Final Recognition Research Articles

In situ detection of spices with medicinal value based on laser‐induced breakdown spectroscopy

AbstractIt is important to detect and distinguish different spices because spices are widely used around the world. In this study, wormwood, artemisia annua, lemongrass and clove are taken as examples. First, laser‐induced breakdown spectroscopy (LIBS) is applied to detect and analyze the ash of different spice samples in situ. In the spectra of the ash of different samples, some characteristic lines of metal elements are observed, such as Ca, Na, Mg, K, and so on. By comparing the spectra of the ash, the relative intensities of the characteristic peaks are different, which can be employed to identify and distinguish different spice samples. Then, using LIBS combined with principal component analysis (PCA) and error back propagation artificial neural network (BP‐ANN), the model of classification is established to distinguish different spices. In PCA, the dimension of the spectra of the ash is reduced, and the cumulative contribution rate of the first two PCs exceeds 90%. The samples after dimension reduction by PCA are classified by BP‐ANN, and the recognition rate can reach 100%. After 10 cross‐verifications, the final recognition accuracy can reach 85.25%. All of the results show the model of classification has the potential in the field of identification and distinction of different spices.

MCF-Net: Fusion Network of Facial and Scene Features for Expression Recognition in the Wild

Nowadays, the facial expression recognition (FER) task has transitioned from a laboratory-controlled scenario to in-the-wild conditions. However, recognizing facial expressions in the wild is challenging due to factors such as variant backgrounds, low-quality facial images, and the subjectiveness of annotators. Therefore, deep neural networks have increasingly been leveraged to learn discriminative representations for FER. In this work, we propose the Multi-cues Fusion Net (MCF-Net), a novel deep learning model with a two-stream structure for FER. Our model first proposes a two-stream coding network to extract face and scene representations. Then, an adaptive fusion module is employed to fuse the two different representations for final recognition. In the face coding stream, a Sparse Mask Attention Learning (SMAL) module is utilized to adaptively generate the corresponding sparse face mask according to the input image. Meanwhile, we employ a Multi-scale Attention (MSA) module for extracting fine-grained feature subsets, which can obtain richer multi-scale interaction information. In the scene coding stream, a Relational Attention (RA) module is applied to construct the hidden relationship between the face and contextual features of non-facial regions by capturing the pairwise similarity. In order to verify the effectiveness and accuracy of our model, a large number of experiments are carried out on two public large-scale static facial expression image datasets, CAER-S and NCAER-S. By comparing the performance of our MCF-Net with other methods, the proposed model achieves superior results on two in-the-wild FER benchmarks: CAER-S with an accuracy of 81.82% and NCAER-S with an accuracy of 45.59%.

Learning multi-scale features for speech emotion recognition with connection attention mechanism

Speech emotion recognition (SER) has become a crucial topic in the field of human–computer interactions. Feature representation plays an important role in SER, but there are still many challenges in feature representation such as the inability to predict which features are most effective for SER and the cultural differences in emotion expression. Most previous studies use a single type of feature for the recognition task or conduct early fusion of features. However, a single type of feature cannot well reflect the emotions of speech signals. Also, different features contain different information, direct fusion cannot integrate the advantages of different features. To overcome these challenges, this paper proposes a parallel network for multi-scale SER based on a connection attention mechanism (AMSNet). AMSNet fuses fine-grained frame-level manual features with coarse-grained utterance-level deep features. Meanwhile, it adopts different speech emotion feature extraction modules according to the temporal and spatial features of speech signals, which enriches features and improves feature characterization. The network consists of a frame-level representation learning module (FRLM) based on the time structure and an utterance-level representation learning module (URLM) based on the global structure. Besides, improved attention-based long short-term memory (LSTM) is introduced into FRLM to focus on the frames that contribute more to the final emotion recognition result. In URLM, a convolutional neural network with the squeeze-and-excitation block (SCNN) is introduced to extract deep features. In addition, the connection attention mechanism is proposed for feature fusion, which applies different weights to different features. Extensive experiments are conducted on the IEMOCAP and EmoDB datasets, and the results demonstrate the effectiveness and performance superiority of AMSNet. Our code will be publicly available at https://codeocean.com/capsule/8636967/tree/v1.

Lightweight individual cow identification based on Ghost combined with attention mechanism

Individual cow identification is a prerequisite for intelligent dairy farming management, and is important for achieving accurate and informative dairy farming. Computer vision-based approaches are widely considered because of their non-contact and practical advantages. In this study, a method based on the combination of Ghost and attention mechanism is proposed to improve ReNet50 to achieve non-contact individual recognition of cows. In the model, coarse-grained features of cows are extracted using a large sensory field of cavity convolution, while reducing the number of model parameters to some extent. ResNet50 consists of two Bottlenecks with different structures, and a plug-and-play Ghost module is inserted between the two Bottlenecks to reduce the number of parameters and computation of the model using common linear operations without reducing the feature map. In addition, the convolutional block attention module (CBAM) is introduced after each stage of the model to help the model to give different weights to each part of the input and extract the more critical and important information. In our experiments, a total of 13 cows' side view images were collected to train the model, and the final recognition accuracy of the model was 98.58%, which was 4.8 percentage points better than the recognition accuracy of the original ResNet50, the number of model parameters was reduced by 24.85 times, and the model size was only 3.61 MB. In addition, to verify the validity of the model, it is compared with other networks and the results show that our model has good robustness. This research overcomes the shortcomings of traditional recognition methods that require human extraction of features, and provides theoretical references for further animal recognition.

Machining feature recognition based on a novel multi-task deep learning network

• Propose a novel learning-based machining feature recognition method that can effectively determine the faces that constitute machining features. • Construct a novel multi-task deep learning network based on point cloud data that can realize machining feature segmentation, machining feature identification, and bottom face identification simultaneously. • Propose a novel loss function that can be applied to the tasks of multi-label classification in deep learning. • Promote the application of deep learning in the manufacturing industry. Machining feature recognition is an essential task to realize the integration of computer-aided design (CAD), process planning (CAPP), and manufacturing (CAM) systems. In general, traditional rule-based machining feature recognition approaches are inflexible and computationally expensive. Moreover, it's not easy to design rules for intersecting machining features. To this end, various learning-based approaches were proposed in recent years. However, some of the existing learning-based approaches are complicated, time-consuming, and hard to deal with intersecting machining features, while some only approximately locate the position of the machining features rather than accurately segmenting them from the parts, which is not conducive to subsequent process planning. To address the above issues, a novel multi-task network named Associatively Segmenting and Identifying Network (ASIN) based on point cloud data is proposed for machining feature recognition. The proposed network performs three tasks: clusters the part faces with high similarities into machining features whose type are unidentified, predicts the semantic class (e.g. hole, pocket) for each face to identify the type of each machining feature, and identifies the bottom faces of the machining features. Through the three tasks, ASIN can realize machining feature segmentation, machining feature identification, and bottom face identification simultaneously. The final machining feature recognition results are obtained by combining the results of the above three tasks. After being trained with a large labeled point cloud dataset of 3D CAD models and a novel loss function, ASIN can be used to recognize machining features. Experimental results demonstrate that the proposed approach can effectively segment the machining features by determining the faces that compose them, and performs well in recognizing intersecting machining features.

Development and validation of the diagnostic accuracy of artificial intelligence-assisted ultrasound in the classification of splenic trauma.

Spleen is the most vulnerable organ in abdominal trauma. Ultrasound (US) has become an important examination method for splenic trauma. However, the sensitivity of routine US in the diagnosis of splenic trauma is low. Contrast-enhanced ultrasound (CEUS) can improve the sensitivity, but it also has some limitations. This study sought to explore the application value of artificial intelligence (AI)-assisted US in the classification of splenic trauma. The splenic injuries of Bama miniature pigs were established. A large number of ultrasonic images were collected. Then, 3-fold cross validation (CV) was used to establish the animal models. Next, clinical ultrasonic images were collected at multiple centers. All injuries were diagnosed by CEUS, enhanced CT or surgery. We used animal models to fine tune a small amount of human data, and then established the final AI splenic trauma recognition model. The whole model was constructed by averaging the prediction ability of the 3 fine-tuned models. Finally, 2 doctors' recognition US results of splenic trauma were compared to the AI recognition results. The area under the curve (AUC), sensitivity, specificity, negative predictive value, and positive predictive value were used to evaluate the diagnostic performance in diagnosis of spleen trauma. (I) Based on the receiver operating characteristic (ROC) curves, the test cohort 1 (AUC =0.90) and 2 (AUC =0.84) had a similar performance. Based on the decision curve analysis (DCA) curves, while threshold smaller than 0.8, the proposed model had better performance on test cohort 1 than test cohort 2. Test cohort 1 had higher sensitivity (0.82 vs. 0.71, P<0.01) and higher specificity (0.88 vs. 0.81, P<0.01) than test cohort 2. (II) The diagnostic accuracy of the AI model was higher than that of doctor 1 (0.82 vs. 0.62, P<0.001) and doctor 2 (0.82 vs. 0.66, P<0.001), and its specificity was higher than that of doctor (0.88 vs. 0.78, P=0.001). AI-assisted US diagnosis of splenic trauma can significantly improve the ultrasonic diagnosis rate. We still need to increase the number of samples to further improve the diagnostic efficiency of the model.

Voltage Sag Causes Recognition with Fusion of Sparse Auto-Encoder and Attention Unet

High-precision voltage sag cause identification is significant in solving the power quality problem. It is challenging for traditional deep learning models to balance training complexity and recognition performance when processing high-dimensional staging data samples, which affects the final recognition effect. This paper proposes a voltage sag identification method that fuses a sparse auto-encoder and Attention Unet. The model uses a sparse auto-encoder to perform unsupervised feature learning on the high-dimensional voltage sag waveform data and automatically obtains the deep low-dimensional features. Attention Unet, fused with cross-layer spatial and channel attention modules, further extracts these features to obtain recognition results with high performance. Compared with other deep learning recognition methods, the noise-adding experiments and the measured data are verified, indicating that the proposed method has low training complexity, higher recall, and better noise immunity. It benefits auxiliary decision-making for power quality management and governance.

Automatic Tandem Dual BlendMask Networks for Severity Assessment of Wheat Fusarium Head Blight

Fusarium head blight (FHB) disease reduces wheat yield and quality. Breeding wheat varieties with resistance genes is an effective way to reduce the impact of this disease. This requires trained experts to assess the disease resistance of hundreds of wheat lines in the field. Manual evaluation methods are time-consuming and labor-intensive. The evaluation results are greatly affected by human factors. Traditional machine learning methods are only suitable for small-scale datasets. Intelligent and accurate assessment of FHB severity could significantly facilitate rapid screening of resistant lines. In this study, the automatic tandem dual BlendMask deep learning framework was used to simultaneously segment the wheat spikes and diseased areas to enable the rapid detection of the disease severity. The feature pyramid network (FPN), based on the ResNet-50 network, was used as the backbone of BlendMask for feature extraction. The model exhibited positive performance in the segmentation of wheat spikes with precision, recall, and MIoU (mean intersection over union) values of 85.36%, 75.58%, and 56.21%, respectively, and the segmentation of diseased areas with precision, recall, and MIoU values of 78.16%, 79.46%, and 55.34%, respectively. The final recognition accuracies of the model for wheat spikes and diseased areas were 85.56% and 99.32%, respectively. The disease severity was obtained from the ratio of the diseased area to the spike area. The average accuracy for FHB severity classification reached 91.80%, with the average F1-score of 92.22%. This study demonstrated the great advantage of a tandem dual BlendMask network in intelligent screening of resistant wheat lines.

Identification of a novel PPARγ modulator with good anti-diabetic therapeutic index via structure-based screening, optimization and biological validation

PPARγ is well-known as the target receptor of TZD anti-diabetic drugs. However, recently the therapeutic benefits of these TZD drugs have been compromised by many severe side effects because of their full PPARγ agonistic action to lock the AF-2 helix. Herein, we conducted a virtual screening in the combination with structure-based design, synthesis and biological evaluation both in vitro and in vivo, leading to the identification of a potent candidate YG-C-20 as the SPPARγM with improved and safer anti-diabetic therapeutics. Mechanistically, this compound presented such desired pharmacological profiles (e.g., preferable anti-diabetic efficiencies and minimized side effects) mainly via selectively inhibiting the CDK5-mediated phosphorylation of PPARγ-Ser273 and up-regulating the expression of insulin-sensitive genes Adiponectin and Glut4, yet lacking the classical full agonism to induce the adipogenesis and the expression of key adipogenic genes including PPARγ, aP2, CD36, LPL, C/EBPα and FASN. Further validation led to the final recognition of its (R)-configured isomer as the potential conformational form. Subsequent molecular docking studies revealed a unique hydrogen-bonding network of (R)-YG-C-20 with three full PPARγ agonism-unrelated residues, especially with PPARγ-Ser273 phosphorylation-associated site Ser342, which not only gives a clear verification for our structure-based design but also provides a proof of concept for the abovementioned molecular mechanism.

Application and improvement of Canny edge-detection algorithm for exterior wall hollowing detection using infrared thermal images

The periodic hollowing inspection of the existing building exterior wall is crucial for public safety and building energy conservation. Due to its non-destructive and intuition advantage, the infrared thermal detection is proposed to be an ideal survey method. However, much manual participation is required to distinguish the hollowing flaw relying on empirical judgment, and a heavy burden comes up when large-area diagnosis required. In order to improve the efficiency of hollowing detection, this investigation developed the Canny algorithm to realize the automatic processing using the computer instead of manual judgment.At first, reasonable pieces of setting advice were given to get more clear hollowing region contours with the final recognition outcome comparison of different processing methods for each step. Besides, it was found that the hollowing contour gradient values are lower and exist in a short interval, and the segmentation threshold value was critical in the Canny edge-detection algorithm, which highly restricted the speed of processing large amounts of infrared images. To improve the efficiency of thermal image recognition, a threshold selection method based on the local maximum inter-class variance algorithm was introduced into the Canny edge-detection algorithm. Compared with Sobel, Roberts, Prewitt, and LoG, the proposed algorithm presented a better performance in the identification of hollowing edge contour according to the verification based on three cases. It revealed that the improved Canny edge-detection algorithm was effective and efficient, which could not only eliminate the influence of subjective factors but also achieve full-automatic and batch processing.

Conventional Neural Network-Based Radio Frequency Fingerprint Identification Using Raw I/Q Data

Radio frequency (RF) fingerprint identification is a nonpassword authentication method based on the physical layer of communication devices. Deep learning methods have thrown new light on RF fingerprint identification. In this paper, a conventional neural network- (CNN-) based RF identification model is proposed. The CNN models are designed to be lightweight. Raw data that reflects the characteristics of theIchannel, theQchannel, and the 2-dimensionalI+Qdata is successively fed into a CNN model. Therefore, three submodels are generated. The final predictive labels are determined by the results of the three submodels through a voting scheme. Experimental results have demonstrated that in the SNR setting at 5 dB, the final recognition accuracy of four transmit devices could achieve as high as 97.25%, while the identification accuracies based on theIchannel data,Qchannel data, andI+Qchannel data are 94.5%, 95%, and 94.5%, respectively. The training time for the 4 devices is around 30 seconds.

Incremental learning of upper limb action pattern recognition based on mechanomyography

Aiming at the problem that the traditional mechanomyography (MMG) pattern recognition model for upper limb rehabilitation action has poor recognition effect on non-identical distribution test data, this research proposes an incremental learning method for deep learning based on MMG. By collecting 8-channels of MMG of 4 types of hand movements (wrist flexion, wrist extension, wrist ulnar flexion, and wrist radial flexion), after signal preprocessing, feature extraction and dimensionality reduction, 12 groups of non-identical distributed data were obtained. The model was trained by using deep neural network (DNN), and five commonly used machine learning algorithms were used as comparison for incremental training. Finally, the recognition rate of DNN was 88.25%, and the final recognition rates of Passive Aggressive (PA), Incremental Support Vector Machine (ISVM), Perceptron, Bernoulli Naive Bayes (BNB) and Multinomial Naive Bayes (MNB) were 85.94%, 84.82%, 81.01%, 73.07% and 61.20% respectively. Among them, both DNN and PA had better upward trends, and DNN had the highest final recognition rate. By comparing the confusion matrices before and after DNN incremental learning, it could be seen that DNN incremental training can significantly improve the accuracy and precision in confusion matrix. The experimental results demonstrate that the method of incremental learning can not only achieve the recognition of non-identically distributed test data, but also improve the recognition rates and the generalization performance of the model.

Action Recognition Model in Choral Conducting Teaching in Colleges and Universities under New Media Environment.

In current college music education, choral conducting is a required course for students. The course implementation aims to cultivate excellent and high-quality choral conductors. The requirements for choral conducting teaching in college music education under the new media environment have been further improved. First, this study gives the value of applying new media technology in choral conducting teaching in colleges and universities. Then, based on the key point that choral conductors' expression of music mainly relies on gestural language, an action recognition model in college choral conducting teaching is proposed. The model is designed with an adaptive deep graph convolution model, and a spatio-temporal convolution submodel with a small number of parameters is created using group convolution. After the trained teacher model is obtained, the spatio-temporal convolutional submodel with fewer parameters is trained using the knowledge distillation method combined with data augmentation techniques. The final action recognition fusion model is obtained using the linear fusion method. The experimental results demonstrate that the proposed model can recognize the movements in college choral conducting teaching with higher performance than other existing models, which provides effective guidance for college choral conducting teaching in the new media environment.

Discriminative Feature Mining and Enhancement Network for Low-Resolution Fine-Grained Image Recognition

Existing fine-grained image recognition methods are difficult to learn complete discriminative features from low-resolution (LR) data, because the original subtle inter-class distinctions become slimmer with the reduction of the image resolution. Besides, existing methods of LR fine-grained image recognition and general LR image recognition only consider the restoration and extraction of global discriminative features, ignoring unreliable local fine-grained details can be detrimental to final recognition. To address the above problems, we propose a multi-tasking framework, discriminative feature mining and enhancement network (DME-Net), for the LR fine-grained image recognition task, which aims to capture the reliable object descriptions from macro and micro perspectives, respectively. Macroscopically, we train the framework’s ability to recover and extract global discriminative features based on the whole images. Microscopically, we purposefully reinforce the framework’s ability to repair and capture the local discriminative details on the mined informative parts. To precisely excavate the most potential parts, we design an informative part mining (IPM) module, in which we firstly employ a part generation layer to predict several part masks that focus on different discriminative parts under the guidance of discrepancy loss and discriminant loss. Then we introduce a part selection (PS) submodule to further screen out a group of most informative parts from the predicted part masks according to their corresponding scores, which measure the semantic correlation degree of each part to the others. Experimental results on three benchmark datasets and one retail product dataset consistently show that our proposed framework can significantly boost the performance of the baseline model. Besides, extensive ablation studies are conducted, which further prove the effectiveness of each component of our designs.

Retrieval practice plus feedback benefits a third language vocabulary learning

ABSTRACT Three experiments explored how retrieval practice and corrective feedback affect a third language (L3) vocabulary learning. In the first two experiments, Chinese–English bilinguals without prior French language experience studied English (Second Language, L2)–French (L3) word pairs in repeated studying or retrieval practice without (Experiment 1)/with (Experiment 2) corrective feedback. A final recognition test was administered after a 30-min retention interval. Experiment 3 was identical to Experiment 2 in all respects, except that Chinese (First Language, L1)–French (L3) word pairs were used as study materials. Experiment 1 found that repeated studying led to greater accuracy and faster response time on the final test than retrieval practice without feedback. However, the opposite results were found in Experiment 2: retrieval practice with feedback yielded superior performance compared with repeated studying. The benefit of retrieval practice plus feedback was also found in Experiment 3, while it produced larger gains than Experiment 2. These results demonstrate that retrieval practice plus feedback is a powerful learning strategy, and L1–L3 learning is more effective than L2–L3 learning for L3 vocabulary learning. Our findings support the bifurcation model of the retrieval practice effect.

Online in situ detection of kitchen environment based on laser-induced breakdown spectroscopy

The damage of kitchen oil fume to the human body and environment cannot be ignored. Based on laser-induced breakdown spectroscopy (LIBS), five kitchen environments are online in situ detected, including the air scene, fry scene, grill scene, steam scene, and stew scene. In the spectra, characteristic elements such as C, H, O, and N are detected in the fry scene containing oil fume, and metal elements such as Mg, Ca, K, and Na are observed in the grill scene containing charcoal smoke. The spectra of five kitchen environments are tested and compared. In the measurement, except for the air scene, obvious carbon–nitrogen molecular spectral lines are detected. LIBS is combined with principal component analysis and backpropagation artificial neural network system to detect and analyze kitchen fumes. Finally, five kitchen scenes are analyzed and identified based on this system, and the final recognition accuracy is 98.60%.

STCA: Utilizing a spatio-temporal cross-attention network for enhancing video person re-identification

Video-based re-identification (ReID) is a crucial task in computer vision that draws increasing attention due to advances in deep learning (DL) and modern computational devices. Despite recent success with CNN architectures, single models (e.g., 2D-CNNs or 3D-CNNs) alone failed to leverage temporal information with spatial cues. This is due to uncontrolled surveillance scenarios and variable poses leading to inevitable misalignment of ROIs across the tracklets, which is accompanied by occlusion and motion blur. In this context, designing temporal and spatial cues for two different models and their combinations can be beneficial, considering the global of a video-tracklet. 3D-CNNs allow encoding of temporal information while 2D-CNNs extract spatial or appearance information. In this paper, we propose a Spatio-Temporal Cross Attention (STCA) network to utilize both 2D-CNNs and 3D-CNNs that calculate the cross attention mapping both from the layer of 3D-CNNs and 2D-CNNs along a person's trajectory to gate the following layers of 2D-CNNs; and highlight relevant appearance features for the person ReID. Given an input tracklet, the proposed cross attention (CA) is able to capture the salient regions that propagate throughout the tracklet to obtain the global view. This provides a spatio-temporal attention approach that can be dynamically aggregated with spatial features of 2D-CNNs to perform finer-grained recognition. Additionally, we exploit the advantage of utilizing cosine similarity while triplet sampling as well as for calculating the final recognition score. Experimental analyses on three challenging benchmark datasets indicate that integrating spatio-temporal cross attention into the state-of-the-art video ReID backbone CNN architecture allows for improving their recognition accuracy.

Sports Biology Seminar of Three-dimensional Movement Characteristics of Yoga Standing Based on Image Recognition

With the exponential improvement of the integration capabilities of electronic hardware devices, digital images have become an indispensable information carrier, thus promoting the development of image recognition, detection, or tracking technologies. At present, there is a blank in the field of image recognition measured by the action standard of yoga standing three-dimensional movement, which will bring huge room for researchers to play. Based on the traditional image recognition technology, there are cases where missed detection or false detection cannot be correctly identified and made a correct judgment. In this paper, the ability to capture the action is improved by optimizing the network for the positioning of the stereoscopic action target of the yoga station. The linear discriminant analysis (LDA) after adding optimization is used to reduce the dimensionality of the captured image data, which is beneficial to improving image recognition rate and reducing image loss rate. Through the analysis of the three-dimensional movement of the yoga station, the model based on the improved algorithm is compared with the traditional model, and the final image recognition accuracy is improved compared with the network model before the improvement. The image recognition error rate steadily tends to 5%, and the loss rate is also below 2%. Through the optimized convolutional neural network, this paper can accurately capture the image position, and the recognition rate has also been greatly improved, which can provide a reference for future research in other fields.

A Multimodal Information Fusion Model for Robot Action Recognition with Time Series

The current robotics field, led by a new generation of information technology, is moving into a new stage of human-machine collaborative operation. Unlike traditional robots that need to use isolation rails to maintain a certain safety distance from people, the new generation of human-machine collaboration systems can work side by side with humans without spatial obstruction, giving full play to the expertise of people and machines through an intelligent assignment of operational tasks and improving work patterns to achieve increased efficiency. The robot’s efficient and accurate recognition of human movements has become a key factor in measuring robot performance. Usually, the data for action recognition is video data, and video data is time-series data. Time series describe the response results of a certain system at different times. Therefore, the study of time series can be used to recognize the structural characteristics of the system and reveal its operation law. As a result, this paper proposes a time series-based action recognition model with multimodal information fusion and applies it to a robot to realize friendly human-robot interaction. Multifeatures can characterize data information comprehensively, and in this study, the spatial flow and motion flow features of the dataset are extracted separately, and each feature is input into a bidirectional long and short-term memory network (BiLSTM). A confidence fusion method was used to obtain the final action recognition results. Experiment results on the publicly available datasets NTU-RGB + D and MSR Action 3D show that the method proposed in this paper can improve action recognition accuracy.

Pulse Wave Signal Identification Method Based on EMD-CNN Model

in the research on the method of using pulse wave signal for identity recognition, this paper uses empirical mode decomposition (EMD) to decompose the pulse signal to obtain multiple intrinsic mode function (IMF) components distributed in different frequency bands. Each IMF component is screened and reconstructed with the correlation coefficient as the index, The IMF components used to reconstruct the pulse wave signal are identified by multi-channel parallel convolution neural network. In this paper, 30 healthy adults were identified, and the final recognition accuracy was 92.38%.