Recognition Performance Research Articles

Detecting emotions through EEG signals based on modified convolutional fuzzy neural network

Emotion is a human sense that can influence an individual’s life quality in both positive and negative ways. The ability to distinguish different types of emotion can lead researchers to estimate the current situation of patients or the probability of future disease. Recognizing emotions from images have problems concealing their feeling by modifying their facial expressions. This led researchers to consider Electroencephalography (EEG) signals for more accurate emotion detection. However, the complexity of EEG recordings and data analysis using conventional machine learning algorithms caused inconsistent emotion recognition. Therefore, utilizing hybrid deep learning models and other techniques has become common due to their ability to analyze complicated data and achieve higher performance by integrating diverse features of the models. However, researchers prioritize models with fewer parameters to achieve the highest average accuracy. This study improves the Convolutional Fuzzy Neural Network (CFNN) for emotion recognition using EEG signals to achieve a reliable detection system. Initially, the pre-processing and feature extraction phases are implemented to obtain noiseless and informative data. Then, the CFNN with modified architecture is trained to classify emotions. Several parametric and comparative experiments are performed. The proposed model achieved reliable performance for emotion recognition with an average accuracy of 98.21% and 98.08% for valence (pleasantness) and arousal (intensity), respectively, and outperformed state-of-the-art methods.

FetchEEG: a hybrid approach combining feature extraction and temporal-channel joint attention for EEG-based emotion classification.

ObjectiveElectroencephalogram (EEG) analysis has always been an important tool in neural engineering, and the recognition and classification of human emotions are one of the important tasks in neural engineering. EEG data, obtained from electrodes placed on the scalp, represent a valuable resource of information for brain activity analysis and emotion recognition. Feature extraction methods have shown promising results, but recent trends have shifted toward end-to-end methods based on deep learning. However, these approaches often overlook channel representations, and their complex structures pose certain challenges to model fitting.ApproachTo address these challenges, this paper proposes a hybrid approach named FetchEEG that combines feature extraction and temporal-channel joint attention. Leveraging the advantages of both traditional feature extraction and deep learning, the FetchEEG adopts a multi-head self-attention mechanism to extract representations between different time moments and channels simultaneously. The joint representations are then concatenated and classified using fully-connected layers for emotion recognition. The performance of the FetchEEG is verified by comparison experiments on a self-developed dataset and two public datasets.ResultsIn both subject-dependent and subject-independent experiments, the FetchEEG demonstrates better performance and stronger generalization ability than the state-of-the-art methods on all datasets. Moreover, the performance of the FetchEEG is analyzed for different sliding window sizes and overlap rates in the feature extraction module. The sensitivity of emotion recognition is investigated for three- and five-frequency-band scenarios. Finally, we perform visualization of the channel-attention to project the model's reliance on different channels onto scalp topographic maps.SignificanceFetchEEG is a novel hybrid method based on EEG for emotion classification, which significantly improves the performance of recognition and is effective and feasible.

Objective Video Quality Assessment Method for Object Recognition Tasks

In the field of video quality assessment for object recognition tasks, accurately predicting the impact of different quality factors on recognition algorithms remains a significant challenge. Our study introduces a novel evaluation framework designed to address this gap by focussing on machine vision rather than human perceptual quality metrics. We used advanced machine learning models and custom Video Quality Indicators to enhance the predictive accuracy of object recognition performance under various conditions. Our results indicate a model performance, achieving a mean square error (MSE) of 672.4 and a correlation coefficient of 0.77, which underscores the effectiveness of our approach in real-world scenarios. These findings highlight not only the robustness of our methodology but also its potential applicability in critical areas such as surveillance and telemedicine.

Toward an Intuitive Device for Construction Hazard Recognition Management: Eye Fixation–Related Potentials in Reinvestigation of Hazard Recognition Performance Prediction

Toward an Intuitive Device for Construction Hazard Recognition Management: Eye Fixation–Related Potentials in Reinvestigation of Hazard Recognition Performance Prediction

Adaptive Scale and Correlative Attention PointPillars: An Efficient Real-Time 3D Point Cloud Object Detection Algorithm

Recognizing 3D objects from point clouds is a crucial technology for autonomous vehicles. Nevertheless, LiDAR (Light Detection and Ranging) point clouds are generally sparse, and they provide limited contextual information, resulting in unsatisfactory recognition performance for distant or small objects. Consequently, this article proposes an object recognition algorithm named Adaptive Scale and Correlative Attention PointPillars (ASCA-PointPillars) to address this problem. Firstly, an innovative adaptive scale pillars (ASP) encoding method is proposed, which encodes point clouds using pillars of varying sizes. Secondly, ASCA-PointPillars introduces a feature enhancement mechanism called correlative point attention (CPA) to enhance the feature associations within each pillar. Additionally, a data augmentation algorithm called random sampling data augmentation (RS-Aug) is proposed to solve the class imbalance problem. The experimental results on the KITTI 3D object dataset demonstrate that the proposed ASCA-PointPillars algorithm significantly boosts the recognition performance and RS-Aug effectively enhances the training effects on an imbalanced dataset.

Privacy-preserving federated transfer learning for defect identification from highly imbalanced image data in additive manufacturing

Defect identification is a crucial task for process monitoring and quality evaluation in additive manufacturing (AM). Deep learning (DL) has shown great potential for diverse fields, but some challenges have hindered the application in AM process monitoring. Firstly, DL-based methods are driven by big data and require a large number of training data. However, in reality, defective data is often rare, so different AM manufacturers can collaborate to train a global model for detecting defects. Nevertheless, due to privacy concerns and conflicts of interest, data-sharing between different manufacturers might be dangerous. Additionally, the heterogeneities among manufacturers’ data leads to the domain shift, making it difficult to obtain a well-generalized model. Moreover, the imbalance issue of powder-spreading defects seriously damages the performance of the defect recognition. In this paper, we proposed FTLAM, a federated transfer learning method, to address the above issues. Concretely, to solve the data privacy concerns, federated learning (FL) is utilized to collaboratively train a global model between different AM manufacturers without exchanging original data. Furthermore, a client-oriented transfer method is proposed to mitigate the heterogeneities across multiple clients and improve the model's generalization. Meanwhile, a multi-loss joint optimization approach is designed to alleviate the data imbalance. Extensive experiments on laser powder bed fusion (LPBF) image datasets of powder-spreading have demonstrated that our FTLAM can obtain satisfactory performance of powder spreading defect identification.

A 4D spontaneous micro‐expression database: Establishment and evaluation

AbstractMicro‐expressions are spontaneous and unconscious facial movements that reveal individuals’ genuine inner emotions. They hold significant potential in various psychological testing fields. As the face is a 3D deformation object, the emergence of facial expression leads to spatial deformation of the face. However, existing databases primarily offer 2D video sequences, limiting descriptions of 3D spatial information related to micro‐expressions. Here, a new micro‐expression database is proposed, which contains 2D image sequences and corresponding 3D point cloud sequences. These samples were classified using both an objective method based on the facial action coding system and a non‐objective emotion classification method that considers video contents and participants’ self‐reports. A variety of feature extraction techniques are applied to 2D data, including traditional algorithms and deep learning methods. Additionally, a novel local curvature‐based algorithm is developed to extract 3D spatio‐temporal deformation features from the 3D data. The authors evaluated the classification accuracies of these two features individually and their fusion results under leave‐one‐subject‐out (LOSO) and tenfold cross‐validation. The results demonstrate that fusing 3D features with 2D features results in improved recognition performance compared to using 2D features alone.

Deep Fake Detecting System

Abstract: Over the past few decades, there have been rapid breakthroughs in AI, machine learning, and deep learning, which have led to the development of new tools and methodologies for manipulating multimedia. Technology has generally been employed for beneficial purposes, such as education and entertainment, but dishonest users have also exploited it for darker or illicit purposes. For example, incredibly realistic-looking, well-produced fake movies, images, or audio have been made to spread false information, incite political unrest and hatred, or even to harass and blackmail people. The extremely replicated, realistic, and edited videos have been dubbed as "Deepfake" in recent times. Since then, several approaches to resolving the problems raised by Deepfake have been described in length in the literature. We conduct a systematic literature review (SLR) in this work to give a current synopsis of the Deepfake detection research projects. We provide an overview of 112 pertinent publications from 2018 to 2020 that showcased various methodologies. For analysis, we divide them into four categories: deep learning-based approaches, traditional machine learning-based approaches, statistical-based approaches, and blockchain-based approaches. We also evaluate the pattern recognition performance of several algorithms on various datasets, and we find that deep learningbased approaches outperform other approaches in Deepfake detection

The Effects of AI-Driven Face Restoration on Forensic Face Recognition

In biometric recognition, face recognition is a mature and widely used technique that provides a fast, accurate, and reliable method for human identification. This paper aims to study the effects of face image restoration for forensic face recognition and then further analyzes the advantages and limitations of the four state-of-the-art face image restoration methods in the field of face recognition for forensic human image identification. In total, 100 face image materials from an open-source face image dataset are used for experiments. The Gaussian blur processing is applied to simulate the effect of blurred face images in actual cases of forensic human image identification. Four state-of-the-art AI-driven face restoration methods are used to restore the blurred face images. We use three mainstream face recognition systems to evaluate the recognition performance changes of the blurred face images and the restored face images. We find that although face image restoration can effectively remove facial noise and blurring effects, the restored images do not significantly improve the recognition performance of the face recognition systems. Face image restoration may change the original features in face images and introduce new made-up image features, thereby affecting the accuracy of face recognition. In current conditions, the improvement in face image restoration on the recognition performance of face recognition systems is limited, but it still has a positive role in the application of forensic human image identification.

Toward Synthetic Physical Fingerprint Targets

Biometric fingerprint identification hinges on the reliability of its sensors; however, calibrating and standardizing these sensors poses significant challenges, particularly in regards to repeatability and data diversity. To tackle these issues, we propose methodologies for fabricating synthetic 3D fingerprint targets, or phantoms, that closely emulate real human fingerprints. These phantoms enable the precise evaluation and validation of fingerprint sensors under controlled and repeatable conditions. Our research employs laser engraving, 3D printing, and CNC machining techniques, utilizing different materials. We assess the phantoms’ fidelity to synthetic fingerprint patterns, intra-class variability, and interoperability across different manufacturing methods. The findings demonstrate that a combination of laser engraving or CNC machining with silicone casting produces finger-like phantoms with high accuracy and consistency for rolled fingerprint recordings. For slap recordings, direct laser engraving of flat silicone targets excels, and in the contactless fingerprint sensor setting, 3D printing and silicone filling provide the most favorable attributes. Our work enables a comprehensive, method-independent comparison of various fabrication methodologies, offering a unique perspective on the strengths and weaknesses of each approach. This facilitates a broader understanding of fingerprint recognition system validation and performance assessment.

HyperFace: A Deep Fusion Model for Hyperspectral Face Recognition

Face recognition has been well studied under visible light and infrared (IR) in both intra-spectral and cross-spectral cases. However, how to fuse different light bands for face recognition, i.e., hyperspectral face recognition, is still an open research problem, which has the advantages of richer information retention and all-weather functionality over single-band face recognition. Thus, in this research, we revisit the hyperspectral recognition problem and provide a deep learning-based approach. A new fusion model (named HyperFace) is proposed to address this problem. The proposed model features a pre-fusion scheme, a Siamese encoder with bi-scope residual dense learning, a feedback-style decoder, and a recognition-oriented composite loss function. Experiments demonstrate that our method yields a much higher recognition rate than face recognition using only visible light or IR data. Moreover, our fusion model is shown to be superior to other general-purpose image fusion methods that are either traditional or deep learning-based, including state-of-the-art methods, in terms of both image quality and recognition performance.

Development of a facial recognition model using optimized convolution neural network with Tiny-ML and fingerprint authentication for safe lock entry

Safe boxes have been used for centuries to protect valuable items, and their design and construction have evolved over time to improve security. In this study, we carried out the design and implementation of a facial recognition technology using optimized convolutional neural networks (CNNs) with TinyML and fingerprint. The facial recognition system is designed using pre-trained CNN architectures then optimized with TinyML to make it fit into resource constrained edge device, the fingerprint recognition is designed using ridge matching biometric algorithm which fits in perfectly into the microcontroller-based system. The safe box was fabricated using sheet metal and insulated using heat resistive material (fiber glass), components like OV2640 camera module, ESP32-S3, R305 Fingerprint sensor, 4.3 inches touch screen TFT display and a charging module to manage and charge the systems battery were used. The system was implemented to be a level two security and two-way verification system using secured Wi-Fi to access the system before the facial and fingerprint recognition functionality is activated, the sequence of operation can either be facial recognition before fingerprint recognition or fingerprint recognition before facial recognition. Evaluating the system’s facial recognition performance with confusion matrix, of all the optimized CNN architectures like VGGNet, SENet, RestNet, MobileNet-V2, EfficientNet-B0, SENet Architecture performed best outputting an accuracy of 98%, precision of 92%, also SENet unoptimized model performed same, MobileNet-V2 performed optimally in the inference speed test with a speed of 15ms/images with the fingerprint also evaluated using confusion matrix and perform with 98% accuracy. It was noticed that the inference speed of the unoptimized model is faster than that of the TinyML optimized model which is justifiable based on the processing clock speed, memory usage, GPU and architecture of the microcontroller/microprocessor. From the result obtained TinyML proves to be suitable mean for edge AI (artificial intelligence) and is recommended for future research and development.

Machine learning for hand pose classification from phasic and tonic EMG signals during bimanual activities in virtual reality

Myoelectric prostheses have recently shown significant promise for restoring hand function in individuals with upper limb loss or deficiencies, driven by advances in machine learning and increasingly accessible bioelectrical signal acquisition devices. Here, we first introduce and validate a novel experimental paradigm using a virtual reality headset equipped with hand-tracking capabilities to facilitate the recordings of synchronized EMG signals and hand pose estimation. Using both the phasic and tonic EMG components of data acquired through the proposed paradigm, we compare hand gesture classification pipelines based on standard signal processing features, convolutional neural networks, and covariance matrices with Riemannian geometry computed from raw or xDAWN-filtered EMG signals. We demonstrate the performance of the latter for gesture classification using EMG signals. We further hypothesize that introducing physiological knowledge in machine learning models will enhance their performances, leading to better myoelectric prosthesis control. We demonstrate the potential of this approach by using the neurophysiological integration of the “move command" to better separate the phasic and tonic components of the EMG signals, significantly improving the performance of sustained posture recognition. These results pave the way for the development of new cutting-edge machine learning techniques, likely refined by neurophysiology, that will further improve the decoding of real-time natural gestures and, ultimately, the control of myoelectric prostheses.

Human Activity Recognition Based on Deep Learning Regardless of Sensor Orientation

In recent years, the continuous progress of wireless communication and sensor technology has enabled sensors to be better integrated into mobile devices. Therefore, sensor-based Human Activity Recognition (HAR) has attracted widespread attention among researchers, especially in the fields of wearable technology and ubiquitous computing. In these applications, mobile devices’ built-in accelerometers and gyroscopes have been typically used for human activity recognition. However, devices such as smartphones were placed in users’ pockets and not fixed to their bodies, and the resulting changes in the orientation of the sensors due to users’ habits or external forces can lead to a decrease in the accuracy of activity recognition. Unfortunately, there is currently a lack of publicly available datasets specifically designed to address the issue of device angle change. The contributions of this study are as follows. First, we constructed a dataset with eight different sensor placement angles using accelerometers and gyroscopes as a prerequisite for the subsequent research. Second, we introduced the Madgwick algorithm to extract quaternion mode features and alleviate the impact of angle changes on recognition performance by fusing raw accelerometer data and quaternion mode features. The resulting study provides a comprehensive analysis. On the one hand, we fine-tuned ResNet and tested its stability on our dataset, achieving a recognition accuracy of 97.13%. We included two independent experiments, one for user-related scenarios and the other for user-independent scenarios. In addition, we validated our research results on two publicly available datasets, demonstrating that our method has good generalization performance.

Modality aware contrastive learning for multimodal human activity recognition

SummaryHuman activity recognition is a well‐established research problem in ubiquitous computing. The increased dependency on various smart devices in our daily lives allows us to investigate the sensor data world produced by multimodal sensors embedded in smart devices. However, the raw sensor data are often unlabeled and annotating this vast amount of data are a costly exercise that can often lead to privacy breaches. Self‐supervised learning‐based approaches are at the forefront of learning semantic representation from unlabeled sensor data, including when applied to human activity recognition tasks. As inferring human activity depends on multimodal sensors, addressing the modality difference and inter‐modality dependencies in a model is an important process. This paper proposes a novel self‐supervised learning approach, modality aware contrastive learning (MACL), for representation learning using multimodal sensor data. The approach uses different sensing modalities to create different views of an input signal. Thus, the model is able to learn the representations by maximizing the similarity among different sensing modalities of the same input signal. Extensive experiments were performed on four publicly available human activity recognition data sets to verify the effectiveness of our proposed MACL method. The experimental evaluation results show that the MACL method attains a comparable performance for human activity recognition to the compared baseline models, directly exceeding the performance of models using standard augmentation transformation strategies.

The underlying neuropsychological and neural correlates of the impaired Chinese reading skills in children with attention deficit hyperactivity disorder.

Impaired basic academic skills (e.g., word recognition) are common in children with Attention Deficit Hyperactivity Disorder (ADHD). The underlying neuropsychological and neural correlates of impaired Chinese reading skills in children with ADHD have not been substantially explored. Three hundred and two children with ADHD (all medication-naïve) and 105 healthy controls underwent the Chinese language skill assessment, and 175 also underwent fMRI scans (84 ADHD and 91 controls). Between-group and mediation analyses were applied to explore the interrelationships of the diagnosis of ADHD, cognitive dysfunction, and impaired reading skills. Five ADHD-related brain functional networks, including the default mode network (DMN) and the dorsal attention network (DAN), were built using predefined regions of interest. Voxel-based group-wise comparisons were performed. The ADHD group performed worse than the control group in word-level reading ability tests, with lower scores in Chinese character recognition (CR) and word chains (WS) (all P < 0.05). With full-scale IQ and sustained attention in the mediation model, the direct effect of ADHD status on the CR score became insignificant (P = 0.066). The underlying neural correlates for the orthographic knowledge (OT) and CR differed between the ADHD and the control group. The ADHD group tended to recruit more DMN regions to maintain their reading performance, while the control group seemed to utilize more DAN regions. Children with ADHD generally presented impaired word-level reading skills, which might be caused by impaired sustained attention and lower IQ. According to the brain functional results, we infer that ADHD children might utilize a different strategy to maintain their orthographic knowledge and character recognition performance.

Lightweight object detection network for multi‐damage recognition of concrete bridges in complex environments

AbstractTo solve the challenges of low recognition accuracy, slow speed, and weak generalization ability inherent in traditional methods for multi‐damage recognition of concrete bridges, this paper proposed an efficient lightweight damage recognition model, constructed by improving the you only look once v4 (YOLOv4) with MobileNetv3 and fused inverted residual blocks, named YOLOMF. First, a novel lightweight network named MobileNetv3 with fused inverted residual (MobileNetv3‐FusedIR) is constructed as the backbone network for YOLOMF. This is achieved by integrating the fused mobile inverted bottleneck convolution (Fused‐MBConv) into the shallow layers of MobileNetv3. Second, the standard convolution in YOLOv4 is replaced with the depthwise separable convolution, resulting in a reduction in the number of parameters and complexity of the model. Third, the effects of different activation functions on the damage recognition performance of YOLOMF are thoroughly investigated. Finally, to verify the effectiveness of the proposed method in complex environments, a data enhancement library named Imgaug is used to simulate concrete bridge damage images under challenging conditions such as motion blur, fog, rain, snow, noise, and color variations. The results indicate that the YOLOMF shows excellent multi‐damage recognition proficiency for concrete bridges across varying field‐of‐view sizes as well as complex environmental conditions. The detection speed of YOLOMF reaches 85f/s, facilitating effective real‐time multi‐damage detection for concrete bridges under complex environments.

Electrocatalyst for hydrogen evolution and H2O2 recognition on composite carbon paste electrodes with three new Ag(I) MOFs

Three new Ag(I) metal-organic frameworks (Ag-MOFs) containing bisbenzimidazole-derived ligand, namely, {[Ag2(L)(TP)](H2O)}n (1), {[Ag2(L)(HBTC))(H2O)](CH3CH2OH)(H2O)}n (2) and {[Ag4(L)2(IP)2](CH3CN)3(H2O)16}n (3) (L = bis(1-(pyridin-4-ylmethyl)-benzimidazol-2-yl methyl) ether, TP = Terephthalate, BTC = Benzenetricarboxylate, IP = Isophthalate), have been synthesized by volatilization method and systematic adjustment of solvent and aromatic polyacid co-ligand. Single crystal structural analysis showed that although three Ag-MOFs have a three-dimensional (3D) network backbone structure, their topological structures are different. Electrocatalytic hydrogen evolution reaction (HER) of carbon paste composite electrodes (Ag-MOF-1∼3@CPE) prepared by mixing graphite powder with Ag-MOFs were investigated in 0.5 M H2SO4 electrolyte. The HER measurements show that the overpotential η10293K of Ag-MOF-1∼3@CPE are -634, -533 and -615 mV compared with sCPCE (blank electrode, -966 mV), and the Tafel slope of sCPCE and Ag-MOF-1∼3@CPE were 276, 251, 123 and 220 mV dec−1, respectively. The results prove that Ag-MOF-1∼3@CPE could effectively catalyze and accelerate the HER behavior, with electrocatalytic activity order being Ag-MOF-2@CPE > Ag-MOF-3@CPE > Ag-MOF-1@CPE > sCPE. The highest HER activity of Ag-MOF-2@CPE can be attributed to the fact that Ag-MOF-2 contains coordination water molecule, while the higher activity of Ag-MOF-3@CPE than Ag-MOF-1@CPE is due to steric hindrance. Furthermore, the recognition performance of Ag-MOF-1∼3@CPE for H2O2 was further studied by chronoamperometry in 0.2 M phosphate buffer solution (PBS, pH = 6). The three sensors can detect H2O2 in a linear range from 0.5 μM to 4 mM with sensitivities of 103, 34.5 and 138 μA·mM−1·cm−2, and also revealed long-term stability and good selectivity. Our study provides a new approach for designing efficient, non-precious metal electrochemical catalysts.

A single-joint multi-task motor imagery EEG signal recognition method based on Empirical Wavelet and Multi‐Kernel Extreme Learning Machine

BackgroundIn the pursuit of finer Brain-Computer Interface commands, research focus has shifted towards classifying EEG signals for multiple tasks. While single-joint multitasking motor imagery provides support, distinguishing between EEG signals from the same joint remains challenging due to their similar brain spatial distribution. New methodWe designed experiments involving three motor imagery tasks—wrist extension, wrist flexion, and wrist abduction—with six participants. Based on this, a single-joint multi-task motor imagery EEG signal recognition method using Empirical Wavelet Decomposition and Multi-Kernel Extreme Learning Machine is proposed. This method employs Empirical Wavelet Decomposition (EWT) for modal decomposition, screening, and reconstruction of raw EEG signals, feature extraction using Common Spatial Patterns (CSP), and classification using Multi-Kernel Extreme Learning Machine (MKELM). ResultsAfter EWT processing, differences in time and frequency characteristics between EEG signals of different classes were enhanced, with the MKELM model achieving an average recognition accuracy of 91.93 %. Comparison with other methods and conclusionsWe compared EWT with Empirical Mode Decomposition (EMD), Variational Mode Decomposition (VMD), Local Mean Decomposition (LMD), and Wavelet Packet Decomposition (WPD). The results showed that the differences between various types of EEG signals processed by EWT were the most pronounced. The MKELM model outperformed traditional machine learning models such as Extreme Learning Machine (ELM), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Linear Discriminant Analysis (LDA) in terms of recognition performance, and also exhibited faster training speeds than deep learning models such as Bayesian Convolutional Neural Network (BCNN) and Attention-based Dual-scale Fusion Convolutional Neural Network (ADFCNN). In summary, the proposed method provides a new approach for achieving finer Brain-Computer Interface commands.

Deep learning-based anatomical position recognition for gastroscopic examination.

The gastroscopic examination is a preferred method for the detection of upper gastrointestinal lesions. However, gastroscopic examination has high requirements for doctors, especially for the strict position and quantity of the archived images. These requirements are challenging for the education and training of junior doctors. The purpose of this study is to use deep learning to develop automatic position recognition technology for gastroscopic examination. A total of 17182 gastroscopic images in eight anatomical position categories are collected. Convolutional neural network model MogaNet is used to identify all the anatomical positions of the stomach for gastroscopic examination The performance of four models is evaluated by sensitivity, precision, and F1 score. The average sensitivity of the method proposed is 0.963, which is 0.074, 0.066 and 0.065 higher than ResNet, GoogleNet and SqueezeNet, respectively. The average precision of the method proposed is 0.964, which is 0.072, 0.067 and 0.068 higher than ResNet, GoogleNet, and SqueezeNet, respectively. And the average F1-Score of the method proposed is 0.964, which is 0.074, 0.067 and 0.067 higher than ResNet, GoogleNet, and SqueezeNet, respectively. The results of the t-test show that the method proposed is significantly different from other methods (p< 0.05). The method proposed exhibits the best performance for anatomical positions recognition. And the method proposed can help junior doctors meet the requirements of completeness of gastroscopic examination and the number and position of archived images quickly.