Network ResNet Research Articles

Employing Transfer Learning in Land-use Land-cover for Risk Management

Abstract. In disaster management, land-use land-cover (LULC) maps are vital for real-time situational awareness and coordinated responses. These maps aid in coordinating field operations, guiding rescue teams, and identifying high vulnerability areas. They ensure accurate spatial information sharing and management during disaster response efforts. Deep transfer learning models have emerged as powerful tools for LULC classification, addressing challenges like insufficient training data and complex classification tasks. In this research instead of building networks from scratch, pre-trained networks that used EuroSAT benchmark dataset are employed. Several deep learning models including 1-layer CNN, 4-layers CNN, VGG16 and an improved ResNet-50 network as proposed method are considered and compared in this study. The results were analyzed in both quantitative and qualitative ways. In the quantitative mode, the measurement criteria such as Overall Accuracy (OA), F1Score, Precision and Recall were calculated, and in the qualitative mode, the class diagram was drawn in the feature space to check the separability of the classes. Finally, the results show high overall accuracy score of 95.9% indicating the high potential of our proposed network for ResNet-50. The proposed method has resolved insufficient training dataset by implementing data augmentation that it can be solved the problem of lack dataset.

Green apple detection method based on multi-dimensional feature extraction network model and Transformer module

To enhance fast and accurate detection of pollution-free green apples for food safety, this paper uses the DETR network as a framework to propose a new method for pollution-free green apple detection based on a multi-dimensional feature extraction network and Transformer module. Firstly, an improved DETR network main feature extraction module adopts the ResNet18 network and replaces some residual layers with deformable convolutions (DCNv2), enabling the model to better adapt to pollution-free fruit changes at different scales and angles, while eliminating the impact of microbial contamination on fruit testing; Subsequently, the extended spatial pyramid pooling model (DSPP) and multi-scale residual aggregation module (FRAM) are integrated, which help reduce feature noise and minimize the loss of underlying features during the feature extraction process. The fusion of the two modules enhances the model’s ability to detect objects of different scales, thereby improving the accuracy of near-color fruit detection; At the same time, in order to solve the problems of slow convergence speed and large calculation amount of the basic network model, the convergence speed of the overall network model is improved by replacing the attention mechanism of Transformer. Experimental results show that compared with the original DETR model, the proposed algorithm has improved in AP, AP50 and AP75 indicators, especially in the AP50 indicator, which has the most obvious improvement reaching a detection accuracy of 97.12%. In the meantime, the trained network model is deployed on the picking robot. Compared with the original DETR network model, its average detection accuracy is as high as 96.58%, and the detection speed is increased by about 51%. Mixed sample detection tests were carried out before and after the model deployment, and the detection rate of the proposed method for non-polluted fruits reached more than 0.95. enabling the picking robot to efficiently complete the task of picking green apples. The test results show that the algorithm proposed in this article exhibits great potential in the task of detecting pollution-free near-color fruits by the picking robot. It ensures pollution-free fruit picking and the application of AI in food safety.

Deep learning applications for real-time and early detection of fall armyworm, African armyworm, and maize stem borer

The application of artificial intelligence for identifying Fall armyworm (Spodoptera frugiperda), African armyworm (Spodoptera exempta), and Maize stem borer (Busseola fusca) is critical due to the threats they pose to global food production. This study aims to evaluate and identify the most accurate and robust DL models in detecting and classifying these three significant agricultural pests. Seven traditional DL models: Convolutional Neural Network, Visual Geometry Group (VGG16), Residual Networks (ResNet50), MobileNetV2, InceptionV3, Deep Neural Network (DNN), and InceptionResNetV2 and the advanced You Look Only Once (YOLOv8) model were trained and tested using pest image datasets. The results showed that all traditional models except DNN had high accuracies ranging from 93.17% (InceptionResNetV2) to 99.43% (MobileNet) in training and testing, with losses ranging from 1.71% (MobileNetV2) to 24.99% (InceptionResNetV2). DNN had a slightly lower accuracy range of 55.27% to 56.39% and a loss range of 85.02% to 89.96% in training and testing. YOLOv8 emerged as the best and most robust model in the pest detection and classification tasks, achieving Precision and Recall scores ranging from 98.4% to 100% on single-class and multi-class classifications, making it highly suitable for real-world pest management applications. This research pioneers the use of DL for the classification and detection of maize stem borer, African armyworm and Fall armyworm, unique and separately addressing a critical gap in agricultural pest management in corn. With early and accurate pest identification, crop protection measures can be implemented efficiently. The findings lead to reduced crop damage and enhanced food security.

A method for recognizing abnormal behaviors of personnel at petroleum stations based on GTB-ResNet

PurposeThe abnormal behaviors of staff at petroleum stations pose significant safety hazards. Addressing the challenges of high parameter counts, lengthy training periods and low recognition rates in existing 3D ResNet behavior recognition models, this paper proposes GTB-ResNet, a network designed to detect abnormal behaviors in petroleum station staff.Design/methodology/approachFirstly, to mitigate the issues of excessive parameters and computational complexity in 3D ResNet, a lightweight residual convolution module called the Ghost residual module (GhostNet) is introduced in the feature extraction network. Ghost convolution replaces standard convolution, reducing model parameters while preserving multi-scale feature extraction capabilities. Secondly, to enhance the model's focus on salient features amidst wide surveillance ranges and small target objects, the triplet attention mechanism module is integrated to facilitate spatial and channel information interaction. Lastly, to address the challenge of short time-series features leading to misjudgments in similar actions, a bidirectional gated recurrent network is added to the feature extraction backbone network. This ensures the extraction of key long time-series features, thereby improving feature extraction accuracy.FindingsThe experimental setup encompasses four behavior types: illegal phone answering, smoking, falling (abnormal) and touching the face (normal), comprising a total of 892 videos. Experimental results showcase GTB-ResNet achieving a recognition accuracy of 96.7% with a model parameter count of 4.46 M and a computational complexity of 3.898 G. This represents a 4.4% improvement over 3D ResNet, with reductions of 90.4% in parameters and 61.5% in computational complexity.Originality/valueSpecifically designed for edge devices in oil stations, the 3D ResNet network is tailored for real-time action prediction. To address the challenges posed by the large number of parameters in 3D ResNet networks and the difficulties in deployment on edge devices, a lightweight residual module based on ghost convolution is developed. Additionally, to tackle the issue of low detection accuracy of behaviors amidst the noisy environment of petroleum stations, a triple attention mechanism is introduced during feature extraction to enhance focus on salient features. Moreover, to overcome the potential for misjudgments arising from the similarity of actions, a Bi-GRU model is introduced to enhance the extraction of key long-term features.

Research on tool wear and breakage state recognition of heavy milling 508III steel based on ResNet-CBAM

Milling water chamber head material 508 III steel belongs to extreme manufacturing, and the tool failure is very serious in the process of machining. How to monitor and identify the tool wear damage state in time and effectively is a key problem to be solved urgently. Identifying chip morphology changes under machining conditions plays a very important role in characterizing tool wear and breakage. Tool wear has an important influence on the chip morphology during the process of milling 508III steel, and changes in chip morphology can also reflect the tool wear and breakage state. Therefore, this paper takes the chip morphology image as an important feature for tool wear and breakage recognition, and uses Gaussian fuzzy estimation morphology method to preprocess the chip morphology image. Build a ResNet network combined by convolutional block attention module (ResNet-CBAM) tool wear and breakage recognition model, and explore the selection of the model backbone network, determination of the ResNet backbone network depth, and fusion methods of different attention mechanisms. Through the ablation experiment and visual analysis of the attention mechanism module, it is verified that the ResNet-CBAM model proposed in this paper has an accuracy rate of 96.67% in tool wear and breakage state recognition, especially in the stage of severe tool wear and breakage. This study realizes the prediction and early warning of tool life, and provides effective guarantee for the efficient cutting of large parts of high-end equipment and the stable operation of manufacturing system.

Deep Convolutional Neural Network for Accurate Classification of Myofibroblastic Lesions on Patch-Based Images.

This study aimed to implement and evaluate a Deep Convolutional Neural Network for classifying myofibroblastic lesions into benign and malignant categories based on patch-based images. A Residual Neural Network (ResNet50) model, pre-trained with weights from ImageNet, was fine-tuned to classify a cohort of 20 patients (11 benign and 9 malignant cases). Following annotation of tumor regions, the whole-slide images (WSIs) were fragmented into smaller patches (224 × 224 pixels). These patches were non-randomly divided into training (308,843 patches), validation (43,268 patches), and test (42,061 patches) subsets, maintaining a 78:11:11 ratio. The CNN training was caried out for 75 epochs utilizing a batch size of 4, the Adam optimizer, and a learning rate of 0.00001. ResNet50 achieved an accuracy of 98.97%, precision of 99.91%, sensitivity of 97.98%, specificity of 99.91%, F1 score of 98.94%, and AUC of 0.99. The ResNet50 model developed exhibited high accuracy during training and robust generalization capabilities in unseen data, indicating nearly flawless performance in distinguishing between benign and malignant myofibroblastic tumors, despite the small sample size. The excellent performance of the AI model in separating such histologically similar classes could be attributed to its ability to identify hidden discriminative features, as well as to use a wide range of features and benefit from proper data preprocessing.

Rhythm classifier: artificial intelligence for rhythm classification in insertable cardiac monitor

Abstract Background Though detections improved significantly, separating atrial tachycardia/atrial flutter (AT/AFL) from atrial fibrillations (AF) in insertable cardiac monitors (ICM) remains a challenge. Objective Using artificial intelligence (AI) algorithms to classify AF and AT/AFL in ICM detected episodes, which may reduce episode review burden. Methods Both raw EGM signal and the QRS diminished version of the raw signal were used to extract 14 AF and AT/AFL features from each episode. Extracted features were then represented as an ensemble of features and used to train and test the AI model. Modified ResNet18 network was used as the AI model. All episodes from a random selection of 80% of patients were used as training dataset, another 10% were used as validation, and 10% were used as test dataset, to evaluate model performance. AI model probability threshold was chosen based on validation set results. Episode detection sensitivity, specificity, and area under the curve (AUC) of receiver operating characteristics (ROC) curve in the independent test dataset are reported. Results AF and AT/AFL episodes detected in 918 LINQ ICM patients (5,253 AF and 4,705 AT/AFL – 745 pts and 8,950 false detection episodes) were included. Training dataset consisted of 15,183 episodes (735 pts), validation set consisted of 1,901 episodes (91 pts), and test set consisted of 1,824 episodes (92 pts). At chosen threshold in validation set, AF and AT/AFL detection had sensitivity of 90% and 73.3%, respectively, and specificity of 93.2% and 95%, respectively, with AUC of 0.966. In test data, AF and AT/AFL obtained relative sensitivity of 84.7% and 85.7%, respectively, with specificity of 88% and 86.5%, respectively (Fig. A). With ROC-AUC of 0.956 (Fig. B), the relative diagnostic yield of AF and AT/AFL on test data was 95.6%. Conclusion The trained AI model provides a relative sensitivity of about 85% for AF vs. AT/AFL classification and reduces inappropriate ICM detections by 79%. Adding more episodes may improve the performance.

Occluded facial expression recognition based on deep learning

Abstract. When employing Convolutional Neural Networks (CNNs) for facial expression recognition, several challenges are often encountered, such as facial occlusions, the limited size of reliable expression datasets, and inadequate precision in recognition outcomes. This paper preprocesses the dataset to enhance its reliability. By leveraging data synthesis and augmentation techniques, it employs a method of randomly generating occlusion blocks to integrate and expand the dataset. Based on the ResNet-18 network, the model is optimized by incorporating an attention mechanism, thereby improving the network's precision and robustness in recognizing facial expressions.

Fire Recognition Method Based on PSO-BP Neural Network and ResNet50

The rapid development of modern society and continuous urbanization have resulted in a proliferation of functional buildings, which offer significant convenience to individuals, but pose significant fire hazards as well. How to detect the fire at the early stage is always the focus of research. This paper proposes a multi-information source fusion fire recognition method based on particle swarm optimization (PSO)-backpropagation (BP) neural networks and ResNet50. The PSO algorithm is applied to optimize the initial parameters of a BP neural network model, while data from three sensors — temperature, humidity and smoke — are integrated, through iterative training of the system, accurate recognition of sensor data can be achieved. Additionally, a method is proposed for the recognition of infrared fire images using ResNet50 and transfer learning. By improving the ResNet50 network model and migrating the ResNet50 pre-trained network weight, infrared fire image recognition accuracy is further enhanced. Then the sensor information recognition results and image information recognition results are input into the fuzzy system for fusion reasoning again, and the final decision is output according to the set fuzzy rules. Experimental findings demonstrate that the multi-information source fusion approach utilizing the PSO-BP neural network and ResNet50 significantly enhances the accuracy and response time of fire recognition, and achieves a remarkable recognition effect.

MaskAppendix: Backbone-Enriched Mask R-CNN Based on Grad-CAM for Automatic Appendix Segmentation

Background: A leading cause of emergency abdominal surgery, appendicitis is a common condition affecting millions of people worldwide. Automatic and accurate segmentation of the appendix from medical imaging is a challenging task, due to its small size, variability in shape, and proximity to other anatomical structures. Methods: In this study, we propose a backbone-enriched Mask R-CNN architecture (MaskAppendix) on the Detectron platform, enhanced with Gradient-weighted Class Activation Mapping (Grad-CAM), for precise appendix segmentation on computed tomography (CT) scans. In the proposed MaskAppendix deep learning model, ResNet101 network is used as the backbone. By integrating Grad-CAM into the MaskAppendix network, our model improves feature localization, allowing it to better capture subtle variations in appendix morphology. Results: We conduct extensive experiments on a dataset of abdominal CT scans, demonstrating that our method achieves state-of-the-art performance in appendix segmentation, outperforming traditional segmentation techniques in terms of both accuracy and robustness. In the automatic segmentation of the appendix region in CT slices, a DSC score of 87.17% was achieved with the proposed approach, and the results obtained have the potential to improve clinical diagnostic accuracy. Conclusions: This framework provides an effective tool for aiding clinicians in the diagnosis of appendicitis and other related conditions, reducing the potential for diagnostic errors and enhancing clinical workflow efficiency.

Study on the classification of benign and malignant breast lesions using a multi-sequence breast MRI fusion radiomics and deep learning model

PurposeTo develop a multi-modal model combining multi-sequence breast MRI fusion radiomics and deep learning for the classification of benign and malignant breast lesions, to assist clinicians in better selecting treatment plans. MethodsA total of 314 patients who underwent breast MRI examinations were included. They were randomly divided into training, validation, and test sets in a ratio of 7:1:2. Subsequently, features of T1-weighted images (T1WI), T2-weighted images (T2WI), and dynamic contrast-enhanced MRI (DCE-MRI) were extracted using the convolutional neural network ResNet50 for fusion, and then combined with radiomic features from the three sequences. The following models were established: T1 model, T2 model, DCE model, DCE_T1_T2 model, and DCE_T1_T2_rad model. The performance of the models was evaluated by the area under the receiver operating characteristic (ROC) curve (AUC), accuracy, sensitivity, specificity, positive predictive value, and negative predictive value. The differences between the DCE_T1_T2_rad model and the other four models were compared using the Delong test, with a P-value < 0.05 considered statistically significant. ResultsThe five models established in this study performed well, with AUC values of 0.53 for the T1 model, 0.62 for the T2 model, 0.79 for the DCE model, 0.94 for the DCE_T1_T2 model, and 0.98 for the DCE_T1_T2_rad model. The DCE_T1_T2_rad model showed statistically significant differences (P < 0.05) compared to the other four models. ConclusionThe use of a multi-modal model combining multi-sequence breast MRI fusion radiomics and deep learning can effectively improve the diagnostic performance of breast lesion classification.

Evaluating facial dermis aging in healthy Caucasian females with LC-OCT and deep learning

Recent advancements in high-resolution imaging have significantly improved our understanding of microstructural changes in the skin and their relationship to the aging process. Line Field Confocal Optical Coherence Tomography (LC-OCT) provides detailed 3D insights into various skin layers, including the papillary dermis and its fibrous network. In this study, a deep learning model utilizing a 3D ResNet-18 network was trained to predict chronological age from LC-OCT images of 100 healthy Caucasian female volunteers, aged 20 to 70 years. The AI-based protocol focused on regions of interest delineated between the segmented dermal-epidermal junction and the superficial dermis, exploiting complex patterns within the collagen network for age prediction. The model achieved a mean absolute error of 4.2 years and exhibited a Pearson correlation coefficient of 0.937 with actual ages. Furthermore, there was a notable correlation (r = 0.87) between quantified clinical scoring, encompassing parameters such as firmness, elasticity, density, and wrinkle appearance, and the ages predicted by deep learning model. This strong correlation underscores how integrating emerging imaging technologies with deep learning can accelerate aging research and deepen our understanding of how alterations in skin microstructure are related to visible signs of aging.

A self-iterative learning with dual attention mechanism fusion method for pumpkin seed quality equipment

To perform fast and accurate classification of pumpkin seeds, three batches of pumpkin seed images were taken. A self-iterative learning large kernel attention-Squeeze and Excitation-Residual Network34 (LKA-SE-ResNet34) model was proposed by adding dual attention mechanism fusion of large kernal attention (LKA) mechanism and Squeeze and Excitation (SE) attention mechanism to each residual module of ResNet34 for recognizing pumpkin seeds. ResNet34 network model was chosen as the basic framework by comparing GoogleNet, AlexNet, EfficientNet, MobileNet, DenseNet, ResNet34 and Residual Network50 (ResNet50) deep learning network models. The comparation result showed that the top two models with the highest accuracy (Acc) are ResNet50 network model and ResNet34 network model among the compared models. The experimental results showed that the self-iterative learning LKA-SE-ResNet34 model outperformed the above-mentioned other network models in pumpkin seed recognition with an accuracy of 91.8%. To further validate the effectiveness of the proposed model for small sample size of data, two-sided images and deformed images of pumpkin seeds were collected and used, and the experimental results still proved that the proposed model had good generalization ability and accuracy, with the accuracies of 93.63% and 94.15% for two sides of pumpkin seeds as well as an accuracy of 99.03% for deformed pumpkin seeds. A set of seed sorting machine and an application (APP) for non-destructive testing of seeds were initially developed.

Research on Intelligent Digital Manufacturing Platform Based on Artificial Intelligence Deep Neural Network

In today’s information age, intelligent manufacturing has become a significant trend of future industrial development. The neural network plays a vital role as the basis of smart manufacturing. This is a kind of intelligent algorithm that can simulate human brain organization, with strong learning ability and computing power, and has achieved remarkable results in the application field; the most typical representative is deep neural network, especially in recent years, an artificial intelligence developed based on deep neural network has entered the application stage, providing a new idea and model for industrial production. Based on this, taking the clothing production process as an example, this paper combined transfer learning and network fine-tuning technology to build a Faster RCNN model based on the ResNet-101 network to realize the positioning and recognition of clothing style map features and create an intelligent generation platform for clothing process flow on this basis.

Magnetic Prediction of Doped Two-Dimensional Nanomaterials Based on Swin–ResNet

Magnetism is an important property of doped two-dimensional nanostructures. By introducing dopant atoms or molecules, the electronic structure and magnetic behavior of the two-dimensional nanostructures can be altered. However, the complexity of the doping process requires different strategies for the preparation and testing of various types, layers, and scales of doped two-dimensional materials using traditional techniques. This process is resource-intensive, inefficient, and can pose safety risks when dealing with chemically unstable materials. Deep learning-based methods offer an effective solution to overcome these challenges and improve production efficiency. In this study, a deep learning-based method is proposed for predicting the magnetism of doped two-dimensional nanostructures. An image dataset was constructed for deep learning using a publicly available database of doped two-dimensional nanostructures. The ResNet model was enhanced by incorporating the Swin Transformer module, resulting in the Swin–ResNet network architecture. A comparative analysis was conducted with various deep learning models, including ResNet, Res2net, ResneXt, and Swin Transformer, to evaluate the performance of the optimized model in predicting the magnetism of doped two-dimensional nanostructures. The optimized model demonstrated significant improvements in magnetism prediction, with a best accuracy of 0.9.

An efficient CNN accelerator for pattern-compressed sparse neural networks on FPGA

Currently, the sparsity of weights and activations are mainly utilized to improve the energy efficiency and computational performance of CNN accelerators. However, the irregular sparsity of weights and activations can lead to problems such as mapping access conflict and imbalanced workload, which poses great challenges to the efficient computation of sparse networks. To address those problems, this paper proposes a CNN accelerator for pattern-compressed sparse neural networks on FPGA to achieve efficient inference acceleration. It mainly includes three parts: first, we propose a convolutional kernel grouping fusion method to achieve a balanced workload of weights. Second, we propose an activation hash mapping conflict reduction method to drastically reduce the mapping access conflict rate. Finally, we further propose a hierarchical mapping computing architecture to achieve efficient processing of mapping access conflicts and balanced adaptation of weight loads. To verify the effectiveness of the method proposed in this paper, our accelerator is designed and implemented on the Zynq UltraScale+ MPSoC ZCU102 evaluation board and evaluated by running VGG16 and ResNet50 networks. The experimental results show that the method proposed in this paper can reduce the mapping conflicts by more than 97 % and improve the utilization rate of processing element (PE) by 1.67×. Furthermore, the accelerator implemented in this paper can achieve 393.2 GOPs throughput, 1.40–3.25× computational performance, and 1.11–7.34× energy efficiency improvement.

RAF-Unet: A Remote Sensing Identification Method for Forest Land Information with Modified Unet

Abstract Carrying out remote sensing refinement identification of forest land in complex environment is of great significance for timely mapping of forest distribution. Aiming at the problem that remote sensing images have bias in the extraction of forest land information data, based on the semantic segmentation algorithm Unet, combining the ResNet50 deep learning network, the attention mechanism module and the feature pyramid structure, we construct RAF-Unet (ResNet+Attention+FPN+Unet) to improve the extraction of forest land information data. The ResNet50 classification network is used as the encoder of the Unet network to extract the feature maps at five different scales; then, the attention mechanism module is introduced in the decoder stage of the Unet network to extract the key task goal information by learning the weight values of the features; finally, the feature pyramid structure is used in the output stage of the encoder to fuse the information from the shallow network and the deep network to extract the remote sensing forest land information in the image. The results show that the RAF-Unet algorithm outperforms the Unet algorithm in all the indexes, with a precision of 95.24%, a recall of 91.80%, an F1-score value of 93.49%, an intersection over union of 87.63%, and an accuracy of 93.68%; the validity of the modules is verified by the ablation experiments, and the ResNet network, the attention mechanism, and the feature pyramid structure are all effective in improve the classification effect. It helps the forestry department to better manage and dynamically monitor forestry information, which is of great significance to the scientific development, utilization and protection of forest land resources.

Rigid-flexible coupling dynamic-assisted imbalanced fault diagnosis for helicopter tail transmission system

Helicopters operate in extreme environments, complicating the balanced and accurate acquisition of operational data. Predicting high-precision failures in helicopter tail-drive systems is particularly challenging due to severe data imbalances. To address this, a fault diagnosis method based on rigid-flexible coupling dynamics is proposed. A high-precision simulation model generates source domain data samples, validated by comparing the frequency domain components and RMS values of simulated and tested vibration responses under healthy conditions. An adaptive Local Maximum Mean Discrepancy (LMMD) mechanism aligns features between simulated and tested data. Additionally, the Coordinate-Separable Convolutional ResNet (CSC-ResNet) network is proposed to efficiently learn high-dimensional feature knowledge. Results show the proposed method outperforms others in fault diagnosis accuracy and smoothness, offering a new perspective for fault diagnosis of helicopter tail transmission systems.

Research on Art Classification Based on the Improved Resnet50 Network

In the field of computer vision, deep learning technology has made breakthroughs in image classification. ResNet50, an efficient deep residual network architecture, has demonstrated excellent performance in many image classification tasks. However, In the task of classifying artworks, ResNet50, still shows some limitations. It has difficulty identifying different artistic styles and is lacking in capturing the delicate details of the artwork. This article proposes an improved resnet50 model. By introducing SE blocks and extended convolution parameters, this article have made a deep revamp to ResNet50 to enhance its performance in art classification tasks. SE blocks increase the sensitivity of the network to differences in art styles by dynamically adjusting the dependencies between channels. Concurrently, the enhanced convolutional operation expands the model's sensory scope, so that the network can capture more delicate artistic details. The experimental results show that the improved model has achieved significant performance improvement on wikiart artwork classification datasets. In upcoming research, the paper plan to enhance art classification by broadening the dataset to encompass diverse art movements and forms, and by optimizing convolutional neural networks for large-scale data. This aims to improve the model's generalization and applicability across various visual arts, ensuring practical effectiveness and reliability.

Post-Stroke Dysarthria Voice Recognition based on Fusion Feature MSA and 1D

Post-stroke Dysarthria (PSD) is one of the common sequelae of stroke. PSD can harm patients’ quality of life and, in severe cases, be life-threatening. Most of the existing methods use frequency domain features to recognize the pathological voice, which makes it hard to completely represent the characteristics of pathological voice. Although some results have been achieved, there is still a long way to go for practical applications. Therefore, an improved deep learning-based model is proposed to classify between the pathological voice and the normal voice, using a novel fusion feature (MSA) and an improved 1D ResNet network hybrid bi-directional LSTM with dilated convolution (named 1D DRN-biLSTM). The experimental results show that our fusion features bring greater improvement in pathological speech recognition than the method that only analyzes the MFCC features, and can better synthesize the hidden features that characterize pathological speech. In terms of model structure, the introduction of dilated convolution and LSTM can further improve the performance of the 1D Resnet network, compared to ordinary networks such as CNN and LSTM. The accuracy of this method reaches 82.41% and 100% at the syllable level and speaker level, respectively. Our scheme outperforms other existing methods in terms of feature learning capability and recognition rate, and will help to play an important role in the assessment and diagnosis of PSD in China.