Deep Learning Network Research Articles

Deep Learning Based Detection of Silicosis from Computed Tomography Images

Artificial intelligence has increasingly been used in interpreting medical images to support the timely treatment of diseases by providing early and accurate diagnosis. Pneumoconiosis is a tissue reaction that develops as a result of the accumulation of inorganic dust in the lungs. The most common types of pneumoconiosis include diseases such as coal worker's pneumoconiosis, silicosis, asbestosis, and siderosis. Silicosis, which has maintained its importance since the 1900s and has seen over 182,000 articles published in the last 10 years, is a global health problem. The automated detection and recognition of silicosis in lung computed tomography (CT) images can be considered the backbone of assisting the silicosis diagnosis process. Automated medical assistance systems developed using artificial intelligence can simplify the medical examination process and reduce the time required to start accurate treatment. Although the literature contains various studies that benefit silicosis diagnosis using chest X-ray images or pneumoconiosis diagnosis using CT images, there is not enough classification study that can particularly aid the diagnosis of silicosis in CT images.The method of early detection of silicosis from chest radiographs and CT images has been a challenging task due to the high variability among pneumoconiosis readers. Based on the success of deep learning in the classification and segmentation of medical images, this study has shown that deep learning networks and transfer learning algorithms can detect silicosis with high accuracy by classifying CT images. The performance of the six algorithms examined in the study is compared, and the algorithm with the best performance is recommended. Performance criteria such as accuracy, precision, specificity, and F1-score of the algorithms used in the study were calculated. The accuracy rates of the models were obtained as 92.62%, 93.03%, 92.76%, 95.38%, 97.29%, and 95.17% for AlexNet, VGG16, ResNet50, InceptionV3, Xception, and DenseNet121, respectively. These results show that Xception outperformed the other algorithms and was the most successful algorithm in the automatic detection of silicosis with an accuracy rate of 97.29%.Additionally, a new dataset consisting of tomography images from silicosis patients is presented in this study. Experimental results have shown that transfer learning algorithms can significantly benefit the diagnosis of silicosis by successfully classifying CT images. The findings of the study highlight the clinical importance of artificial intelligence methods in medical image analysis and early disease diagnosis.

Image restoration through few-mode fiber using a new comprehensive attention model

Few-mode fibers (FMF) has unique advantages in single fiber imaging (SFI) in special applications. However, the limited number of transmission modes restricts the information carriers, hence the high-quality image restoration is a major challenge. A deep learning network with new comprehensive attention, called few-mode fiber Swin Transformer (FFST), is developed based on Swin Transformer V2 to realize successful image restoration with a FMF on handwriting letters and numbers. This network is proved outperforming nine widely used convolutional neural networks or Transformer-based models with four evaluation criteria. In addition, single-fiber-multi-location (SFML) experiment is conducted to investigate the effect of different view on imaging feasibility, the result of which shows that SFML can improve the restoration quality despite one object corresponds to multiple different speckle images, indicating that multi-mapping relationships has constructive effect in single FMF imaging. This paper provides a method for imaging in very narrow cavities.

Identification of apple watercore based on ConvNeXt and Vis/NIR spectra

This paper proposes a method for discriminating between normal apples and watercore apples using visible/near-infrared spectroscopy technology, which combines Gramian Angular Fields (GAF) encoding technology and the ConvNeXt deep learning network. The existing feature extraction methods for visible/near-infrared spectroscopy data do not perform deeper information mining on the extracted features, which results in the quality of the established model being entirely determined by the extracted features. Additionally, the process of building a visible/near-infrared spectroscopy data classification model is complex and time-consuming, and the accuracy of the established model is not high. To address these issues, the experimental visible/near-infrared spectroscopy data of apples was first transformed into two-dimensional images using Gramian Angular Summation Fields (GASF) and Gramian Angular Difference Fields (GADF) with sizes of 64, 128, 256, and 512. These images were then input into the ConvNeXt network, and the performance of different encoding methods and sizes was compared. The results showed that, under the conditions provided in this paper, the GADF encoding method with a size of 256 achieved the highest classification accuracy of 98.48%. Next, ResNet, EfficientNet, and RegNet deep learning networks were selected to classify the encoded images under the same conditions. The results above indicate that the apple variety discrimination method based on GAF encoding technology and ConvNeXt network combined with visible/near-infrared spectroscopy technology can achieve deep information mining of visible/near-infrared spectroscopy data and provide a relatively simple method for establishing qualitative classification models of visible/near-infrared spectroscopy. This method has a relatively excellent discrimination effect between normal apples and watercore apples.

Application of machine learning for the differentiation of thymomas and thymic cysts using deep transfer learning: A multi-center comparison of diagnostic performance based on different dimensional models.

This study aimed to evaluate the feasibility and performance of deep transfer learning (DTL) networks with different types and dimensions in differentiating thymomas from thymic cysts in a retrospective cohort. Based on chest-enhanced computed tomography (CT), the region of interest was delineated, and the maximum cross section of the lesion was selected as the input image. Five convolutional neural networks (CNNs) and Vision Transformer (ViT) were used to construct a 2D DTL model. The 2D model constructed by the maximum section (n) and the upper and lower layers (n - 1, n + 1) of the lesion was used for feature extraction, and the features were selected. The remaining features were pre-fused to construct a 2.5D model. The whole lesion image was selected for input and constructing a 3D model. In the 2D model, the area under curve (AUC) of Resnet50 was 0.950 in the training cohort and 0.907 in the internal validation cohort. In the 2.5D model, the AUCs of Vgg11 in the internal validation cohort and external validation cohort 1 were 0.937 and 0.965, respectively. The AUCs of Inception_v3 in the training cohort and external validation cohort 2 were 0.981 and 0.950, respectively. The AUC values of 3D_Resnet50 in the four cohorts were 0.987, 0.937, 0.938, and 0.905. The DTL model based on multiple different dimensions can be used as a highly sensitive and specific tool for the non-invasive differential diagnosis of thymomas and thymic cysts to assist clinicians in decision-making.

Solutions to elliptic and parabolic problems via finite difference based unsupervised small linear convolutional neural networks

In recent years, there has been a growing interest in leveraging deep learning and neural networks to address scientific problems, particularly in solving partial differential equations (PDEs). However, many neural network-based methods like PINNs rely on auto differentiation and sampling collocation points, leading to a lack of interpretability and lower accuracy than traditional numerical methods. As a result, we propose a fully unsupervised approach, requiring no training data, to estimate finite difference solutions for PDEs directly via small linear convolutional neural networks. Our proposed approach uses substantially fewer parameters than similar finite difference-based approaches while also demonstrating comparable accuracy to the true solution for several selected elliptic and parabolic problems compared to the finite difference method.

Toward high-quality pseudo masks from noisy or weak annotations for robust medical image segmentation

Deep learning networks excel in image segmentation with abundant accurately annotated training samples. However, in medical applications, acquiring large quantities of high-quality labeled images is prohibitively expensive. Thus, learning from imperfect annotations (e.g. noisy or weak annotations) has emerged as a prominent research area in medical image segmentation. This work aims to extract high-quality pseudo masks from imperfect annotations with the assistance of a small number of clean labels. Our core motivation is based on the understanding that different types of flawed imperfect annotations inherently exhibit unique noise patterns. Comparing clean annotations with corresponding imperfectly annotated labels can effectively identify potential noise patterns at minimal additional cost. To this end, we propose a two-phase framework including a noise identification network and a noise-robust segmentation network. The former network implicitly learns noise patterns and revises labels accordingly. It includes a three-branch network to identify different types of noises. The latter one further mitigates the negative influence of residual annotation noises based on parallel segmentation networks with different initializations and a label softening strategy. Extensive experimental results on two public datasets demonstrate that our method can effectively refine annotation flaws and achieve superior segmentation performance to the state-of-the-art methods.

Small aircraft detection in infrared aerial imagery based on deep neural network

Detection of aerial target is an important part of infrared image processing. Both neural network method and traditional method can be used in infrared object detection. Neural network method has many advantages such as high accuracy and good portability compared with traditional object detection method. Since the features extracted by neural network method can change over detection target, automatic feature extraction makes neural network based detection method more effective. In recent years deep learning method has been also found wide use for object detection in images. In this paper, an object detection model based on the deep learning network YOLO is constructed for solving the infrared aircraft detection problem. We construct the dataset used for training and testing with recognized features being iteratively learned. The task of infrared otject detection is sensitive to model size and detection speed. There is a requirement of using quantization method to reduce the storage space and the computation complexity. We propose a quantized model with appropriate accuracy for infrared object detection task. To solve the detection task for multiple extremely small aircrafts, model adjustment and quantization are used in proposed model and it gets a better performance. Experimental results on the constructed dataset show that the storage space for weight after quantization shrinks to a quarter, and there is no precision loss for extremely small aircrafts compared to the original model. The optimized YOLO-based deep learning model is effective to detect the small aircraft target in infrared aerial imagery.

A Lightweight Future Skeleton Generation Network(FSGN) Based on Spatio-Temporal Encoding and Decoding

Since early warning in industrial applications is far more valuable than post-event analysis, human activity prediction based on partially observed skeleton sequences has become a popular research area. Recent studies focus on building complex deep learning networks to generate accurate future skeleton data, but overlook the requirement for timeliness. Different from such frame-by-frame generation methods, we propose a Future Skeleton Generation Network (FSGN) based on spatio-temporal encoding and decoding framework. Firstly, we design a dynamically regulated input module to ensure equal-length input of partially observed data, and set modules like discrete cosine transform(DCT) and low-pass filtering(LPF) to filter important information. Then, we employ an improved multi-layer perceptron(MLP) structure as the basic computational unit for the encoding and decoding framework to extract spatio-temporal information, and propose using multi-dimensional motion error of human skeleton to form the loss function. Finally, we use an output module symmetrical to the input module to achieve the generation of future activity data. Results show that the proposed FSGN achieves fewer parameters(0.12M) and higher generation accuracy, which can effectively provide future information for human activity prediction tasks.

Machine vision-based detection of surface defects in cylindrical battery cases

Automotive 21700 series lithium batteries are prone to surface defects during production and transportation, thus affecting their performance, so we propose a full-surface defect detection method for battery cases based on the synthesis of traditional image processing and deep learning to address this problem. First, the mechanism of surface defects on a battery case is analysed, and the types of surface defects are summarized. A suitable platform for image acquisition is designed for the severely reflective surface. Since there is no publicly available defect dataset for cylindrical battery cases, a defect dataset is established, and the dataset is augmented and expanded via the traditional method and the ACGAN model. For the defects on the top of the battery case, a traditional image processing algorithm is used to combine the roundness and the area of the area pixels to make a comprehensive judgement. For the defects on the bottom and side of the battery case, after comparing and analysing a variety of deep learning networks, the YOLOv7 model is selected to address the data characteristics of the large experimental inputs and the small targeted defects. To improve the accuracy of the model to meet the needs of real-time detection in industry, the CA attention mechanism, the DYHEAD dynamic detector head, and the slicing-assisted super inference (SAHI) method are used, with a final map accuracy reaching 98.1 %, which is 2.7 % higher than that of the initial model. Compared with mainstream target detection algorithms, the algorithm in this paper has good detection performance for cylindrical battery case defect detection and can be better applied to real-time detection in industry.

A soft scanning electron microscopy for efficient segmentation of alloy microstructures based on a new self-supervised pre-training deep learning network

To provide an on-site metallographic segmentation using only optical microscopy images, sSEM-Net, a soft scanning electron microscopy network, is developed based on a self-supervised pre-training deep learning framework. During model training, only a sparse collection of SEM images is necessary for annotation assistance. By integrating CNN and Transformer, sSEM-Net efficiently utilizes global context information while mitigating data dependency and computational resource constraints. Using only readily available optical microscopy images as input, sSEM-Net achieves metallographic segmentation comparable to SEM images, catering to rapid and cost-effective industrial needs. This methodology leverages non-destructive inspection attributes, catering to rapid and cost-sensitive industrial requirements. The efficacy of the proposed sSEM-Net is demonstrated through metallographic structure analysis of TC4 titanium alloy, with potential extensions to other alloy types.

Fluorescence spectroscopy combined with multilayer perceptron deep learning to identify the authenticity of monofloral honey—Rape honey

Honey authenticity is critical to honey quality. The development of a quick, easy, and non-destructive technique for determining the authenticity of honey encourages an improvement in honey quality. Here, the authenticity of monofloral honey—rape honey was determined using fluorescence spectroscopy combined with multilayer perceptron (MLP) deep learning, without the need for any prior feature extraction or pre-processing. A total of 91 raw fluorescence intensity data of the real and adulterated honey samples at a fixed excitation wavelength of 280 nm were first matrixed, and all data were then categorized into a training set, a validation set, and a test set with numbers of 64, 16, and 11, respectively. The connection with dropout was selected to build and link the MLP internal network. The activation function, learning rate, optimizer, and number of epochs were among the hyperparameters of the MLP neural network that were tuned. A good MLP deep learning network model for determining the authenticity of monofloral honey, rape honey, was developed after constant validation and debugging. According to the accuracy curve of the MLP model, the accuracy of the training set increased with the number of epochs and eventually converged to 100 %, while the accuracy of the validation set could be well stabilized at about 100 % after 5000 epochs. Finally, the accuracy of the MLP model on the test set was close to 100 %. According to our findings, the MLP neural network and fluorescence intensity have great potential applications in identifying the authenticity of honey.

CNN Intelligent diagnosis method for bearing incipient faint faults based on adaptive stochastic resonance-wave peak cross correlation sliding sampling

As a representative of deep learning networks, convolutional neural networks (CNN) have been widely used in bearing fault diagnosis with good results. However, the signal length and segmentation of the input CNN can have a significant impact on diagnostic accuracy. In addition, the signal-to-noise ratio of early bearing faults is usually very low, which makes it difficult for traditional CNNs to accurately identify and classify these faults. To solve this problem, this paper proposes an adaptive stochastic resonance wave peak cross-correlation sliding sampling method. Firstly, the adaptive stochastic resonance is used to reduce the noise of the original signal, and then the data is divided from the position of the signal wave peak, the correlation coefficient between the divided signals is calculated, and the maximum value is found to determine the size of the division window. Finally, it is converted into a 2D image by Gramian Angular Field and input into CNN for diagnostic classification. The design methodology was validated using the Case Western Reserve University bearing dataset. Subsequently, three validation strategies were established on a self-built platform, including mixed diagnosis of 10 different bearing states, variable speed diagnosis, and low sampling data diagnosis. The proposed method outperforms the conventional CNN by 10% in the Case Western Reserve University dataset test set. The variable speed test set is 24.67% and 31.17% higher, respectively. It is 30% higher in low sampling data diagnosis.

Research on Pavement Defect Detection Algorithm Based on SEM-YOLOv8n

Automated detection and identification of pavement distresses is essential for timely pavement repair. Subtle pavement defects and multiple defects detection is a challenging task under complex background. With the deepening of the deep learning network, some subtle features tend to disappear and are more difficult to detect under the influence of the complex background. To solve the above problems, this paper proposes the SEM-YOLOv8n pavement defect detection algorithm. Firstly, SPD-Conv is used to replace the traditional convolution, which is conducive to retaining more defect detail information in the image and improving the detection ability of subtle defects; then an efficient multi-scale attention mechanism is added to the fusion network, so that the network suppresses the background information and focuses more on the defect information. Finally, MPDIoU is introduced as a loss function, which optimizes the minimum perpendicular distance between the predicted bounding box and the real bounding box and improves the localization ability, thus improving the accuracy of the network. Finally, the effectiveness of the proposed network is verified on the IRRDD dataset, and the results show that the method achieves 91.9% (Precision), 91.3% (Recall), and 71.3% (mAP) for the classification and detection of road multi-scale minor defects, which meets the demand of real-time road defect detection.

A Multi-Scale Feature Fusion Deep Learning Network for the Extraction of Cropland Based on Landsat Data

In the face of global population growth and climate change, the protection and rational utilization of cropland are crucial for food security and ecological balance. However, the complex topography and unique ecological environment of the Qinghai-Tibet Plateau results in a lack of high-precision cropland monitoring data. Therefore, this paper constructs a high-quality cropland dataset for the YarlungZangbo-Lhasa-Nyangqv River region of the Qinghai-Tibet Plateau and proposes an MSC-ResUNet model for cropland extraction based on Landsat data. The dataset is annotated at the pixel level, comprising 61 Landsat 8 images in 2023. The MSC-ResUNet model innovatively combines multiscale features through residual connections and multiscale skip connections, effectively capturing features ranging from low-level spatial details to high-level semantic information and further enhances performance by incorporating depthwise separable convolutions as part of the feature fusion process. Experimental results indicate that MSC-ResUNet achieves superior accuracy compared to other models, with F1 scores of 0.826 and 0.856, and MCC values of 0.816 and 0.847, in regional robustness and temporal transferability tests, respectively. Performance analysis across different months and band combinations demonstrates that the model maintains high recognition accuracy during both growing and non-growing seasons, despite the study area’s complex landforms and diverse crops.

Enhancing Biometric Security with Bimodal Deep Learning and Feature-level Fusion of Facial and Voice Data

Recent research in biometric technologies underscores the benefits of multimodal systems that use multiple traits to enhance security by complicating the replication of samples from genuine users. To address this, we present a bimodal deep learning network (BDLN or BNet) that integrates facial and voice modalities. Voice features are extracted using the SincNet architecture, and facial image features are obtained from convolutional layers. Proposed network fuses these feature vectors using either averaging or concatenation methods. A dense connected layer then processes the combined vector to produce a dual-modal vector that encapsulates distinctive user features. This dual-modal vector, processed through a softmax activation function and another dense connected layer, is used for identification. The presented system achieved an identification accuracy of 99% and a low equal error rate (EER) of 0.13% for verification. These results, derived from the VidTimit and BIOMEX-DB datasets, highlight the effectiveness of the proposed bimodal approach in improving biometric security.

Fast inter-frame motion correction in contrast-free ultrasound quantitative microvasculature imaging using deep learning

Contrast-free ultrasound quantitative microvasculature imaging shows promise in several applications, including the assessment of benign and malignant lesions. However, motion represents one of the major challenges in imaging tumor microvessels in organs that are prone to physiological motions. This study aims at addressing potential microvessel image degradation in in vivo human thyroid due to its proximity to carotid artery. The pulsation of the carotid artery induces inter-frame motion that significantly degrades microvasculature images, resulting in diagnostic errors. The main objective of this study is to reduce inter-frame motion artifacts in high-frame-rate ultrasound imaging to achieve a more accurate visualization of tumor microvessel features. We propose a low-complex deep learning network comprising depth-wise separable convolutional layers and hybrid adaptive and squeeze-and-excite attention mechanisms to correct inter-frame motion in high-frame-rate images. Rigorous validation using phantom and in-vivo data with simulated inter-frame motion indicates average improvements of 35% in Pearson correlation coefficients (PCCs) between motion corrected and reference data with respect to that of motion corrupted data. Further, reconstruction of microvasculature images using motion-corrected frames demonstrates PCC improvement from 31 to 35%. Another thorough validation using in-vivo thyroid data with physiological inter-frame motion demonstrates average improvement of 20% in PCC and 40% in mean inter-frame correlation. Finally, comparison with the conventional image registration method indicates the suitability of proposed network for real-time inter-frame motion correction with 5000 times reduction in motion corrected frame prediction latency.

Ancient Yi Script Handwriting Sample Repository

The ancient Yi script has been used for over 8000 years, which can be ranked with Oracle,Sumerian,Egyptian,Mayan and Harappan,and is one of the six ancient scripts in the world. In this article, we collected 2922 handwritten single word samples of commonly used ancient Yi characters. Each character was written by 310 people respectively, with a total of 427,939 valid characters. We completed continuous handwritten text sampling, written by 250 people, with 5 texts per person, covering topics such as Yi astronomy, geography, rituals, and agriculture. In the process of data collection, we proposed an automatic sampling method for ancient Yi script, and completed the automatic cutting and labeling of handwritten samples. Furthermore, we tested the recognition performance of the sorted data set under different deep learning network models. The results show that ancient Yi script has diverse shape structures and rich writing styles, which can be used as a benchmark data set in related fields such as handwritten text recognition and handwritten text generation.

Enhanced Multi-Scale Attention-Driven 3D Human Reconstruction from Single Image

Due to the inherent limitations of a single viewpoint, reconstructing 3D human meshes from a single image has long been a challenging task. While deep learning networks enable us to approximate the shape of unseen sides, capturing the texture details of the non-visible side remains difficult with just one image. Traditional methods utilize Generative Adversarial Networks (GANs) to predict the normal maps of the non-visible side, thereby inferring detailed textures and wrinkles on the model’s surface. However, we have identified challenges with existing normal prediction networks when dealing with complex scenes, such as a lack of focus on local features and insufficient modeling of spatial relationships.To address these challenges, we introduce EMAR—Enhanced Multi-scale Attention-Driven Single-Image 3D Human Reconstruction. This approach incorporates a novel Enhanced Multi-Scale Attention (EMSA) mechanism, which excels at capturing intricate features and global relationships in complex scenes. EMSA surpasses traditional single-scale attention mechanisms by adaptively adjusting the weights between features, enabling the network to more effectively leverage information across various scales. Furthermore, we have improved the feature fusion method to better integrate representations from different scales. This enhanced feature fusion allows the network to more comprehensively understand both fine details and global structures within the image. Finally, we have designed a hybrid loss function tailored to the introduced attention mechanism and feature fusion method, optimizing the network’s training process and enhancing the quality of reconstruction results. Our network demonstrates significant improvements in performance for 3D human model reconstruction. Experimental results show that our method exhibits greater robustness to challenging poses compared to traditional single-scale approaches.

Sign language recognition using modified deep learning network and hybrid optimization: a hybrid optimizer (HO) based optimized CNNSa-LSTM approach.

Speech impairment limits a person's capacity for oral and auditory communication. Improvements in communication between the deaf and the general public can be progressed by a real-time sign language detector. Recent studies have contributed to make progress in motion and gesture identification processes using Deep Learning (DL) methods and computer vision. But the development of static and dynamic sign language recognition (SLR) models is still a challenging area of research. The difficulty is in obtaining an appropriate model that addresses the challenges of continuous signs that are independent of the signer. Different signers' speeds, durations, and many other factors make it challenging to create a model with high accuracy and continuity. This study mainly focused on SLR using a modified DL and hybrid optimization approach. Notably, spatial and geometric-based features are extracted via the Visual Geometry Group 16 (VGG16), and motion features are extracted using the optical flow approach. A new DL model, CNNSa-LSTM, is a combination of a Convolutional Neural Network (CNN), Self-Attention (SA), and Long-Short-Term Memory (LSTM) to identify sign language. This model is developed for feature extraction by combining CNNs for spatial analysis with SA mechanisms for focusing on relevant features, while LSTM effectively models temporal dependencies. The proposed CNNSa-LSTM model enhances performance in tasks involving complex, sequential data, such as sign language processing. Besides, a Hybrid Optimizer (HO) is proposed using the Hippopotamus Optimization Algorithm (HOA) and the Pathfinder Algorithm (PFA). The proposed model has been implemented in Python, and it has been evaluated over the existing models in terms of accuracy (98.7%), sensitivity (98.2%), precision (98.5%), Word Error Rate (WER) (0.131), Sign Error Rate (SER) (0.114), and Normalized Discounted Cumulative Gain (NDCG) (98%) as well. The proposed model has recorded the highest accuracy of 98.7%.

A Plantar Pressure Detection and Gait Analysis System Based on Flexible Triboelectric Pressure Sensor Array and Deep Learning.

Gait detection is essential for the assessment of human health status and early diagnosis of diseases. The current gait analysis systems are bulky, limited in the scope of use, and cause interference with the movement of the measured person. Hence, it is necessary to develop a wearable gait detection system that is soft, breathable, lightweight, and self-powered. Here, a plantar pressure sensor array and gait analysis system based on a flexible triboelectric pressure sensor (FTPS) array is developed. Soft, breathable, and wearable electrospinning nanofiber film with excellent triboelectric properties is used as the plantar pressure sensor, achieving a high sensitivity of 45.1 mVkPa-1 in the range of 40-200 kPa and 19.4 mVkPa-1 in the range of 200-400 kpa. 32 FTPSs are integrated into an intelligent insole, which has the characteristics of soft, easy production, good air permeability, long-time stability, no external power supply, and etc. Based on the long short-term memory artificial neural network deep learning model, the accuracy of gait judgment can reach 94.23%. This work provides a feasible solution for real-time gait detection, which will have potential applications in human health assessment and early diagnosis of disease.