Convolution Kernel Size Research Articles

A deep learning model for steel surface defect detection

Industrial defect detection is a hot topic in the field of computer vision. It is a challenging task due to complex features and many categories of industrial defects. In this paper, a deep learning model based on the multiscale feature extraction module is introduced for steel surface defect detection. The main focus on the feature extraction capability of the model and feature fusion capability to improve the accuracy of the model for steel surface defect detection. First, to improve the feature extraction ability of the model, a multiscale feature extraction (MSFE) module is introduced. The MSFE module can effectively extract multiscale features through three branches that have different convolution kernel sizes. Second, an efficient feature fusion (EFF) module is proposed to optimize feature fusion by adding features from the backbone network to the neck network. Third, this paper puts forward a new Bottleneck module by reducing the normalization layer and activation function in the original Bottleneck module. Finally, the backbone network is deepened to further enhance the feature extraction ability of the model. Extensive experiments are conducted on the public NEU-DET dataset. The experimental results validate the effectiveness of the designed modules and the proposed model. Compared with other state-of-the-art methods, the proposed model achieves optimal accuracy(73.08% mAP@0.5) while maintaining a small number of parameters.

Research on Mechanical Equipment Fault Diagnosis Method Based on Deep Learning and Information Fusion.

Solving the problem of the transmission of mechanical equipment is complicated, and the interconnection between equipment components in a complex industrial environment can easily lead to faults. A multi-scale-sensor information fusion method is proposed, overcoming the shortcomings of fault diagnosis methods based on the analysis of one signal, in terms of diagnosis accuracy and efficiency. First, different sizes of convolution kernels are applied to extract multi-scale features from original signals using a multi-scale one-dimensional convolutional neural network (1DCNN); this not only improves the learning ability of the features but also enables the fine characterization of the features. Then, using Dempster-Shafer (DS) evidence theory, improved by multi-sensor information fusion strategy, the feature signals extracted by the multi-scale 1DCNN are fused to realize the fault detection and location. Finally, the experimental results of fault detection on a flash furnace show that the accuracy of the proposed method is more than 99.65% and has better fault diagnosis, which proves the feasibility and effectiveness of the proposed method.

Land Cover Extraction in the Typical Black Soil Region of Northeast China Using High-Resolution Remote Sensing Imagery

The black soil region of Northeast China is one of the most fertile soil areas in the world and serves as a crucial grain-producing region in China. However, excessive development and improper utilization have led to severe land use issues. Conducting land cover extraction in this region can provide essential data support for monitoring and managing natural resources effectively. This article utilizes GF-6 remote sensing imagery as the data source and adopts the U-Net model as the backbone network. By incorporating residual modules and adjusting the convolution kernel size, a high-precision land cover extraction model called RAT-UNet is developed. Taking Qiqihar City as an example, the RAT-UNet model is applied to extract land cover information. The results are as follows: (1) The RAT-UNet model achieves high accuracy in land cover extraction, with the following accuracies for different land types: cropland (95.11%), forestland (93.61%), grassland (68.41%), water bodies (94.67%), residential land (89.40%), and unused land (87.25%). (2) The land cover extraction performance of the RAT-UNet model is superior to DeepLabV3, U-Net, SegNet, and LinkNet34 models. This research outcome provides methodological support for the intelligent and high-precision extraction of land cover information and also offers timely data for Qiqihar city’s land use planning.

Learned spatiotemporal correlation priors for CEST image denoising using incorporated global-spectral convolution neural network.

To develop a deep learning-based method, dubbed Denoising CEST Network (DECENT), to fully exploit the spatiotemporal correlation prior to CEST image denoising. DECENT is composed of two parallel pathways with different convolution kernel sizes aiming to extract the global and spectral features embedded in CEST images. Each pathway consists of a modified U-Net with residual Encoder-Decoder network and 3D convolution. Fusion pathway with 1 × 1 × 1 convolution kernel is utilized to concatenate two parallel pathways, and the output of DECENT is noise-reduced CEST images. The performance of DECENT was validated in numerical simulations, egg white phantom experiments, and ischemic mouse brain and human skeletal muscle experiments in comparison with existing state-of-the-art denoising methods. Rician noise was added to CEST images to mimic a low SNR situation for numerical simulation, egg white phantom experiment, and mouse brain experiments, while human skeletal muscle experiments were of inherently low SNR. From the denoising results evaluated by peak SNR (PSNR) and structural similarity index (SSIM), the proposed deep learning-based denoising method (DECENT) can achieve better performance compared to existing CEST denoising methods such as NLmCED, MLSVD, and BM4D, avoiding complicated parameter tuning or time-consuming iterative processes. DECENT can well exploit the prior spatiotemporal correlation knowledge of CEST images and restore the noise-free images from their noisy observations, outperforming state-of-the-art denoising methods.

A Novel Diagnosis Method Based on the Feature Selection Algorithm and Convolutional Neural Network for SF6 Gas-Insulated Equipment

SF6-decomposition-based insulation degradation of gas-insulated equipment has attracted considerable research attention in the power industry. However, because limited feature mining has been performed for SF6 decomposition components, diagnostic accuracy is reduced. Therefore, in this study, a novel diagnosis strategy was proposed based on various feature selection algorithms and convolutional neural network (CNN). In all, 19 characteristic products were obtained by combining original characteristic products. Various feature selection algorithms were used to mine feature information. A novel CNN structure was designed. The 1-D input layer was reorganized into the 2-D layer to improve the operation efficiency of the CNN. The robustness and generalization ability of the network were improved by adjusting the training step size and convolution kernel size or number. The diagnosis accuracy of the SF6 decomposition component was improved to 99%. Thus, a set of accurate and effective diagnosis schemes was proposed for the power industry.

A Method to Reduce the Intra-Frame Prediction Complexity of HEVC Based on D-CNN

Among a series of video coding standards jointly developed by ITU-T, VCEG, and MPEG, high-efficiency video coding (HEVC) is one of the most widely used video coding standards today. Therefore, it is still necessary to further reduce the coding complexity of HEVC. In the HEVC standard, a flexible partitioning procedure entitled “quad-tree partition” is proposed to significantly improve the coding efficiency, which, however, leads to high coding complexity. To reduce the coding complexity of the intra-frame prediction, this paper proposes a scheme based on a densely connected convolution neural network (D-CNN) to predict the partition of coding units (CUs). Firstly, a densely connected block was designed to improve the efficiency of the CU partition by fully extracting the pixel features of CTU. Then, efficient channel attention (ECA) and adaptive convolution kernel size were applied to a fast CU partition for the first time to capture the information of the D-CNN convolution channels. Finally, a threshold optimization strategy was formulated to select the best threshold for each depth to further balance the computation complexity of video coding and the performance of RD. The experimental results show that the proposed method reduces the encoding time of HEVC by 60.14%, with a negligible reduction in RD performance, which is better than the existing fast partitioning methods.

ADU-Net: An Attention Dense U-Net based deep supervised DNN for automated lesion segmentation of COVID-19 from chest CT images

An automatic method for qualitative and quantitative evaluation of chest Computed Tomography (CT) images is essential for diagnosing COVID-19 patients. We aim to develop an automated COVID-19 prediction framework using deep learning.We put forth a novel Deep Neural Network (DNN) composed of an attention-based dense U-Net with deep supervision for COVID-19 lung lesion segmentation from chest CT images. We incorporate dense U-Net where convolution kernel size 5×5 is used instead of 3×3. The dense and transition blocks are introduced to implement a densely connected network on each encoder level. Also, the attention mechanism is applied between the encoder, skip connection, and decoder. These are used to keep both the high and low-level features efficiently. The deep supervision mechanism creates secondary segmentation maps from the features. Deep supervision combines secondary supervision maps from various resolution levels and produces a better final segmentation map. The trained artificial DNN model takes the test data at its input and generates a prediction output for COVID-19 lesion segmentation. The proposed model has been applied to the MedSeg COVID-19 chest CT segmentation dataset. Data pre-processing methods help the training process and improve performance.We compare the performance of the proposed DNN model with state-of-the-art models by computing the well-known metrics: dice coefficient, Jaccard coefficient, accuracy, specificity, sensitivity, and precision. As a result, the proposed model outperforms the state-of-the-art models.This new model may be considered an efficient automated screening system for COVID-19 diagnosis and can potentially improve patient health care and management system.

Artificial Intelligence-Based Digital Media Design Effect Enhancement Mechanism

The advancement of AI technology has promoted the development speed of digital multimedia and brought a new experience to the digital media experience effect. In this paper, we aim to using artificial intelligence methods to enhance the digital media design experience. Specifically, we propose a method for low-light image enhancement using generative adversarial networks as a model framework. To better solve the problem, we design the following strategies in our proposed method. First, we preprocess the images into patches with a proper size. Second, we introduce the overall network structure of GAN. Third, we designed a multifeature extraction module with different sizes of convolution kernels to enhance the model’s ability for extracting features. Fourth, we propose a loss that combines the mean square error distance function with the adversarial error function to enable the model to learn the distributional features of the image. Finally, we verify the effectiveness of the proposed method on two datasets, namely, PASCAL VOC and MIT-Adobe FiveK. The results show that our proposed method performs well in the process of evaluation.

Time series classification based on convolutional network with a Gated Linear Units kernel

Time series data are ubiquitous in human society and nature, and classification is one of the most significant problems in the field of time series mining. Although it has been intensively studied, and has achieved significant results and successful applications, it is still a challenging problem, which requires capturing of multi-scale features of one-dimensional or multi-dimensional time series in variable length. In this paper, we propose a novel time series feature extraction block named Convolutional Gated Linear Units (CGLU), which is a combination of convolutional operations and Gated Linear Units for adaptively extracting local temporal features of time series. Combined with a temporal maxpooling block, it can extract global temporal features. To capture more diverse features, the Inception architecture is adopted to organize the CGLUs with different convolution kernel sizes, which result in the Convolutional GLU network. In order to evaluate the performance, we conduct extensive experiments on the UCR time series datasets (one-dimension) and UEA datasets (multi-dimension). Compared with baselines, our model obtains best results in terms of classification accuracy and training speed, which demonstrate effectiveness and efficiency of CGLUs and Conv-GLU network on time series classification tasks.

MFE-ResNet: A new extraction framework for land cover characterization in mining areas

Obtaining accurate information about land cover is a critical aspect of environmental monitoring in mining areas. However, the land cover in these areas is often complex and poses a challenge for analysis compared to other landscapes. This results in the difficulty of acquiring effective features and an unbalanced proportion of available training samples. We introduce a new approach to obtaining effective image features and enhancing the extraction of open-pit (OP) and waste-dump area (WDA) by using a mining feature-enhanced ResNet (MFE-ResNet). Our framework utilizes image objects as the basic units of high spatial resolution images (HSRIs) to better incorporate spatial and geometric information. The MFE-ResNet is an improvement of the ResNet18 architecture, forming a 26-layer network, and reducing the computational complexity by changing the convolutional kernel size. A squeeze-and-excitation (SE) module is also introduced to form an SE-residual structure to improve the network’s ability to perceive the detailed mining area features. To alleviate the sample imbalance problem, a switchable normalization (SN) layer is applied to optimize the internal network structure. In addition, focal loss is introduced in the training process. We build an evaluation system that considers both the number of objects and their area. Compared with other six deep learning methods, our MFE-ResNet achieves the best overall accuracy in both large and small study areas. Specifically, MFE-ResNet obtained F values (F) of 87.73% and 97.28%, and quality (Q) of 81.37% and 95.18% in OP extraction. The F of WDA extraction improved by 0.99% and 2.13%, and the Q improved by 0.98% and 9.56% compared to the second-ranked model. We show that our framework exhibits promising performance in monitoring mining areas.

Multiscale cascade recurrent dilation convolution network for fault diagnosis of rolling bearing under cross-load conditions

Recently, deep learning (DL) models based on convolutional neural networks have achieved satisfactory results in rolling bearing fault diagnosis. However, the bearings usually work in variable loading conditions, and their feature distribution could vary with load. The important features cannot be effectively captured in the convolution process using the existing diagnosis models, resulting in poor generalization performance. In this paper, a novel DL model, named multiscale cascade recurrent dilation convolution network, is proposed by introducing the dilated convolution and global average pooling (GAP) layer. Firstly, a new multiscale cascade structure with different convolution kernel sizes is introduced to extract multiscale features contained in the vibration signal. Secondly, a recurrent dilation convolution strategy is designed in each branch of the multiscale cascade structure to extract abundant feature information. Finally, the GAP is employed to reduce redundant feature vectors and output them, while a classifier of multilayer perceptron is used to automatically identify the fault types. The effectiveness of the proposed algorithm is evaluated by two experimental cases. The results show that the proposed method can successfully identify the labels of fault samples under unknown load conditions using the fault samples with labels under existing load conditions. Compared with other methods, this method exhibits excellent robustness and generalization performance for bearing fault diagnosis under cross-load conditions.

Predicting gene and protein expression levels from DNA and protein sequences with Perceiver

Background and Objective: The functions of an organism and its biological processes result from the expression of genes and proteins. Therefore quantifying and predicting mRNA and protein levels is a crucial aspect of scientific research. Concerning the prediction of mRNA levels, the available approaches use the sequence upstream and downstream of the Transcription Start Site (TSS) as input to neural networks. The State-of-the-art models (e.g., Xpresso and Basenjii) predict mRNA levels exploiting Convolutional (CNN) or Long Short Term Memory (LSTM) Networks. However, CNN prediction depends on convolutional kernel size, and LSTM suffers from capturing long-range dependencies in the sequence. Concerning the prediction of protein levels, as far as we know, there is no model for predicting protein levels by exploiting the gene or protein sequences. Methods: Here, we exploit a new model type (called Perceiver) for mRNA and protein level prediction, exploiting a Transformer-based architecture with an attention module to attend to long-range interactions in the sequences. In addition, the Perceiver model overcomes the quadratic complexity of the standard Transformer architectures. This work’s contributions are 1. DNAPerceiver model to predict mRNA levels from the sequence upstream and downstream of the TSS; 2. ProteinPerceiver model to predict protein levels from the protein sequence; 3. Protein&DNAPerceiver model to predict protein levels from TSS and protein sequences. Results: The models are evaluated on cell lines, mice, glioblastoma, and lung cancer tissues. The results show the effectiveness of the Perceiver-type models in predicting mRNA and protein levels. Conclusions: This paper presents a Perceiver architecture for mRNA and protein level prediction. In the future, inserting regulatory and epigenetic information into the model could improve mRNA and protein level predictions. The source code is freely available at https://github.com/MatteoStefanini/DNAPerceiver

Scene Text Detection Based on Multi-Headed Self-Attention Using Shifted Windows

Scene text detection has become a popular topic in computer vision research. Most of the current research is based on deep learning, using Convolutional Neural Networks (CNNs) to extract the visual features of images. However, due to the limitations of convolution kernel size, CNNs can only extract local features of images with small perceptual fields, and they cannot obtain more global features. In this paper, to improve the accuracy of scene text detection, a feature enhancement module is added to the text detection model. This module acquires global features of an image by computing the multi-headed self-attention of the feature map. The improved model extracts local features using CNNs, while extracting global features through the feature enhancement module. The features extracted by both of these are then fused to ensure that visual features at different levels of the image are extracted. A shifted window is used in the calculation of the self-attention, which reduces the computational complexity from the second power of the input image width-height product to the first power. Experiments are conducted on the multi-oriented text dataset ICDAR2015 and the multi-language text dataset MSRA-TD500. Compared with the pre-improvement method DBNet, the F1-score improves by 0.5% and 3.5% on ICDAR2015 and MSRA-TD500, respectively, indicating the effectiveness of the model improvement.

Automatic modulation recognition based on a multiscale network with statistical features

Automatic modulation recognition (AMR) can be used in dynamic spectrum access (DSA) techniques to reduce the pressure on spectrum resources. In this paper, we propose a multiscale convolution-based network model called MSNet-SF, which combines traditional statistical features and deep learning (DL) to balance recognition accuracy and complexity. In the model, the feature information is extracted by two multiscale modules, which consist of unit convolution and three different sizes of convolution kernels arranged in parallel. Additionally, the sparse connectivity of unit convolution enables the network to be more lightweight. Five statistical features (four higher-order cumulants (HOCs) and one zero-centered normalized instantaneous magnitude tightness) are also input into the model and are fully fused with the main feature map by multiplication to achieve complementarity of long-term and short-term features. This approach yields a large performance gain at a small cost and greatly reduces the confusion between QAM16 and QAM64. Simulation results in the RML2018.10A dataset show that the average recognition accuracy of the model improved by 4% after adding the statistical features and achieved an accuracy of more than 97% from 12 dB.

CNN_FunBar: Advanced Learning Technique for Fungi ITS Region Classification.

Fungal species identification from metagenomic data is a highly challenging task. Internal Transcribed Spacer (ITS) region is a potential DNA marker for fungi taxonomy prediction. Computational approaches, especially deep learning algorithms, are highly efficient for better pattern recognition and classification of large datasets compared to in silico techniques such as BLAST and machine learning methods. Here in this study, we present CNN_FunBar, a convolutional neural network-based approach for the classification of fungi ITS sequences from UNITE+INSDC reference datasets. Effects of convolution kernel size, filter numbers, k-mer size, degree of diversity and category-wise frequency of ITS sequences on classification performances of CNN models have been assessed at all taxonomic levels (species, genus, family, order, class and phylum). It is observed that CNN models can produce >93% average accuracy for classifying ITS sequences from balanced datasets with 500 sequences per category and 6-mer frequency features at all levels. The comparative study has revealed that CNN_FunBar can outperform machine learning-based algorithms (SVM, KNN, Naïve-Bayes and Random Forest) as well as existing fungal taxonomy prediction software (funbarRF, Mothur, RDP Classifier and SINTAX). The present study will be helpful for fungal taxonomy classification using large metagenomic datasets.

Deep-Learning-Based Low-Frequency Reconstruction in Full-Waveform Inversion

Low frequencies are vital for full-waveform inversion (FWI) to retrieve long-scale features and reliable subsurface properties from seismic data. Unfortunately, low frequencies are missing because of limitations in seismic acquisition steps. Furthermore, there is no explicit expression for transforming high frequencies into low frequencies. Therefore, low-frequency reconstruction (LFR) is imperative. Recently developed deep-learning (DL)-based LFR methods are based on either 1D or 2D convolutional neural networks (CNNs), which cannot take full advantage of the information contained in 3D prestack seismic data. Therefore, we present a DL-based LFR approach in which high frequencies are transformed into low frequencies by training an approximately symmetric encoding-decoding-type bridge-shaped 3D CNN. Our motivation is that the 3D CNN can naturally exploit more information that can be effectively used to improve the LFR result. We designed a Hanning-based window for suppressing the Gibbs effect associated with the hard splitting of the low- and high-frequency data. We report the significance of the convolutional kernel size on the training stage convergence rate and the performance of CNN’s generalization ability. CNN with reasonably large kernel sizes has a large receptive field and is beneficial to long-wavelength LFR. Experiments indicate that our approach can accurately reconstruct low frequencies from bandlimited high frequencies. The results of 3D CNN are distinctly superior to those of 2D CNN in terms of precision and highly relevant low-frequency energy. FWI on synthetic data indicates that the DL-predicted low frequencies nearly resemble those of actual low frequencies, and the DL-predicted low frequencies are accurate enough to mitigate the FWI’s cycle-skipping problems. Codes and data of this work are shared via a public repository.

Optimised U-Net for Land Use–Land Cover Classification Using Aerial Photography

Convolutional Neural Networks (CNN) consist of various hyper-parameters which need to be specified or can be altered when defining a deep learning architecture. There are numerous studies which have tested different types of networks (e.g. U-Net, DeepLabv3+) or created new architectures, benchmarked against well-known test datasets. However, there is a lack of real-world mapping applications demonstrating the effects of changing network hyper-parameters on model performance for land use and land cover (LULC) semantic segmentation. In this paper, we analysed the effects on training time and classification accuracy by altering parameters such as the number of initial convolutional filters, kernel size, network depth, kernel initialiser and activation functions, loss and loss optimiser functions, and learning rate. We achieved this using a well-known top performing architecture, the U-Net, in conjunction with LULC training data and two multispectral aerial images from North Queensland, Australia. A 2018 image was used to train and test CNN models with different parameters and a 2015 image was used for assessing the optimised parameters. We found more complex models with a larger number of filters and larger kernel size produce classifications of higher accuracy but take longer to train. Using an accuracy-time ranking formula, we found using 56 initial filters with kernel size of 5 × 5 provide the best compromise between training time and accuracy. When fully training a model using these parameters and testing on the 2015 image, we achieved a kappa score of 0.84. This compares to the original U-Net parameters which achieved a kappa score of 0.73.

Research on transformer internal defect detection based on large scale model

Aiming at the problem that the traditional convolution network can not be used to detect and identify targets in the more complex target detection application scenarios in power grid business, a target detection scheme based on the ability of large model is proposed. The effects of convolution neural network layers, convolution kernel size and feature dimension on the recognition effect are systematically analyzed and compared. The differences between traditional convolution network and large model in data volume, accuracy and model construction efficiency are compared. The experimental results show that for the specific target detection business scenario downstream of power grid, the model constructed by this scheme has the advantages of fast and efficient, high accuracy and less data dependence, and the comprehensive performance is better than the traditional convolution network detection method.

Anti‐noise diesel engine misfire diagnosis using a multi‐scale CNN‐LSTM neural network with denoising module

AbstractCurrently, accuracy of existing diesel engine fault diagnosis methods under strong noise and generalisation performance between different noise levels are still limited. A novel multi‐scale CNN‐LSTM neural network (MSCNN‐LSTMNet) is proposed with a residual‐CNN denoising module for anti‐noise diesel engine misfire diagnosis. First, a residual‐CNN module is designed for denoising the original vibration signal measured from the diesel engine cylinder and residual loss for constructing a new loss function is utilised. Considering the essential characteristics of measured vibration signals at different scales, a multi‐scale convolutional NN (CNN) block is designed to realize multi‐scale feature extraction. Specifically, multiple convolution layers with different branches and different convolution kernel sizes are utilised to extract different time scales features, enhancing the robustness of the model. On this basis, the LSTM is utilised to further extract sequential features for improving anti‐noise and generalisation performances. The effectiveness of MSCNN‐LSTMNet is validated by experimental results of both one‐ and hybrid‐cylinder diesel engine misfires diagnosis under various noise levels and working conditions. The results demonstrate that MSCNN‐LSTMNet achieved much better anti‐noise and generalisation performances than the existing methods. Under strong noise conditions (−10 dB signal‐to‐noise ratio) for four datasets, MSCNN‐LSTMNet obtained 97.561% average accuracy, while average accuracy for random forest, deep neural network, CNN and MSCNNNet were 73.828%, 79.544%, 82.247%, and 89.741%, respectively. Moreover, for 11 noise generalisation tasks between different noise levels, MSCNN‐LSTMNet obtained at least 96.679%, 97.849%, 98.892%, and 94.010% accuracy on the four datasets, which are much higher than those of the existing methods.

CNN parameter design based on fault signal analysis and its application in bearing fault diagnosis

As a representative deep learning network, Convolutional Neural Network (CNN) has been extensively used in bearing fault diagnosis and many good results have been reported. In Prognostics and Health Management (PHM) field, the CNN’s input size is usually designed as a 1D vector or 2D square matrix, and the convolution kernel size is also defined as a square shape like 3 × 3 and 5 × 5, which are directly adopted from the image recognition. Though satisfying results can be obtained, CNN with such parameter specifications is not optimal and efficient. To this end, this paper elaborated the physical characteristics of bearing acceleration signals to guide the CNN design. First, the fault period under different fault types and shaft rotation frequency were used to determine the size of CNN’s input. Next, an exponential function was involved in fitting the envelope of decaying acceleration signal during each fault period, and signal length within different decaying ratios was used to define the CNN’s kernel size. Finally, the designed CNN was validated with the Case Western Reserve University bearing dataset and Paderborn University bearing dataset. Results confirm that the physics-guided CNN (PGCNN) with rectangular input shape and rectangular convolution kernel works better than the baseline CNN with higher accuracy and smaller uncertainty. The feasibility of designing CNN parameters with physics-guided rules derived from bearing fault signal analysis has also been verified.