNN) integrating genetic algorithm (GA) and a particle swarm optimization (PSO). The model can establish feature maps between factors affecting wind power such as wind speed, wind direction, and temperature. Moreover, a mix-encoding GA–PSO algorithm is introduced to optimize the network hyperparameters and weights collaboratively, which solves the problem of subjective determination of the optimal network in the CNN and effectively prevents local optimization in the training process. The prediction effectiveness of the proposed model is verified using data from a wind farm in Ningxia, China. The results show that the MAE, MSE, and MAPE of the proposed GA–PSO–CNN model decreased by 1.13–9.55%, 0.46–7.98%, and 3.28–19.29%, respectively, in different seasons, compared with Single–CNN, PSO–CNN, ISSO–CNN, and CHACNN models. The convolution kernel size and number in each convolution layer were reduced by 5–18.4% in the GA–PSO–CNN model.

NL-VTON: a non-local virtual try-on network with feature preserving of body and clothes

In an image based virtual try-on network, both features of the target clothes and the input human body should be preserved. However, current techniques failed to solve the problems of blurriness on complex clothes details and artifacts on human body occlusion regions at the same time. To tackle this issue, we propose a non-local virtual try-on network NL-VTON. Considering that convolution is a local operation and limited by its convolution kernel size and rectangular receptive field, which is unsuitable for large size non-rigid transformations of persons and clothes in virtual try-on, we introduce a non-local feature attention module and a grid regularization loss so as to capture detailed features of complex clothes, and design a human body segmentation prediction network to further alleviate the artifacts on occlusion regions. The quantitative and qualitative experiments based on the Zalando dataset demonstrate that our proposed method significantly improves the ability to preserve features of bodies and clothes compared with the state-of-the-art methods.

A Novel Anti-Noise Fault Diagnosis Approach for Rolling Bearings Based on Convolutional Neural Network Fusing Frequency Domain Feature Matching Algorithm.

The development of deep learning provides a new research method for fault diagnosis. However, in the industrial field, the labeled samples are insufficient and the noise interference is strong so that raw data obtained by the sensor are occupied with noise signal. It is difficult to recognize time-domain fault signals under the severe noise environment. In order to solve these problems, the convolutional neural network (CNN) fusing frequency domain feature matching algorithm (FDFM), called CNN-FDFM, is proposed in this paper. FDFM extracts key frequency features from signals in the frequency domain, which can maintain high accuracy in the case of strong noise and limited samples. CNN automatically extracts features from time-domain signals, and by using dropout to simulate noise input and increasing the size of the first-layer convolutional kernel, the anti-noise ability of the network is improved. Softmax with temperature parameter T and D-S evidence theory are used to fuse the two models. As FDFM and CNN can provide different diagnostic information in frequency domain, and time domain, respectively, the fused model CNN-FDFM achieves higher accuracy under severe noise environment. In the experiment, when a signal-to-noise ratio (SNR) drops to -10 dB, the diagnosis accuracy of CNN-FDFM still reaches 93.33%, higher than CNN’s accuracy of 45.43%. Besides, when SNR is greater than -6 dB, the accuracy of CNN-FDFM is higher than 99%.

Extreme Image Classification Algorithm Based on Multicore Dense Connection Network

Extreme images refer to low-quality images taken under extreme environmental conditions such as haze, heavy rain, strong light, and shaking, which will lead to the failure of the visual system to effectively recognize the target. Most of the existing extreme image restoration algorithms only handle the restoration work of a certain type of image; how to effectively recognize all kinds of extreme images is still a challenge. Therefore, this paper proposes a classification and restoration algorithm for extreme images. Due to the large differences in the features on extreme images, it is difficult for the existing models such as DenseNet to effectively extract depth features. In order to solve the classification problem in the algorithm, we propose a Multicore Dense Connection Network (MDCNet). MDCNet is composed of dense part, attention part, and classification part. Dense Part uses two dense blocks with different convolution kernel sizes to extract features of different sizes; attention part uses channel attention mechanism and spatial attention mechanism to amplify the effective information in the feature map; classification part is mainly composed of two convolutional layers and two fully connected layers to extract and classify feature images. Experiments have shown that the recall of MDCNet can reach 92.75% on extreme image dataset. At the same time, the mAP value of target detection can be improved by about 16% after the image is processed by the classification and recovery algorithm.

DeepnsSNPs: Accurate prediction of non-synonymous single-nucleotide polymorphisms by combining multi-scale convolutional neural network and residue environment information

Non-synonymous single-nucleotide polymorphisms (nsSNPs) is a typical kind of genetic variant, and more than 6000 diseases have been detected to be caused by nsSNPs. Accordingly, the accurate prediction of nsSNPs is of great importance for a better understanding of their functional mechanisms and disease treatment. Till now, many computational studies have been developed to identify disease-causing nsSNPs from the neutral ones; however, there is still some gap existing for further improvement in terms of overall prediction performance. In this work, we proposed a novel deep learning model, called multi-scale convolutional neural network (MSCNN). It utilized multi-scale convolution with different kernel sizes for feature processing, which can collect more effective characteristics than using a single convolution kernel size. Moreover, we applied three types of nominal structural features for further improving the nsSNPs prediction performance. Notably, the nsSNPs sequence and structural features were extracted based on the “residue environment” method we proposed, which has proved to be effective for protein nsSNPs prediction in our previous research. Based on the proposed MSCNN model and the extracted informative feature matrix, we implemented a new nsSNPs predictor, named DeepnsSNPs. The DeepnsSNPs was tested on three nsSNPs datasets collected from the PredictSNP1 website and achieved an average Matthews correlation coefficient of 0.507, which is 18.28% higher than the individual classifiers and 11.37% higher than the consensus classifier on average. Detailed dataset analyses have demonstrated that the DeepnsSNPs would be useful in the nsSNPs prediction. We provide the source python codes and benchmark datasets at https://github.com/sera616/DeepnsSNPs.git for academic use.

SEMG-Based Continuous Estimation of Finger Kinematics via Large-Scale Temporal Convolutional Network

Since continuous motion control can provide a more natural, fast and accurate man–machine interface than that of discrete motion control, it has been widely used in human–robot cooperation (HRC). Among various biological signals, the surface electromyogram (sEMG)—the signal of actions potential superimposed on the surface of the skin containing the temporal and spatial information—is one of the best signals with which to extract human motion intentions. However, most of the current sEMG control methods can only perform discrete motion estimation, and thus fail to meet the requirements of continuous motion estimation. In this paper, we propose a novel method that applies a temporal convolutional network (TCN) to sEMG-based continuous estimation. After analyzing the relationship between the convolutional kernel’s size and the lengths of atomic segments (defined in this paper), we propose a large-scale temporal convolutional network (LS-TCN) to overcome the TCN’s problem: that it is difficult to fully extract the sEMG’s temporal features. When applying our proposed LS-TCN with a convolutional kernel size of 1 × 31 to continuously estimate the angles of the 10 main joints of fingers (based on the public dataset Ninapro), it can achieve a precision rate of 71.6%. Compared with TCN (kernel size of 1 × 3), LS-TCN (kernel size of 1 × 31) improves the precision rate by 6.6%.

A CNN identified by reinforcement learning-based optimization framework for EEG-based state evaluation

Objective. Electroencephalogram (EEG) data, as a kind of complex time-series, is one of the most widely-used information measurements for evaluating human psychophysiological states. Recently, numerous works applied deep learning techniques, especially the convolutional neural network (CNN), into EEG-based research. The design of the hyper-parameters of the CNN model has a great influence on the performance of the model. Therefore, automatically designing these hyper-parameters can save the time and labor of experts. This leads to the appearance of the neural architecture search technique. In this paper, we propose a reinforcement learning (RL)-based step-by-step framework to efficiently search for CNN models. Approach. Specifically, the deep Q network in RL is first used to determine the depth of convolutional layers and the connection modes among layers. Then particle swarm optimization algorithm is used to fine-tune the number and size of convolution kernels. Through this step-by-step strategy, the search space can be narrowed in each step for saving the overall time cost. This framework is employed for both EEG-based sleep stage classification and driver drowsiness evaluation tasks. Main results. The results show that compared with state-of-the-art methods, the high-performance CNN models identified by the proposed optimization framework, can achieve high overall accuracy and better root mean squared error in the two tasks. Significance. Therefore, the proposed optimization framework has a great potential to provide high-performance results for other kinds of classification and prediction tasks. In this way, it can greatly save researchers’ time cost and promote broader applications of CNNs.

Multiscale Random Convolution Broad Learning System for Hyperspectral Image Classification

Deep learning is a powerful technique for image processing. Convolution neural network (CNN) is one of the widely used approaches for hyperspectral image (HSI) classification. These methods mostly need a time-consuming pretraining process to obtain deep features. Random patches networks (RPNets) provide a novel approach that the convolution kernel can be the original image without any pretraining process. In this letter, we propose a novel HSI classification method, multiscale random convolution broad learning system (MRC-BLS), which takes the spatial feature learning by an adaptive weighted mean filter as the convolution kernel to extract local spatial feature in the first layer. Different sizes of random convolution kernels can obtain a multiscale feature map. The weighted fusion of multiscale spatial features extracted by different sizes kernels can get better performance in HSI classification. A broad learning system (BLS) is an efficient classifier to classify images by the multiscale random feature. Experiments in three HSI data sets fully testify to the efficiency and satisfactory performance of the proposed method.

Ultra-lightweight CNN design based on neural architecture search and knowledge distillation: A novel method to build the automatic recognition model of space target ISAR images

In this paper, a novel method of ultra-lightweight convolution neural network (CNN) design based on neural architecture search (NAS) and knowledge distillation (KD) is proposed. It can realize the automatic construction of the space target inverse synthetic aperture radar (ISAR) image recognition model with ultra-lightweight and high accuracy. This method introduces the NAS method into the radar image recognition for the first time, which solves the time-consuming and labor-consuming problems in the artificial design of the space target ISAR image automatic recognition model (STIIARM). On this basis, the NAS model's knowledge is transferred to the student model with lower computational complexity by the flow of the solution procedure (FSP) distillation method. Thus, the decline of recognition accuracy caused by the direct compression of model structural parameters can be effectively avoided, and the ultra-lightweight STIIARM can be obtained. In the method, the Inverted Linear Bottleneck (ILB) and Inverted Residual Block (IRB) are firstly taken as each block's basic structure in CNN. And the expansion ratio, output filter size, number of IRBs, and convolution kernel size are set as the search parameters to construct a hierarchical decomposition search space. Then, the recognition accuracy and computational complexity are taken as the objective function and constraint conditions, respectively, and the global optimization model of the CNN architecture search is established. Next, the simulated annealing (SA) algorithm is used as the search strategy to search out the lightweight and high accuracy STIIARM directly. After that, based on the three principles of similar block structure, the same corresponding channel number, and the minimum computational complexity, the more lightweight student model is designed, and the FSP matrix pairing between the NAS model and student model is completed. Finally, by minimizing the loss between the FSP matrix pairs of the NAS model and student model, the student model's weight adjustment is completed. Thus the ultra-lightweight and high accuracy STIIARM is obtained. The proposed method's effectiveness is verified by the simulation experiments on the ISAR image dataset of five types of space targets.

Performance for rotor system of hybrid electromagnetic bearing and elastic foil gas bearing with dynamic characteristics analysis under deep learning.

The bearing-rotor system is prone to faults during operation, so it is necessary to analyze the dynamic characteristics of the bearing-rotor system to discuss the optimal structure of the convolutional neural network (CNN) in system fault detection and classification. The turbo expander is undertaken as the research object. Firstly, the hybrid magnetic bearing-rotor system is modeled into the form of four stiffness coefficients and four damping coefficients, so as to analyze and explain the dynamic characteristics of the system. Secondly, the ambient pressure is introduced to analyze the dynamic characteristics of the elastic foil gas bearing-rotor system based on the changes in the dynamic stiffness and dynamic damping of the gas bearing. Finally, the CNN is introduced to be applied in the detection of faults of bearing-rotor system through determining the parameters of the constructed CNN. The results show that the displacement of the rotor increases and the stiffness decreases with the acceleration of the speed of the electromagnetic bearing. The maximum displacement of the rotor can reach 135μm, and the maximum stiffness can be reduced to 35×105N/m. Increase of ambient pressure causes enhancement of main stiffness of the gas bearing, and the main damping decreases accordingly. Analysis of the classification accuracy and loss function based on the CNN model shows that the convolution kernel size of 7*1 and the batch size of 128 can realize the best performance of CNN in fault classification. This provides a data support and reference for studying the dynamic characteristics of the bearing-rotor system and for the optimization of CNN structure in fault classification and detection.

Computer-Aided Diagnosis Research of a Lung Tumor Based on a Deep Convolutional Neural Network and Global Features.

Based on the better generalization ability and the feature learning ability of the deep convolutional neural network, it is very significant to use the DCNN on the computer-aided diagnosis of a lung tumor. Firstly, a deep convolutional neural network was constructed according to the fuzzy characteristics and the complexity of lung CT images. Secondly, the relation between model parameters (iterations, different resolution) and recognition rate is discussed. Thirdly, the effects of different model structures for the identification of a lung tumor were analyzed by changing convolution kernel size, feature dimension, and depth of the network. Fourthly, the different optimization methods on how to influence the DCNN performance were discussed from three aspects containing pooling methods (maximum pooling and mean pooling), activation function (sigmoid and ReLU), and training algorithm (batch gradient descent and gradient descent with momentum). Finally, the experimental results verified the feasibility of DCNN used on computer-aided diagnosis of lung tumors, and it can achieve a good recognition rate when selecting the appropriate model parameters and model structure and using the method of gradient descent with momentum.