Average Recognition Accuracy Research Articles

Smart agriculture: An intelligent approach for apple leaf disease identification based on convolutional neural network

AbstractPlant diseases pose a significant threat to global agricultural productivity and food safety. Early detection and accurate identification of these diseases are essential for effective disease management strategies. Traditional plant disease identification mainly relies on manual observation and experienced expert judgement, which has the disadvantages of being time‐consuming, labour‐intensive and low efficiency. Given the above problems, this study proposes a method for identifying apple leaf diseases based on a convolutional neural network combining hybrid attention and bidirectional long short‐term memory (BiLSTM). Appropriate apple leaf disease samples were selected from multiple public data sets to form an experimental data set. Then, the data set is imported into the improved convolutional neural network for training. Based on the original ResNet18 model, a new convolutional neural network, AppleNet, is constructed by adding a hybrid attention module and modifying the classifier structure. The experimental results show that the average recognition accuracy of AppleNet is 94.66%, which is 2.47% higher than that of the ResNet18 network. In addition, the training time of the model is only slightly increased. The ablation experiment further verified the effectiveness of the model modification. Compared with other advanced models in recognition accuracy and model training time, the superiority of AppleNet is confirmed. This study verifies that deep learning has great potential and application prospects in plant disease identification and provides a new technical solution for intelligent and convenient plant disease identification.

SIMCB-Yolo: An Efficient Multi-Scale Network for Detecting Forest Fire Smoke

Forest fire monitoring plays a crucial role in preventing and mitigating forest disasters. Early detection of forest fire smoke is essential for a timely response to forest fire emergencies. The key to effective forest fire monitoring lies in accounting for the various levels of forest fire smoke targets in the monitoring images, enhancing the model’s anti-interference capabilities against mountain clouds and fog, and reducing false positives and missed detections. In this paper, we propose an improved multi-level forest fire smoke detection model based on You Only Look Once v5s (Yolov5s) called SIMCB-Yolo. This model aims to achieve high-precision detection of forest fire smoke at various levels. First, to address the issue of low precision in detecting small target smoke, a Swin transformer small target monitoring head is added to the neck of Yolov5s, enhancing the precision of small target smoke detection. Then, to address the issue of missed detections due to the decline in conventional target smoke detection accuracy after improving small target smoke detection accuracy, we introduced a cross stage partial network bottleneck with three convolutional layers (C3) and a channel block sequence (CBS) into the trunk. These additions help extract more surface features and enhance the detection accuracy of conventional target smoke. Finally, the SimAM attention mechanism is introduced to address the issue of complex background interference in forest fire smoke detection, further reducing false positives and missed detections. Experimental results demonstrate that, compared to the Yolov5s model, the SIMCB-Yolo model achieves an average recognition accuracy (mAP50) of 85.6%, an increase of 4.5%. Additionally, the mAP50-95 is 63.6%, an improvement of 6.9%, indicating good detection accuracy. The performance of the SIMCB-Yolo model on the self-built forest fire smoke dataset is also significantly better than that of current mainstream models, demonstrating high practical value.

Selection of hand features based on Random Forest algorithm and hand shape recognition

To obtain effective features applicable to hand morphology recognition, the method of obtaining effective feature indicators based on the Random Forest (RF) algorithm to downscale hand morphology parameters is proposed. Firstly, 232 female university students collected three-dimensional hand information, constructed auxiliary point, line, and surface standardised measurement methods, obtained 33 characteristic parts of human dimensions, and used k-means clustering for hand morphology subdivision. Hand morphology can be divided into three categories: short, slender and broad. The RF algorithm is used for feature index importance assessment and hand shape recognition model. The accuracy of the feature metrics determined by the RF algorithm, PCA, and VC method applied to hand shape recognition is compared and analysed to verify the effectiveness of the dimensionality reduction of the RF algorithm. The results showed that the feature indexes used for hand shape recognition were five items: hand length, thickness at the metacarpal, thenar width, the distance between the thumb and index finger, and distance from the root of the little finger to the centre of the wrist. Using the RF algorithm to reduce the dimensionality is more effective; the average recognition accuracy of the four hand shape recognition models is 93.78% on average, compared with PCA and VC reduction methods, the average accuracy of hand shape recognition models is increased by 19.17%, and 14.86% respectively. The study's results can provide methodological references for the objective selection of characteristic indicators and morphological recognition of human body parts.

Plant leaf disease recognition based on improved SinGAN and improved ResNet34.

The identification of plant leaf diseases is crucial in precision agriculture, playing a pivotal role in advancing the modernization of agriculture. Timely detection and diagnosis of leaf diseases for preventive measures significantly contribute to enhancing both the quantity and quality of agricultural products, thereby fostering the in-depth development of precision agriculture. However, despite the rapid development of research on plant leaf disease identification, it still faces challenges such as insufficient agricultural datasets and the problem of deep learning-based disease identification models having numerous training parameters and insufficient accuracy. This paper proposes a plant leaf disease identification method based on improved SinGAN and improved ResNet34 to address the aforementioned issues. Firstly, an improved SinGAN called Reconstruction-Based Single Image Generation Network (ReSinGN) is proposed for image enhancement. This network accelerates model training speed by using an autoencoder to replace the GAN in the SinGAN and incorporates a Convolutional Block Attention Module (CBAM) into the autoencoder to more accurately capture important features and structural information in the images. Random pixel Shuffling are introduced in ReSinGN to enable the model to learn richer data representations, further enhancing the quality of generated images. Secondly, an improved ResNet34 is proposed for plant leaf disease identification. This involves adding CBAM modules to the ResNet34 to alleviate the limitations of parameter sharing, replacing the ReLU activation function with LeakyReLU activation function to address the problem of neuron death, and utilizing transfer learning-based training methods to accelerate network training speed. This paper takes tomato leaf diseases as the experimental subject, and the experimental results demonstrate that: (1) ReSinGN generates high-quality images at least 44.6 times faster in training speed compared to SinGAN. (2) The Tenengrad score of images generated by the ReSinGN model is 67.3, which is improved by 30.2 compared to the SinGAN, resulting in clearer images. (3) ReSinGN model with random pixel Shuffling outperforms SinGAN in both image clarity and distortion, achieving the optimal balance between image clarity and distortion. (4) The improved ResNet34 achieved an average recognition accuracy, recognition precision, recognition accuracy (redundant as it's similar to precision), recall, and F1 score of 98.57, 96.57, 98.68, 97.7, and 98.17%, respectively, for tomato leaf disease identification. Compared to the original ResNet34, this represents enhancements of 3.65, 4.66, 0.88, 4.1, and 2.47%, respectively.

Construction of English corpus oral instant translation model based on internet of things and deep learning of information security

In order to improve the security and performance of the oral English instant translation model, this paper optimizes the instant translation model through the Internet of Things (IoT) security technology and deep learning technology. In this paper, the real-time translation model based on deep learning and IoT technology is analyzed in detail to show the application of these two technologies in the real-time translation model, and the related information security issues are discussed. Meanwhile, this paper proposes a method combining deep learning network and IoT technology to further improve the security of instant translation model. The experimental results show that under the optimized model, the parameter upload time is 60 seconds, the aggregation calculation time is 6.5 seconds, and the authentication time is 7.5 seconds. Moreover, the average recognition accuracy of the optimized model reaches 93.1%, and it is superior to the traditional machine translation method in accuracy and real-time, which has wide practical value and application prospects. Therefore, the research has certain reference significance for improving the security of the English corpus oral instant translation model.

CoreViT: A new vision transformer model for lithofacies identification in cores

Lithofacies identification is crucial in fields such as geology, geotechnical engineering, rock mechanics, and petroleum engineering, as it reveals the physical, chemical, and mineralogical characteristics of rocks that aid in identifying oil and gas reservoir types, volumes, and productivity. Traditional methods of identifying rock types, such as manual sample identification, micrograph identification, and experimental identification, are costly and time-consuming. Several studies have endeavored to enhance the degree of lithofacies identification automation using machine learning (ML) and deep learning (DL) techniques. However, their capacity to recognize intricate data remains restricted. To overcome this challenge, we introduce CoreViT, a modified architecture of the advanced vision transformer (ViT) that includes the parallel transformer encoder (PTE) for exchanging information among image patches, and the class encoder (CE) for improved automatic lithofacies identification. Our study, focusing on core images including algal limestone, mudstone, and non-algal limestone from the Fengxi Well 1 in the Qaidam Basin, demonstrates CoreViT's average recognition accuracy at approximately 97.5% with an average loss of 0.071. Compared with other classical convolutional neural network (CNN) models, CoreViT achieves relatively high accuracy and low loss. This study highlights the superiority of the ViT model over traditional deep convolutional neural networks (DCNNs) and suggests the great potentiality of applying the ViT model in lithofacies identification in core samples.

A Case Study on (EAVIBA) Enhanced Accessibility for Visually Impaired & Blind Artists using Machine Learning

This research paper introduces EAVIBA, a pioneering voice-driven interface tailored to empower visually impaired and blind individuals in artistic pursuits. Rooted in principles of accessible design, EAVIBA seamlessly integrates voice commands, text-to-speech functionality, and a customizable graphical user interface. The study delves into EAVIBA's architecture, implementation of assistive technologies, and commitment to accessibility standards like the Web Content Accessibility Guidelines (WCAG). Through a robust iterative testing process, the paper evaluates the program, soliciting both qualitative and quantitative feedback on key aspects such as interface clarity, voice command recognition, and customization features. Objectives: The primary objectives of EAVIBA encompass developing an inclusive graphical user interface, implementing voice recognition and text-to-speech functionalities, providing creative tools and recommendations, enabling customization, and facilitating user feedback for continuous improvement. Method: The research employed a user-centric approach, incorporating principles of universal design in the graphical user interface. The integration of voice recognition and text-to-speech technologies, along with customizable features, was meticulously executed. Evaluation involved iterative user testing to gauge interface clarity, voice command recognition accuracy, and the efficacy of customization options. Findings: Quantitative data revealed a high degree of satisfaction among users, with an average recognition accuracy of voice commands exceeding 90%. Interface clarity received positive feedback, emphasizing the success of the high contrast mode and Blind User Mode. Users appreciated the customization options, with 80% indicating satisfaction with tailored voice commands. Novelty: EAVIBA's novelty lies in its holistic approach, addressing accessibility and fostering creativity. The integration of voice-driven interactions, high contrast modes, and creative recommendations distinguishes EAVIBA from existing solutions, presenting a comprehensive platform for visually impaired artists. Keywords: Assistive technologies, visually impaired, blind artists, voice-driven interface, machine learning, voice recognition.

Embedded monitoring system and teaching of artificial intelligence online drug component recognition.

Drug testing has many test elements. It aims to prevent unqualified drugs from entering the market and ensure drug safety. The existing artificial intelligence (AI) online monitoring system identifies active ingredients in the process of use. Owing to their openness, data are easy to be lost, failing to meet user needs and inducing a specific impact on the use of the monitoring system. With the continuous development of computer and measurement technologies, various biochemical data are increasing at an unprecedented speed, and numerous databases are emerging. Extracting patterns from considerable known data and experimental facts is an essential task for a wide range of biological and chemical workers. Pattern recognition is one of the essential technologies for data mining. It is widely used in industry, agriculture, national defense, biomedicine, meteorology, astronomy, and other fields. To improve the effect of the online drug ingredient recognition system, this study used AI to design an online drug ingredient recognition-embedded monitoring system and applied AI to the teaching field to improve teaching efficiency. First, this study constructed the framework of the AI online drug ingredient recognition-embedded monitoring system and introduced the process of online drug ingredient recognition. Then, it introduced the pattern recognition method, constructed the pattern recognition system, and presented the pattern recognition algorithm and the algorithm evaluation index. Afterward, it used pattern recognition to conduct a qualitative analysis of the infrared spectrum of drug components and introduced the overall process of the qualitative analysis. In addition, this study employed AI to implement changes to the embedded system instruction in colleges and universities, summarizing the current issues. The impact of drug component recognition and the educational impact of embedded systems were investigated in the experimental portion. The experimental findings demonstrated the excellent accuracy, sensitivity, specificity, and Matthew correlation coefficient of the online drug component recognition-integrated monitoring system in this work. Compared with that of other systems, its average drug component recognition accuracy was above 0.85. Students in five majors reported high levels of satisfaction with the embedded system teaching, which is better for delivering college instruction.

Application of emotion recognition technology in psychological counseling for college students

Abstract In traditional psychological counseling, limitations of patients’ subjective feelings and expressive abilities, as well as the varying levels of professional skills and conversation skills of counselors, make it difficult to properly solve psychological problems. Emotional recognition technology can help psychological counselors better understand the inner world of patients, thereby providing more targeted psychological counseling services for college students. 2D convolutional neural network-long short-term memory (2DCNN-LSTM) and 1D convolutional neural network-long short-term memory network (1DCNN-LSTM) were used for emotion recognition. Wavelet packet coefficient sequences were utilized to analyze multimodal data; multi-core convolutional neural network models were applied for multimodal feature layer fusion, and LSTM recurrent neural networks were used to effectively identify valence and college student emotions. The experimental results showed that under multimodal recognition, the average recognition accuracy of 2DCNN-LSTM for fourth-year university students was 18.7% higher than that of 1DCNN-LSTM for fourth-year university students. Compared with 1DCNN-LSTM, 2DCNN-LSTM can achieve better recognition accuracy. The research results can help psychological counselors better understand the emotional state and needs of students and provide more personalized and targeted counseling services.

Automatic Modulation Recognition Method Based on Phase Transformation and Deep Residual Shrinkage Network

Automatic Modulation Recognition (AMR) is currently a research hotspot, and research under low Signal-to-Noise Ratio (SNR) conditions still poses certain challenges. This paper proposes an AMR method based on phase transformation and deep residual shrinkage network to improve recognition accuracy. Firstly, the raw I/Q data from the benchmark dataset RML2016.10a are used as the input. Then, an end-to-end modulation recognition is performed using the model. Phase transformation is used to correct the raw I/Q data and reduce the interference of phase shift on modulation recognition. Convolutional neural network (CNN) and Gate Recurrent Unit (GRU) extract the spatial and temporal features of the modulation signal, respectively. The improved deep residual shrinkage network is added after CNN to eliminate unimportant features through soft thresholding. Finally, the proposed model is trained and tested. The experimental results show that the proposed model notably reduces the number of parameters compared to other models, effectively improving the recognition accuracy under low SNR conditions. The average recognition accuracy reaches 62.46%, and the highest recognition accuracy reaches 92.41%.

A fiber recognition framework based on multi-head attention mechanism

Convolutional neural networks have been extensively studied in textile fiber recognition. However, the down-sampling performed by the convolutional and pooling layers on the extracted features result in a significant loss of fine-grained fiber features. To address this issue, a multi-head attention framework for fiber recognition has been proposed. First, textile surface images are captured using optical magnifiers and smart devices such as smartphones. Second, fiber features and label embeddings are learned separately by multi-head self-attention. Finally, the label embeddings are used to query the presence of fiber types, and pool type-related features in the multi-head cross-attention module to classify fibers. The experimental results demonstrate that the proposed method performs exceptionally well on the textile surface image dataset, with a mean average precision accuracy improvement of 6% compared with the convolutional neural network-based fiber recognition method Cu-Net, and a 5% improvement compared with the fiber recognition method FiberCT combining convolutional neural networks and attention mechanisms. It is worth mentioning that the fiber images captured at 200× magnification in the collected dataset are most favorable for models to recognize fiber. The proposed method achieves a mean average precision fiber recognition accuracy of 80.2% on images at 200× magnification.

Shallow defects identification for urban roads using interpretable dynamic broad network

Deep learning is extensively utilised in transport geotechnical engineering. However, deep architectures have large computational costs and update times, while failing to understand decisions. To this regard, we propose an interpretable dynamic broad network combined with ground-penetrating radar for internal defect identification in roadbeds. The method is more suitable for feature characterisation of two-dimensional data and satisfies incremental updates. The test results indicated that the proposed method has an average recognition accuracy of 0.9124 for the four types of internal defects in roadbeds. Compared to the other four classical machine learning methods, it balances training efficiency and recognition accuracy. Robustness analysis results demonstrated that the method is noise-resistant. However, comprehending the recognition results of intelligent algorithms is a key topic. Local interpretation approach is introduced to quantify the feature importance that affects the decision of the model. Based on the feature importance calculation, it is possible to distinguish between positive and negative regions in one sample that influence the decision of the detection model. These interpretative analyses can assist us in better understanding the reasons for decisions generated by the detection model that provide technical support for subsequent enhancements.

A novel relation‐aware global attention network for sentiment analysis of students in the wisdom classroom

SummaryThe ability of facial expression recognition, the ability to decipher human emotions from facial features, and we can judge the emotional state of human beings by analysing facial expressions. Empowered by this technology, the related research of student sentiment analysis has become a focal point in the realm of educational technology. Teachers can now read the emotional state of their students and estimate the effectiveness of their teaching strategies, enabling them to implement appropriate intervention techniques to enhance teaching outcomes. However, facial expression recognition techniques are currently limited by their shortcomings. Network performance degradation and loss of feature information are key issues that hinder the effectiveness of sentiment analysis of student feedback. To overcome these limitations, in this paper, we propose RFMNet, a novel network model based on deep learning theory. RFMNet, with its higher extraction capability, is designed to accurately analyse student expressions. It introduces a relation‐aware global attention (RGA) module, which facilitates the integration of more discriminative expression features in the image, leading to more refined sentiment analysis. This innovative model also employs a Mish activation function and a focal loss function. The Mish activation function is of particular interest because it helps to avoid the loss of feature information due to neuron deactivation when the ReLU gradient is . This approach results in a more robust and accurate facial expression recognition model. Our model was tested on the FERPlus public dataset. The model achieved an average recognition accuracy of 89.62%, which is a testament to its performance in decoding the facial expressions of students. This precision allows for intelligent processing of teaching information and real‐time feedback, enabling teachers to adapt their teaching strategies to better suit the needs of their students.

LPI radar waveform recognition based on semi-supervised model all mean teacher

Low probability of intercept (LPI) radar signal identification plays an important role in electronic warfare, but most existing algorithms are proposed under the condition of sufficient samples, ignoring the problem of a small amount of labeled data in the actual electromagnetic environment. To solve the problem, in this paper, a semi-supervised learning model All Mean Teacher (AMT) based on Mean Teacher (MT) is proposed. First, the LPI radar signal is transformed into Time-frequency images (TFIs) by using the Choi-Williams distribution, and Random Erasing is used for TFIs which improves the generalization ability of the model. Then the Multi-headed Self-Attention Network (MSA-Net) is aimed to extract features, combined with AMT to realize the automatic waveform recognition of radar signals. MSA-Net facilitates feature information propagation by computing contrast costs on TFIs between the student and teacher networks. It solves the problem that TFIs are not easy to train for small amounts of labeled data, improving the accuracy of signal recognition in semi-supervised learning scenarios. Experimental results show that the average recognition accuracy of the proposed method is up to 85.7% at a signal-to-noise ratio of -8 dB.

Multimodal deep learning-based drought monitoring research for winter wheat during critical growth stages

Wheat is a major grain crop in China, accounting for one-fifth of the national grain production. Drought stress severely affects the normal growth and development of wheat, leading to total crop failure, reduced yields, and quality. To address the lag and limitations inherent in traditional drought monitoring methods, this paper proposes a multimodal deep learning-based drought stress monitoring S-DNet model for winter wheat during its critical growth periods. Drought stress images of winter wheat during the Rise-Jointing, Heading-Flowering and Flowering-Maturity stages were acquired to establish a dataset corresponding to soil moisture monitoring data. The DenseNet-121 model was selected as the base network to extract drought features. Combining the drought phenotypic characteristics of wheat in the field with meteorological factors and IoT technology, the study integrated the meteorological drought index SPEI, based on WSN sensors, and deep image learning data to build a multimodal deep learning-based S-DNet model for monitoring drought stress in winter wheat. The results show that, compared to the single-modal DenseNet-121 model, the multimodal S-DNet model has higher robustness and generalization capability, with an average drought recognition accuracy reaching 96.4%. This effectively achieves non-destructive, accurate, and rapid monitoring of drought stress in winter wheat.

Design of a flexible tactile sensor for material and texture identification utilizing both contact-separation and surface sliding modes for real-life touch simulation

Haptic feedback plays a crucial role, continuously offering individuals vital information about their external surroundings through physical contact. Therefore, rapid advances in human-friendly biomimetic electronics and flexible devices, have allowed the development of advance electronic skins for artificial haptic sensation. In particular, triboelectric nanogenerator-based tactile sensors have garnered significant interest, pursuing enhanced performance in material and texture recognition, coupled with simplicity in fabrication, cost-effectiveness, high resolution and sensitivity. Thus, in this work, we propose a flexible triboelectric tactile sensor, utilizing the freestanding single-electrode triboelectric nanogenerator principle, realizing concurrent material and texture detection in a single device. We leverage a novel hybrid operational regime for tactile sensing, combining the advantages of both contact-separation and surface sliding modes in a single contact-sliding-separation motion. This enables a realistic representation of the surface employing an artificial finger, by gathering material information from the contact-separation events and texture details from the sliding motion. Common materials such as wood, paper, copper and glass were used for the identification study along with textured surfaces made of the same materials with microscopic surface features. Simultaneous classification and recognition of materials and textures was performed by a single, 1D convolutional artificial neural network, achieving an average recognition accuracy of 98.4% across all different surfaces.

A two-stage feature selection method for hob state recognition

In mechanical field, tool state seriously affects the production quality and efficiency. Recently, machine learning is an important technique to realize tool state recognition. Extracting and selecting features from the raw input data of the machine learning model can effectively eliminate useless information and improve the model accuracy. Therefore, a two-stage feature selection method is proposed in this paper. We first qualitatively propose the non-discreteness and separability of signal features. For these two properties, quantitative calculation methods are then developed in two stages. The proposed two-stage feature selection method is applied to machine learning models for tool state recognition. In engineering application verification, the proposed method achieves an average recognition accuracy of 84.4% in hob wear recognition, and an average recognition accuracy of 99.8% and 94.4% for two data sets in hob fault recognition, which is significantly higher than other feature selection methods. The verification results indicate that the proposed method has good generalization and engineering application value.

How to use one surface electromyography sensor to recognize six hand movements for a mechanical hand in real time: a method based on Morse code.

Surface electromyography (sEMG) signal is a kind of physiological signal reflecting muscle activity and muscle force. At present, the existing methods of recognizing human motion intention need more than two sensors to recognize more than two kinds of movements, the sensor pasting positions are special, and the hardware conditions for execution are high. In this work, a real-time motion intention recognition method based on Morse code is proposed and applied to the mechanical hand. The short-time and long-term muscle contraction signals collected by a single sEMG sensor were extracted and encoded with the Morse code method, and then the developed mapping method from Morse code to six hand movements were used to recognize hand movements. The average recognition accuracy of hand movements was 94.8704 ± 2.3556%, the average adjusting time was 34.89s for all subjects, and the execution time of a single movement was 381ms. The corresponding experiment video can be found in the attachment to show the experiment. The method proposed in this work is a method with one sensor to recognize six movements, low hardware conditions, high recognition accuracy, and insensitive to the difference of sensor pasting position.

REAL TIME FACE ATTENDANCE SYSTEM USING DEEP LEARNING

One's face is a physical depiction of who they are. As a result, we have suggested a facial recognition- based automatic student attendance system. The usage of face recognition technology in everyday life is particularly beneficial for security control systems. Face recognition is used by the FBI (Federal Bureau of Investigation) for criminal investigations as well as the airport security system to identify criminals. In the first step of our suggested method, video framing, the camera is activated using a user-friendly interface. The Viola-Jones method is used to identify and segment the face ROI from the video frame. If necessary, image size scaling is done at the pre-processing step to prevent information loss. After using the median filter to reduce noise, color pictures are transformed into grayscale versions. Then, to improve the contrast of the pictures, contrast-limited adaptive histogram equalization (CLAHE) is applied. Principal component analysis (PCA) and improved local binary pattern (LBP) are used appropriately in the face identification stage to extract the characteristics from facial pictures. By minimizing the lighting impact and boosting identification rates, the improved local binary pattern outperforms the original local binary pattern in our suggested method. The features that were retrieved from the training pictures and the features that were extracted from the test images are then compared. The face pictures are then identified and categorized using the upgraded LBP, PCA, and algorithm that produced the best results. The acknowledged student's attendance will then be recorded and saved in the excel file. A warning will be sent if a student signs in more than once. The student who is not enrolled will also be allowed to register right away. When two photographs per person are trained, the average recognition accuracy is 100% for high-quality images, 94.12% for low-quality images, and 95.76% for the face database.

Multi-type classification and 3D visualization of internal defects in concrete plates based on a hybrid method of deep learning algorithm and bending mode analysis for acoustic vibration signals

In this paper, a hybrid method combining the deep learning algorithm and bending mode analysis of acoustic vibration signals is proposed for the multi-type classification and 3D visualization of internal defects in concrete plates. A novel deep learning model termed fully convolutional network based on principal component analysis and attention-embedded long-short term memory (PCA-ALSTM-FCN) is established. The PCA-ALSTM-FCN model successfully classifies multi-type internal defects, and the two-dimensional (2D) defect contour maps are generated based on the predicted state by the trained model. The defect depths are calculated according to the analytical formula of bending vibration mode of acoustic vibration signals. Combining the 2D defect map and depth information, the three-dimensional (3D) visualization images of defects are obtained. The average recognition accuracy of PCA-ALSTM-FCN model for different types of defects reaches 94.8 %, and the defect depth calculation error range is approximately 10 %–20 %. The experimental results show that the proposed method in this paper can accurately distinguish different types of internal defects and effectively locate the 3D position of defects in concrete.