Convolution Kernel Size Research Articles

Inspection of Defects Depth for Stainless-Steel Sheets Using Four-Coil Excitation Sensor and Deep Learning

The detection of lift-off variation presents a significant challenge in many eddy current testing (ECT) applications, where it is often considered an undesirable signal. To address this issue, the paper proposes a novel 4-coil excitation sensor, which can effectively suppress the lift-off effect when detecting defects of varying depths in planar structures. The proposed sensor design leverages opposite excitation directions between any two arbitrary adjacent excitation coils, which can effectively cancel out the lift-off signal due to a defect-free plate. However, this 4-coil excitation design has an angular sensitivity – i.e., different signal response for different angles between the same defect and the sensor. This angular sensitivity will cause trouble to identify the depth of defects without knowing the angle in prior. To overcome this challenge, the current study introduces deep learning (DL) models as a potential solution. Signals representing different defect depths and angles between the sensor and the defect are acquired using an experimental platform in our lab. The real and imaginary parts of the induced voltage on the receiver coil labelled with depths and angles are utilized to train a 1-D convolutional neural network (1D CNN) model, in which different convolution kernel sizes are used to extract more information and speed up convergence. The high performance of the proposed structure of 1D CNN were compared with convolutional long-short term memory (LSTM) model, recurrent neural network (RNN) model, LSTM+RNN model and ResNet18 model. The proposed 1D CNN model not only achieves good performances with a higher accuracy of 99.8% but also requires a smaller model size and less computational time.

Motor Imagery EEG Decoding Based on Multi-scale Hybrid Networks and Feature Enhancement.

Motor Imagery (MI) based on Electroencephalography (EEG), a typical Brain-Computer Interface (BCI) paradigm, can communicate with external devices according to the brain's intentions. Convolutional Neural Networks (CNN) are gradually used for EEG classification tasks and have achieved satisfactory performance. However, most CNN-based methods employ a single convolution mode and a convolution kernel size, which cannot extract multi-scale advanced temporal and spatial features efficiently. What's more, they hinder the further improvement of the classification accuracy of MI-EEG signals. This paper proposes a novel Multi-Scale Hybrid Convolutional Neural Network (MSHCNN) for MI-EEG signal decoding to improve classification performance. The two-dimensional convolution is used to extract temporal and spatial features of EEG signals and the one-dimensional convolution is used to extract advanced temporal features of EEG signals. In addition, a channel coding method is proposed to improve the expression capacity of the spatiotemporal characteristics of EEG signals. We evaluate the performance of the proposed method on the dataset collected in the laboratory and BCI competition IV 2b, 2a, and the average accuracy is at 96.87%, 85.25%, and 84.86%, respectively. Compared with other advanced methods, our proposed method achieves higher classification accuracy. Then we use the proposed method for an online experiment and design an intelligent artificial limb control system. The proposed method effectively extracts EEG signals' advanced temporal and spatial features. Additionally, we design an online recognition system, which contributes to the further development of the BCI system.

Improved Three-Dimensional Inception Networks for Hyperspectral Remote Sensing Image Classification

Convolutional neural networks (CNNs) have been applied to the categorization of hyperspectral images. Typically, CNNs employ fixed convolution kernel sizes. This may be complicated by the high-dimensional features and the spatial-spectral features variability of hyperspectral data. In such situation, dealing with various dataset types flexibly is hampered by the fixed kernel size. Therefore, we chose a parallel design, an upgraded three-dimensional (3D) inception network, for hyperspectral image classification. Additionally, the adaptive band selection of hyperspectral remote sensing data employs dimensionality reduction based on interactive information entropy, allowing the selection of low-redundant bands with more abundant and discriminative information. Experimental results based on the publicly available hyperspectral datasets demonstrated that the enhanced 3D inception network can perform well with limited training samples. Moreover, the combinations of different dimensional convolutions in the 3D-2D and 3D-2D-1D inception networks achieved comparable classification results. According to the experimental findings, the suggested networks are generalised classification models with high classification rates.

Using Hybrid Penalty and Gated Linear Units to Improve Wasserstein Generative Adversarial Networks for Single-Channel Speech Enhancement

Recently, speech enhancement methods based on Generative Adversarial Networks have achieved good performance in time-domain noisy signals. However, the training of Generative Adversarial Networks has such problems as convergence difficulty, model collapse, etc. In this work, an end-to-end speech enhancement model based on Wasserstein Generative Adversarial Networks is proposed, and some improvements have been made in order to get faster convergence speed and better generated speech quality. Specifically, in the generator coding part, each convolution layer adopts different convolution kernel sizes to conduct convolution operations for obtaining speech coding information from multiple scales; a gated linear unit is introduced to alleviate the vanishing gradient problem with the increase of network depth; the gradient penalty of the discriminator is replaced with spectral normalization to accelerate the convergence rate of the model; a hybrid penalty term composed of L1 regularization and a scale-invariant signal-to-distortion ratio is introduced into the loss function of the generator to improve the quality of generated speech. The experimental results on both TIMIT corpus and Tibetan corpus show that the proposed model improves the speech quality significantly and accelerates the convergence speed of the model.

Research on Plant Species Identification Based on Improved Convolutional Neural Network

Plant species recognition is an important research area in image recognition in recent years. However, the existing plant species recognition methods have low recognition accuracy and do not meet professional requirements in terms of recognition accuracy. Therefore, ShuffleNetV2 was improved by combining the current hot concern mechanism, convolution kernel size adjustment, convolution tailoring, and CSP technology to improve the accuracy and reduce the amount of computation in this study. Six convolutional neural network models with sufficient trainable parameters were designed for differentiation learning. The SGD algorithm is used to optimize the training process to avoid overfitting or falling into the local optimum. In this paper, a conventional plant image dataset TJAU10 collected by cell phones in a natural context was constructed, containing 3000 images of 10 plant species on the campus of Tianjin Agricultural University. Finally, the improved model is compared with the baseline version of the model, which achieves better results in terms of improving accuracy and reducing the computational effort. The recognition accuracy tested on the TJAU10 dataset reaches up to 98.3%, and the recognition precision reaches up to 93.6%, which is 5.1% better than the original model and reduces the computational effort by about 31% compared with the original model. In addition, the experimental results were evaluated using metrics such as the confusion matrix, which can meet the requirements of professionals for the accurate identification of plant species.

A novel Deep Capsule Neural Network for Vowel Imagery patterns from EEG signals

Speech imagery has recently been included in the design of Brain–Computer Interfaces to develop novel communication or control systems based on brain activity that does not need external stimulation like evoked potentials. Three types of speech imagery exist: imagining words, syllables, or vowels. Words are composed of syllables and syllables by consonants and vowels. However, imagining just vowels generates Speech-Related Potentials that reduce the complexity of the brain activity in Electroencephalographic signals. This paper proposes a new classifier method for communication or control purposes based on a novel Deep Capsule Neural Network for Vowel Imagery recognition. The method is named Capsules for Vowel Imagery (CapsVI). CapsVI has the appropriate number and size of convolution kernels to find the relevant features of the input. The size of the capsules is estimated based on the feature patterns found during the training. The class capsules were developed based on prototypes patterns of /a/, /u/, and/no vowel/ classes. The experiments were developed with the DaSalla dataset. Results indicate that capsules model the Speech-Related Potentials of Vowel Imagery correctly enough to generate the necessary information for vowel pairwise classification. Furthermore, the results also demonstrate that CapsVI recognizes vowels with an average accuracy of 93.32% on the pairwise classification, and the best precision by subject is 94.68%. These results are the best in Vowel Imagery recognition of the English language reported in the literature.

Rapid nondestructive detecting of sorghum varieties based on hyperspectral imaging and convolutional neural network.

The purity of sorghum varieties is an important indicator of the quality of raw materials used in the distillation of liquors. Different varieties of sorghum may be mixed during the acquisition process, which will affect the flavor and quality of liquor. To facilitate the rapid identification of sorghum varieties, this study proposes a sorghum variety identification model using hyperspectral imaging (HSI) technology combined with convolutional neural network (AlexNet). First, the watershed algorithm, which was modified with the extended-maxim transform, was used to segment the hyperspectral images of a single sorghum grain. The isolated forest algorithm was used to eliminate abnormal spectral data from the complete spectral data. Secondly, the AlexNet model of sorghum variety identification was established based on the two-dimensional gray image data of sorghum grain in group 1. The effects of different preprocessing methods and different convolution kernel sizes on the performance of the AlexNet model were discussed. The eigenvalues of the last layer of the AlexNet model were visualized using the t-distributed random neighborhood embedding method, which is used to evaluate the separability of features extracted by the AlexNet model. The performance differences between the optimal AlexNet model and traditional machine learning models for sorghum variety identification were compared. Finally, the varieties of sorghum grains in groups 2 and 3 were identified based on the optimal AlexNet model, and the average accuracy values of the test set reached 95.62% and 95.91% respectively. The results in this study demonstrated that HSI combined with the AlexNet model could provide a feasible technical approach for the detection of sorghum varieties. © 2022 Society of Chemical Industry.

A Mixed Visual Encoding Model Based on the Larger-Scale Receptive Field for Human Brain Activity.

Research on visual encoding models for functional magnetic resonance imaging derived from deep neural networks, especially CNN (e.g., VGG16), has been developed. However, CNNs typically use smaller kernel sizes (e.g., 3 × 3) for feature extraction in visual encoding models. Although the receptive field size of CNN can be enlarged by increasing the network depth or subsampling, it is limited by the small size of the convolution kernel, leading to an insufficient receptive field size. In biological research, the size of the neuronal population receptive field of high-level visual encoding regions is usually three to four times that of low-level visual encoding regions. Thus, CNNs with a larger receptive field size align with the biological findings. The RepLKNet model directly expands the convolution kernel size to obtain a larger-scale receptive field. Therefore, this paper proposes a mixed model to replace CNN for feature extraction in visual encoding models. The proposed model mixes RepLKNet and VGG so that the mixed model has a receptive field of different sizes to extract more feature information from the image. The experimental results indicate that the mixed model achieves better encoding performance in multiple regions of the visual cortex than the traditional convolutional model. Also, a larger-scale receptive field should be considered in building visual encoding models so that the convolution network can play a more significant role in visual representations.

DRPN: Making CNN dynamically handle scale variation

Based on our observations of infrared targets, significant scale variation along sequence frames occurs with high frequency. This paper proposes a dynamic re-parameterization network (DRPN) to deal with the scale variation and balance the detection precision between small and large targets in infrared datasets. DRPN adopts the multiple branches with different sizes of convolution kernels and the dynamic convolution strategy. Multiple branches with different sizes of convolution kernels have different sizes of receptive fields. The dynamic convolution strategy makes DRPN adaptively weight multiple branches. DRPN can dynamically adjust the receptive field according to the scale variation of the target. Besides, in order to maintain efficient inference in the test phase, the multi-branch structure is further converted to a single-branch structure via the re-parameterization technique after training. Extensive experiments on FLIR, KAIST, and InfraPlane datasets demonstrate the effectiveness of our proposed DRPN. The experimental results show that the detector using DRPN as the basic structure rather than SKNet or TridentNet obtains the best performances.

LiSeNet: multitask lightweight segmentation network for accurate and complete iris segmentation

The demand for applying the iris segmentation model on mobile devices has been growing rapidly. Most current segmentation networks have an enormous amount of parameters, hence unsuitable for mobile devices, while other small memory footprint models follow the spirit of classification networks and ignore the inherent characteristic of segmentation. To address the challenge, we propose a lightweight segmentation network (LiSeNet) for iris segmentation of noisy images. Unlike previous studies that only focus on improving the accuracy of segmentation masks, LiSeNet can simultaneously obtain segmentation masks, parameterized pupillary and limbic boundaries of the iris, further enabling CNN-based iris segmentation to be applied in any regular iris recognition systems. We first propose a multiscale concatenate (MSC) Block, which connects multiple sizes of convolution kernels in a dense manner, gradually reduces the dimension of feature maps and uses the aggregation of them for image representation. Based on the MSC block, we develop a two-stage refinement encoder to aggregate discriminative features through subnetwork feature reuse and substage feature reassess, thus obtaining a sufficient receptive field and enhancing the model learning ability. To exploit object contextual information more efficiently, we further devise a grouped spatial attention to emphasize the important features and suppress irrelevant noises through a gating mechanism in the decoder. Extensive experiments on three challenging iris datasets show that LiSeNet, without any complicated postprocessing, achieves competitive or state-of-the-art performance with only 2.2M parameters, being 14 × smaller than the previous best method. Code will be publicly available.

A novel oil pipeline leakage detection method based on the sparrow search algorithm and CNN

A novel sparrow search algorithm and convolutional neural network (SSA-CNN) method is proposed for oil pipeline leakage detection. Firstly, the proposed SSA-CNN method converts the input data from time series to two-dimensional matrix, and the classification conditions of different convolution kernel sizes and different pooling sizes are compared. Then, the SSA algorithm is used to optimize the parameters of the CNN. The simulation results show that, compared with the traditional machine learning method, using two-dimensional data as input can enhance the neural network's extraction of eigenvalues. The proposed SSA-CNN method was able to accurately classify 148 sample points in 150 test sets with an accuracy rate of 98.67%, which is not only higher than traditional machine learning methods, but also further improves the classification capability of CNN, while the SSA-CNN method can use the parameters already learned in the test set to ensure real-time detection.

Toward masked face recognition: An effective facial feature extraction and refinement model in multiple scenes

AbstractWith the impact of the COVID‐19 epidemic, the demand for masked face recognition technology has increased. In the process of masked face recognition, some problems such as less feature information and poor robustness to the environment are obvious. The current masked face recognition model is not quantified enough for feature extraction, there are large errors for faces with high similarity, and the categories cannot be clustered during the detection process, resulting in poor classification of masks, which cannot be well adapted to changes in multiple environments. To solve current problems, this paper designs a new masked face recognition model, taking improved Single Shot Multibox Detector (SSD) model as a face detector, and replaces the input layer VGG16 of SSD with Deep Residual Network (ResNet) to increase the receptive field. In order to better adapt to the network, we adjust the convolution kernel size of ResNet. In addition, we fine‐tune the Xception network by designing a new fully connected layer, and reduce the training cycle. The weights of the three input samples including anchor, positive and negative are shared and clustered together with triplet network to improve recognition accuracy. Meanwhile, this paper adjusts alpha parameter in triplet loss. A higher value of alpha can improve the accuracy of model recognition. We further adopt a small trick to classify and predict face feature vectors using multi‐layer perceptron (MLP), and a total of 60 neural nodes are set in the three neural layers of MLP to get higher classification accuracy. Moreover, three datasets of MFDD, RMFRD and SMFRD are fused to obtain high‐quality images in different scenes, and we also add data augmentation and face alignment methods for processing, effectively reducing the interference of the external environment in the process of model recognition. According to the experimental results, the accuracy of masked face recognition reaches 98.3%, it achieves better results compared with other mainstream models. In addition, the hyper‐parameters tuning experiment is carried out to improve the utilization of computing resources, which shows better results than the indicators of different networks.

Image classification and identification for rice leaf diseases based on improved WOACW_SimpleNet.

In view of the problem that manual selection of hyperparameters may lead to low performance and large consumption of manpower cost of the convolutional neural network (CNN), this paper proposes a nonlinear convergence factor and weight cooperative self-mapping chaos optimization algorithm (WOACW) to optimize the hyperparameters in the identification and classification model of rice leaf disease images, such as learning rate, training batch size, convolution kernel size and convolution kernel number. Firstly, the opposition-based learning is added to the whale population initialization with improving the diversity of population initialization. Then the algorithm improves the convergence factor, increases the weight coefficient, and calculates the self-mapping chaos. It makes the algorithm have a strong ability to find optimization in the early stage of iteration and fast convergence rate. And disturbance is carried out to avoid falling into local optimal solution in the late stage of iteration. Next, a polynomial mutation operator is introduced to correct the current optimal solution with a small probability, so that a better solution can be obtained in each iteration, thereby enhancing the optimization performance of the multimodal objective function. Finally, eight optimized performance benchmark functions are selected to evaluate the performance of the algorithm, the experiment results show that the proposed WOACW outperforms than 5 other common improved whale optimization algorithms. The WOACW_SimpleNet is used to identify rice leaf diseases (rice blast, bacterial leaf blight, brown spot disease, sheath blight and tungro disease), and the experiment results show that the identification average recognition accuracy rate reaches 99.35%, and the F1-score reaches 99.36%.

Improving Hybrid CTC/Attention Architecture for Agglutinative Language Speech Recognition.

Unlike the traditional model, the end-to-end (E2E) ASR model does not require speech information such as a pronunciation dictionary, and its system is built through a single neural network and obtains performance comparable to that of traditional methods. However, the model requires massive amounts of training data. Recently, hybrid CTC/attention ASR systems have become more popular and have achieved good performance even under low-resource conditions, but they are rarely used in Central Asian languages such as Turkish and Uzbek. We extend the dataset by adding noise to the original audio and using speed perturbation. To develop the performance of an E2E agglutinative language speech recognition system, we propose a new feature extractor, MSPC, which uses different sizes of convolution kernels to extract and fuse features of different scales. The experimental results show that this structure is superior to VGGnet. In addition to this, the attention module is improved. By using the CTC objective function in training and the BERT model to initialize the language model in the decoding stage, the proposed method accelerates the convergence of the model and improves the accuracy of speech recognition. Compared with the baseline model, the character error rate (CER) and word error rate (WER) on the LibriSpeech test-other dataset increases by 2.42% and 2.96%, respectively. We apply the model structure to the Common Voice—Turkish (35 h) and Uzbek (78 h) datasets, and the WER is reduced by 7.07% and 7.08%, respectively. The results show that our method is close to the advanced E2E systems.

TDMSAE: A transferable decoupling multi-scale autoencoder for mechanical fault diagnosis

When the working environment of mechanical equipment changes, the fault diagnosis model trained in the source domain cannot adapt well to the new environment, meaning the model drift has occurred. This paper proposes a transferable decoupling multi-scale autoencoder (TDMSAE) to address the above problem. Firstly, the feature extraction module is constructed using the multi-scale residual network, which can extract features of vibration signals at different scales by setting the different sizes of convolution kernels. Secondly, the distribution alignment module based on the transposed convolution is constructed to offset the distribution discrepancy of the source and target domains by transferring features and parameters of the feature extraction module so that the fault diagnosis model can achieve higher accuracy in the target domain. Finally, the experimental results show that the TDMSAE, compared to the other three existing methods, achieves a maximum of 100% accuracy when the model is transferred from the source domain to the target domain in eight sub-datasets.

Discrimination of Minced Mutton Adulteration Based on Sized-Adaptive Online NIRS Information and 2D Conventional Neural Network.

Single-probe near-infrared spectroscopy (NIRS) usually uses different spectral information for modelling, but there are few reports about its influence on model performance. Based on sized-adaptive online NIRS information and the 2D conventional neural network (CNN), minced samples of pure mutton, pork, duck, and adulterated mutton with pork/duck were classified in this study. The influence of spectral information, convolution kernel sizes, and classifiers on model performance was separately explored. The results showed that spectral information had a great influence on model accuracy, of which the maximum difference could reach up to 12.06% for the same validation set. The convolution kernel sizes and classifiers had little effect on model accuracy but had significant influence on classification speed. For all datasets, the accuracy of the CNN model with mean spectral information per direction, extreme learning machine (ELM) classifier, and 7 × 7 convolution kernel was higher than 99.56%. Considering the rapidity and practicality, this study provides a fast and accurate method for online classification of adulterated mutton.

Hyperspectral image super-resolution employing nonlocal block and hybrid multiscale three-dimensional convolution

Due to the special characteristics of hyperspectral images (HSIs), hundreds of continuous bands, and low spatial resolution, it is of great importance to explore the coherence among hyperspectral bands and extract the spatial information as far as possible to reconstruct the high-resolution (HR) HSIs, in which most of the methods failed. We propose an HSI super-resolution (SR) method termed NLB-HMS3D, which consists of two main parts named the spatial–similarity features module and the spatial and spectral correlation utilization module. Different from the majority of existing methods that stack multiple parallel two-dimensional convolution layers to blindly extract more spatial features, we introduce the nonlocal block to expand the receptive fields to thoroughly dig the spatial–similarity features from the image itself. This block not only greatly improves the effectiveness but also reduces tons of parameters. To better preserve the spectral details, we further propose a new block called multiscale spectral features fusion block using the separated three-dimensional convolution with different convolution kernel sizes to explore the diverse spatial–spectral features and fuse them to recover better spectral details. The experiments and data analysis demonstrate that NLB-MS3D can obtain superior performance over many existing state-of-the-art algorithms.

Convolutional Neural Network‐Based Method for Predicting Oxygen Content at the End Point of Converter

The main factor affecting the cleanliness of steel is inclusions, most of which are deoxidation products. The greater the oxygen content fluctuation at the converter's end point, the more unstable the inclusion control. A large amount of charging, operation, and sampling data is collected during the converter smelting, but it is not fully utilized, and the deoxidizing alloy addition still relies on manual experience. Herein, the representative convolutional neural network (CNN) of deep learning is adopted, and the number of convolutional layers, convolutional kernel size, and the number of convolutional kernels are optimized to establish the best prediction model of the converter end point. Its root mean square error, mean absolute error, and mean absolute percentage error are 35.29%, 25.59%, and 7.30%, respectively, which are superior to the back propagation neural network. This CNN prediction model uses 1300 sets of preprocessed data containing 23 indicator variables, of which 1200 sets of data are used for training and 100 sets for model validation. Within the ±50 ppm error scope, the model prediction hit rate can reach 87%, and when within the ±70 ppm error scope, the hit rate is 93%.

Scale-Space Feature Recalibration Network for Single Image Deraining.

Computer vision technology is increasingly being used in areas such as intelligent security and autonomous driving. Users need accurate and reliable visual information, but the images obtained under severe weather conditions are often disturbed by rainy weather, causing image scenes to look blurry. Many current single image deraining algorithms achieve good performance but have limitations in retaining detailed image information. In this paper, we design a Scale-space Feature Recalibration Network (SFR-Net) for single image deraining. The proposed network improves the image feature extraction and characterization capability of a Multi-scale Extraction Recalibration Block (MERB) using dilated convolution with different convolution kernel sizes, which results in rich multi-scale rain streaks features. In addition, we develop a Subspace Coordinated Attention Mechanism (SCAM) and embed it into MERB, which combines coordinated attention recalibration and a subspace attention mechanism to recalibrate the rain streaks feature information learned from the feature extraction phase and eliminate redundant feature information to enhance the transfer of important feature information. Meanwhile, the overall SFR-Net structure uses dense connection and cross-layer feature fusion to repeatedly utilize the feature maps, thus enhancing the understanding of the network and avoiding gradient disappearance. Through extensive experiments on synthetic and real datasets, the proposed method outperforms the recent state-of-the-art deraining algorithms in terms of both the rain removal effect and the preservation of image detail information.

Online chatter detection considering beat effect based on Inception and LSTM neural networks

Chatter is an unstable and self-excited vibration that adversely affects part quality and tool life in various machining processes. To achieve high-performance machining, chatter identification has attracted considerable interest from many researchers in recent decades. Nevertheless, most existing chatter detection approaches fail to consider the presence of the beat effect, which is an interference pattern caused by slightly different chatter frequencies. The neglect of the beat effect would seriously degrade the effectiveness of these methods and even result in false alarms. In this paper, a novel deep neural network combining the Inception module, long short-term memory (LSTM) and residual networks (ILR-DNN) is proposed for online chatter detection considering the presence of the beat effect. The ILR-DNN automatically extracts insightful features from two branches. One branch is a two-layer LSTM network that can capture temporal characteristics of the chatter development process from the raw cutting force. In the other branch, a two-layer Inception network consisting of multiple convolutional kernel size filters extracts multiscale features from the FFT spectrum of the cutting force. Afterwards, the extracted features are concatenated and go through a residual network to alleviate the gradient disappearing during deep-layer network training. Finally, the fully-connected and softmax layers are employed to make a final classification. Machining tests are carried out, and cutting forces are collected to validate the feasibility and effectiveness of the proposed chatter detection method. Results show that the proposed ILR-DNN achieves much better performance than other methods in distinguishing machining states, i.e., stable machining, chatter with the beat effect, and chatter without the beat effect.