Inception Blocks Research Articles

Bio-inspired multi-dimensional deep fusion learning for predicting dynamical aerospace propulsion systems

Rapid and precise forecasting of dynamical systems is critical to ensuring safe aerospace missions. Previous forecasting research has primarily concentrated on global trend analysis using full-scale inputs. However, time series arising from real-world applications such as aerospace propulsion, exhibit a distinct dynamical periodicity over a limited timeframe. Here we develop a deep learning model, TimeWaves, to capture both global trends and local variations, through 3D spectrum-oriented interval extraction from an integrated viewpoint of biological perceptions. Specifically, a shared parameter fusion algorithm is employed to effectively integrate Fourier and Wavelet analyses, providing full and sliced 1D sequences to form 2D tensors that can be seamlessly processed by parameter-efficient inception blocks. Additionally, a dual-way learning workflow using TwinBlock, inspired by the cooperative behavior of visual cells, is implemented to enhance perception of dynamical multi-scale features at a reduced computational cost. TimeWaves demonstrates reliable and robust performance in predicting rocket combustion instability, a key challenge in the aerospace industry.

Modeling fission gas release at the mesoscale using multiscale DenseNet regression with attention mechanism and inception blocks

Mesoscale simulations of fission gas release (FGR) in nuclear fuel provide a powerful tool for understanding how microstructure evolution impacts FGR, but they are computationally intensive. In this study, we present an alternate, data-driven approach, using deep learning to predict instantaneous FGR flux from 2D simulated nuclear fuel microstructure images. Four convolutional neural network (CNN) architectures with multiscale regression are trained and evaluated on simulated FGR data generated using a hybrid phase field/cluster dynamics model. All four networks show high predictive power, with R2 values above 98%. The best performing network combines a Convolutional Block Attention Module (CBAM) and InceptionNet mechanisms to provide superior accuracy (mean absolute percentage error of 4.4%), training stability, and robustness on very low instantaneous FGR flux values. The trained model demonstrates the utility of deep learning in this context, but has no validity outside the 2D model and conditions used to generate the training data.

Classification of Grapevine Varieties Using UAV Hyperspectral Imaging

Classifying grapevine varieties is crucial in precision viticulture, as it allows for accurate estimation of vineyard row growth for different varieties and ensures authenticity in the wine industry. This task can be performed with time-consuming destructive methods, including data collection and analysis in the laboratory. In contrast, unmanned aerial vehicles (UAVs) offer a markedly more efficient and less restrictive method for gathering hyperspectral data, even though they may yield data with higher levels of noise. Therefore, the first task is the processing of these data to correct and downsample large amounts of data. In addition, the hyperspectral signatures of grape varieties are very similar. In this study, we propose the use of a convolutional neural network (CNN) to classify seventeen different varieties of red and white grape cultivars. Instead of classifying individual samples, our approach involves processing samples alongside their surrounding neighborhood for enhanced accuracy. The extraction of spatial and spectral features is addressed with (1) a spatial attention layer and (2) inception blocks. The pipeline goes from data preparation to dataset elaboration, finishing with the training phase. The fitted model is evaluated in terms of response time, accuracy and data separability and is compared with other state-of-the-art CNNs for classifying hyperspectral data. Our network was proven to be much more lightweight by using a limited number of input bands (40) and a reduced number of trainable weights (560 k parameters). Hence, it reduced training time (1 h on average) over the collected hyperspectral dataset. In contrast, other state-of-the-art research requires large networks with several million parameters that require hours to be trained. Despite this, the evaluated metrics showed much better results for our network (approximately 99% overall accuracy), in comparison with previous works barely achieving 81% OA over UAV imagery. This notable OA was similarly observed over satellite data. These results demonstrate the efficiency and robustness of our proposed method across different hyperspectral data sources.

Grow-light smart monitoring system leveraging lightweight deep learning for plant disease classification

This work focuses on a novel lightweight machine learning approach to the task of plant disease classification, posing as a core component of a larger grow-light smart monitoring system. To the extent of our knowledge, this work is the first to implement lightweight convolutional neural network architectures leveraging down-scaled versions of inception blocks, residual connections, and dense residual connections applied without pre-training to the PlantVillage dataset. The novel contributions of this work include the proposal of a smart monitoring framework outline; responsible for detection and classification of ailments via the devised lightweight networks as well as interfacing with LED grow-light fixtures to optimize environmental parameters and lighting control for the growth of plants in a greenhouse system. Lightweight adaptation of dense residual connections achieved the best balance of minimizing model parameters and maximizing performance metrics with accuracy, precision, recall, and F1-scores of 96.75%, 97.62%, 97.59%, and 97.58% respectively, while consisting of only 228,479 model parameters. These results are further compared against various full-scale state-of-the-art model architectures trained on the PlantVillage dataset, of which the proposed down-scaled lightweight models were capable of performing equally to, if not better than many large-scale counterparts with drastically less computational requirements.

Improving Deep CNN-Based Radar Target Classification Performance by Applying a Denoise Filter

This paper presents a novel method for removing noise from range-Doppler images by using a filter prior to conducting target classification using a deep neural network. Specifically, Kuan, Frost, and Lee filters are employed to eliminate speckle noise components from radar data images. Furthermore, a neural network that combines residual and inception blocks (RINet) is proposed. The RINet model is trained and tested on the RAD-DAR dataset—a collection of range-Doppler feature maps. The analysis results show that the application of a Lee filter with a window size of 7 in the RAD-DAR dataset demonstrates the most improvement in the model’s classification performance. On applying this noise filter to the dataset, the RINet model successfully classified radar targets, exhibiting a 4.51% increase in accuracy and a 14.07% decrease in loss compared to the classification results achieved for the original data. Furthermore, a comparison of the RINet model with the noise filtering solution with five other networks was conducted, the results of which show that the proposed model significantly outperforms the others.

Differentiable neural architecture search for domain adaptation in fault diagnosis

In recent years, the application of deep transfer learning in fault diagnosis has gained considerable traction. Notably, the utilization of domain adaptation methods to address the challenges of unsupervised transfer learning has emerged as a focal point of research inquiry. However, prevailing investigations primarily emphasize the construction of loss functions, leaving the design of network structures reliant on manual intervention. To overcome this limitation, a novel approach employing differentiable neural architecture search is proposed. This method facilitates the automatic exploration of network structures that are well-suited for domain adaptation in fault diagnosis. The proposed method explores two key components within the Inception search space: the arrangement of Inception blocks and the feature mask. By stacking Inception blocks, a comprehensive network architecture is constructed, while the feature mask selectively incorporates domain-relevant features. Moreover, to enhance search efficiency, simultaneous exploration and adaptation of multiple target domains are undertaken, thereby expanding the potential application domains of domain adaptation. The effectiveness of the proposed method is evaluated using 3 public datasets through a series of experiments, which demonstrate that this method exhibits a remarkable capacity for generating networks of satisfied performance.

Tomato leaf disease identification via two–stage transfer learning approach

In the last ten years, there has been an upsurge in focus on sustainable agribusiness as a response to the bio-hazards posed by the effects of climate change, severe weather events, population growth, increased demands for food security, and a lack of natural resources. Generally, tomato plants are very susceptible to a wide variety of diseases. So, when it comes to maintaining the quality of tomato crops, having a quick diagnosis that is also correct plays a significant role. Nowadays, deep learning (DL), notably convolutional neural networks (CNNs), have achieved exceptional results in numerous applications, including categorizing tomato plant diseases. This study aims to enhance the disease identification algorithm for tomatoes by using transfer learning to reduce the time needed for the model to be trained and improve its identification accuracy. The model is based on the VGGNet that has been pre-trained using ImageNet and two inception blocks. In addition, the improved categorical cross-entropy loss function for the multi-attribute identification task and two-stage transfer learning were included in the model training process. Our model obtains a better identification accuracy for tomato diseases in the test set than other state-of-the-art techniques, as shown by the experiments' outcomes. The findings indicate substantial values produced by the suggested approach, with 99.23% of accuracy. The suggested model is a beneficial tool for farmers in helping to detect and protect tomatoes from disease due to the considerably high success rate achieved by the model.

Testing different supervised machine learning architectures for the classification of liquid crystals

ABSTRACTDifferent convolutional neural network (CNN) and inception network architectures were trained for the classification of isotropic, nematic, cholesteric and smectic liquid crystal phase textures to test the prediction accuracy for each one of these models. Varying the number of layers and inception blocks, as well as the regularisation, and application to different phase transitions and classification tasks, it is shown that in general the architecture of an inception network with two blocks leads to the best classification results. Regularisation, such as image flipping, and dropout layers additionally somewhat increase the classification accuracy. Even for simple tasks like the isotropic-nematic transition, which is of importance for applications in the automatic readout of sensors, convolutional neural networks need more than one layer. Care must be taken to not apply architectures of too large complexity, as this will again reduce the classification accuracy due to overfitting. Architecture complexity needs to be adjusted to the given classification task

A Novel Image Classification Method Based on Residual Network, Inception, and Proposed Activation Function

In deeper layers, ResNet heavily depends on skip connections and Relu. Although skip connections have demonstrated their usefulness in networks, a major issue arises when the dimensions between layers are not consistent. In such cases, it is necessary to use techniques such as zero-padding or projection to match the dimensions between layers. These adjustments increase the complexity of the network architecture, resulting in an increase in parameter number and a rise in computational costs. Another problem is the vanishing gradient caused by utilizing Relu. In our model, after making appropriate adjustments to the inception blocks, we replace the deeper layers of ResNet with modified inception blocks and Relu with our non-monotonic activation function (NMAF). To reduce parameter number, we use symmetric factorization and 1×1 convolutions. Utilizing these two techniques contributed to reducing the parameter number by around 6 M parameters, which has helped reduce the run time by 30 s/epoch. Unlike Relu, NMAF addresses the deactivation problem of the non-positive number by activating the negative values and outputting small negative numbers instead of zero in Relu, which helped in enhancing the convergence speed and increasing the accuracy by 5%, 15%, and 5% for the non-noisy datasets, and 5%, 6%, 21% for non-noisy datasets.

HIUNET: A Hybrid Inception U-Net for Diagnosis of Diabetic Retinopathy

Type 2 diabetes patients often suffer from microvascular complications of diabetes. These complications, in turn, often lead to vision impairment. Diabetic Retinopathy (DR) detection in its early stage can rescue people from long-term complications that could lead to permanent blindness. In this study, we propose a complex deep convolutional neural network architecture with an inception module for automated diagnosis of DR. The proposed novel Hybrid Inception U-Net (HIUNET) comprises various inception modules connected in the U-Net fashion using activation maximization and filter map to produce the image mask. First, inception blocks were used to enlarge the model’s width by substituting it with primary convolutional layers. Then, aggregation blocks were used to deepen the model to extract more compact and discriminating features. Finally, the downsampling blocks were adopted to reduce the feature map size to decrease the learning time, and the upsampling blocks were used to resize the feature maps. This methodology ensured high prominence to lesion regions compared to the non-lesion regions. The performance of the proposed model was assessed on two benchmark competitive datasets called Asia Pacific Tele-Ophthalmology Society (APTOS) and KAGGLE, attaining accuracy rates of 95% and 92%, respectively.

Machine learning classification of polar sub-phases in liquid crystal MHPOBC.

Experimental polarising microscopy texture images of the fluid smectic phases and sub-phases of the classic liquid crystal MHPOBC were classified as paraelectric (SmA*), ferroelectric (SmC*), ferrielectric (SmC1/3*), and antiferroelectric (SmCA*) using convolutional neural networks, CNNs. Two neural network architectures were tested, a sequential convolutional neural network with varying numbers of layers and a simplified inception model with varying number of inception blocks. Both models are successful in binary classifications between different phases as well as classification between all four phases. Optimised architectures for the multi-phase classification achieved accuracies of (84 ± 2)% and (93 ± 1)% for sequential convolutional and inception networks, respectively. The results of this study contribute to the understanding of how CNNs may be used in classifying liquid crystal phases. Especially the inception model is of sufficient accuracy to allow automated characterization of liquid crystal phase sequences and thus opens a path towards an additional method to determine the phases of novel liquid crystals for applications in electro-optics, photonics or sensors. The outlined procedure of supervised machine learning can be applied to practically all liquid crystal phases and materials, provided the infrastructure of training data and computational power is provided.

DeepParcellation: A novel deep learning method for robust brain magnetic resonance imaging parcellation in older East Asians.

Accurate parcellation of cortical regions is crucial for distinguishing morphometric changes in aged brains, particularly in degenerative brain diseases. Normal aging and neurodegeneration precipitate brain structural changes, leading to distinct tissue contrast and shape in people aged >60 years. Manual parcellation by trained radiologists can yield a highly accurate outline of the brain; however, analyzing large datasets is laborious and expensive. Alternatively, newly-developed computational models can quickly and accurately conduct brain parcellation, although thus far only for the brains of Caucasian individuals. To develop a computational model for the brain parcellation of older East Asians, we trained magnetic resonance images of dimensions 256 × 256 × 256 on 5,035 brains of older East Asians (Gwangju Alzheimer's and Related Dementia) and 2,535 brains of Caucasians. The novel N-way strategy combining three memory reduction techniques inception blocks, dilated convolutions, and attention gates was adopted for our model to overcome the intrinsic memory requirement problem. Our method proved to be compatible with the commonly used parcellation model for Caucasians and showed higher similarity and robust reliability in older aged and East Asian groups. In addition, several brain regions showing the superiority of the parcellation suggest that DeepParcellation has a great potential for applications in neurodegenerative diseases such as Alzheimer's disease.

A Lightweight Two-Stream Model for Driver Emotion Recognition

In recent years, more and more attention has been paid to the traffic safety of drivers. Studies show that the motor vehicle speed will affect the incidence of traffic accidents and traffic accident casualty rate. In this paper, machine learning vision technology is used for real-time recognition of driver emotions, and a lightweight SIC-NET+RC-NET dual-stream recognition model is designed to cope with different scenarios of vehicle speed during driving. SIC-NET is a lightweight CNN network model, which greatly reduces the number of model parameters and computational power requirements, sacrifices part of accuracy, and greatly improves the real-time performance of model recognition. To improve the residual convolutional network model of inception blocks, RC-NET nested multiple micro-convolutional neural networks are used to identity mapping between different inception blocks to enhance the network training effect, expand network width and convolutional depth, and alleviate network degradation. This article stipulates that when the vehicle speed is greater than 60km/h, SIC-NET is used. When the vehicle speed is less than or equal to 60km/h, SIC-NET+RC-NRT is used. The FER2013 dataset was used for testing and validation in this paper. When SIC-NET is used, the model has better real-time performance and can provide a more timely warning when the vehicle speed is higher. When SIC-NET+RC-NET is used, the accuracy of the model in the FER2013 dataset is 69.65%, and the model can better cope with the scenario of a low vehicle speed. The experiment shows that the lightweight dual-flow model has good robustness and identification effect, which solves the problem of using the same model for different vehicle speeds and can effectively ensure the safety of the driver’s life and property.

Driver identification through vehicular CAN bus data: An ensemble deep learning approach

AbstractDriver identification using in‐vehicle data is receiving considerable attention in the field of intelligent transportation owing to the advances in deep learning (DL). In order to improve accuracy and robustness of identification, this paper proposes an ensemble deep learning framework that integrates a modified one‐dimensional convolutional neural network (M 1‐D CNN) and bidirectional long short‐term memory (BLSTM) to improve the performance and robustness of driver identification using information extracted from vehicular CAN‐bus signals. The M 1‐D CNN architecture is developed by adopting inception blocks, residual connection, and global average pooling to obtain optimal deep‐feature representations of local time series. The BLSTM is used to learn the bidirectional long‐term temporal dependencies. Results of extensive experiments using real driving data show that the proposed ensemble DL model can improve the accuracy and robustness of driver identification. Furthermore, four data augmentation methods, namely up‐sampling, adding noise, data reversal, and random drifting, are used to expand the original training data to improve the performance of the ensemble method. Especially, few‐shot learning is performed using the four data augmentation methods, and it shows excellent potential for driver identification with limited data.

Characterization of the in-focus droplets in shadowgraphy systems via deep learning-based image processing method

It is important to accurately identify and measure in-focus droplets from shadowgraph droplet images that typically contain a large number of defocused droplets for the research of multiphase flow. However, conventional in-focus droplet identification methods are time-consuming and laborious due to the noise and background illumination in experimental data. In this paper, a deep learning-based method called focus-droplet generative adversarial network (FocGAN) is developed to automatically detect and characterize the focused droplets in shadow images. A generative adversarial network framework is adopted by our model to output binarized images containing only in-focus droplets, and inception blocks are used in the generator to enhance the extraction of multi-scale features. To emulate the real shadow images, an algorithm based on the Gauss blur method is developed to generate paired datasets to train the networks. The detailed architecture and performance of the model were investigated and evaluated by both the synthetic data and spray experimental data. The results show that the present learning-based method is far superior to the traditional adaptive threshold method in terms of effective extraction rate and accuracy. The comprehensive performance of FocGAN, including detection accuracy and robustness to noise, is higher than that of the model based on a convolutional neural network. Moreover, the identification results of spray images with different droplet number densities clearly exhibit the feasibility of FocGAN in real experiments. This work indicates that the proposed learning-based approach is promising to be widely applied as an efficient and universal tool for processing particle shadowgraph images.

Automatic cyst and kidney segmentation in autosomal dominant polycystic kidney disease: Comparison of U-Net based methods

Autosomal Dominant Polycystic Kidney Disease is a genetic disease that causes uncontrolled growth of fluid-filled cysts in the kidney. Kidney enlargement resulting from the expansion of cysts is continuous and often associated with decreased renal function and kidney failure. Mouse and rat models are necessary to discover new drugs able to halt the progression of the disease. The analysis of the effects of pharmacological interventions in these models is based on renal morphology and quantification of changes in total renal volume and cyst volume. This requires a proper, reproducible and fast segmentation of the kidney images. We propose a set of fully convolutional networks for kidney and cyst segmentation in micro-CT images, based on the U-Net architecture, to compare them and analyze which ones perform better on contrast-enhanced micro-CT images from normal rats and rats with Autosomal Dominant Polycystic Kidney Disease. Networks have been tested on a series images, and the performance has been evaluated in terms of Intersection over Union and Dice coefficients. Results showed that the best performing networks are the U-Net in which a batch normalization layer is applied after each pair of 3 × 3 convolutions, and the U-Net in which convolutional layers are replaced by inception blocks. Results also showed accurate cyst-to-kidney volume ratios obtained from the segmented images, which is one of main metrics of interest. Finally, segmentation performance has been found to be stable as the images in the training set vary. Therefore, the proposed automatic methodology is suitable and immediately applicable to segment cysts and kidney from micro-CT images, and directly provides the cyst-to-kidney volume ratio.

ILU-Net: Inception-Like U-Net for retinal vessel segmentation

High-precision segmentation of retinal vessels is a critical procedure for clinicians to analyze retinal diseases. However, previous U-Net based segmentation methods are lost in the preservation of capillaries and vascular structures of complex bifurcation. To solve these problems, this paper proposes a novel U-Net based network, named Inception-Like U-Net (ILU-Net), including the following key points: First, two different types of inception-like block (Down-sampling Inception Blocks (DIB) in encoder and up-sampling Inception Blocks (UIB) in decoder) replace the two convolution operations in U-Net to learn the characteristics from different sizes of receptive field. Second, a novel skip connection strategy is proposed and can better connect low-level features to high-level features for reproducing more micro-scale vascular structures. Third, the segmentation performance is further improved by using the combination of two activation functions (ReLU and ELU). Qualitative and quantitative results of experiments on the public DRIVE dataset demonstrate that the proposed ILU-Net approach is effective and even better than the state-of-the-arts in preserving capillaries and complex bifurcation structures.

A comparative study of Deep Learning architectures for Classification of Natural and Human-made Sea Events in SAR images

Sea monitoring is essential for a better understanding of its dynamics and to measure the impact of human activities. In this context, remote sensing plays an important role by providing satellite imagery every day, even in critical climate conditions, for the detection of sea events with a potential risk to the environment. The present work proposes a comparative study of Deep Learning architectures for classification of natural and man-made sea events using SAR imagery. The evaluated architectures comprises models based on convolutional networks, inception blocks, and attention modules. Two datasets are employed for this purpose: the first one encompasses a series of natural events (geophysical phenomena), while the second describes a real oil spill scenario in the Gulf of Mexico from 2018 to 2021. As a result, through experimental analysis, it is demonstrated how the Xception and Deep Attention sampling architectures obtained the highest performance metrics, presenting Recall values of 94.2% and 87.4% for the classification of natural and human-made events, respectively.

Cloud Removal in Remote Sensing Images Using Generative Adversarial Networks and SAR-to-Optical Image Translation

Satellite images are often contaminated by clouds. Cloud removal has received special attention due to the wide range of satellite image applications. As the clouds thicken, the process of removing them becomes more challenging. In such cases, using auxiliary images, such as near-infrared or synthetic aperture radar (SAR), for reconstructing is common. In this study, we attempt to solve the problem using two generative adversarial networks (GANs): the first translates SAR images to optical images and the second removes clouds using the translated images of prior GAN. Also, we propose dilated residual inception blocks (DRIBs) instead of vanilla U-net in the generator networks and use structural similarity index measure (SSIM) in addition to the L1 loss function. Reducing the number of downsamplings and expanding receptive fields by dilated convolutions increased the quality of output images. We used the SEN1-2 dataset to train and test both GANs, and we made cloudy images by adding synthetic clouds to optical images. In addition, we used the SEN12MS-CR dataset to test network performance to remove real clouds. The restored images are evaluated using PSNR, SSIM, SAM, MAE, RMSE, and <inline-formula> <tex-math notation="LaTeX">$Q$ </tex-math></inline-formula>. We compared the proposed method with state-of-the-art deep learning models and achieved more accurate results in both SAR-to-optical image translation and cloud removal parts.

Deep Residual Inception Encoded‐Decoder Neural Network for amyloid PET harmonization

AbstractBackgroundAmyloid PET is an in vivo technology to visualize and quantify beta‐amyloid deposition in the brain. However, the use of multiple amyloid tracers with varied characteristics compounded by processing variabilities poses significant challenges to interpret or combine results from cross‐center studies, and to define common positivity threshold. In this research, we developed an encoder‐decoder based deep model as a harmonization strategy to render imputed amyloid PET images of one amyloid tracer to the images of another.Method91 PiB‐florbetapir (FBP) image pairs from the Open Access Series of Imaging Studies (OASIS) were processed using established pipelines to extract regional standard uptake value ratios (SUVRs), mean cortical SUVRs (mcSUVRs), and SUVR images. Residual Inception Encoder‐Decoder Neural Network (RIED‐Net) was implemented to learn the nonlinear mappings from the image pairs in 2D using encoding‐decoding architecture in conjunction with residual inception blocks and generating imputed images. A 10‐fold cross‐validation was implemented on axial, coronal and sagittal views separately to generate imputed PiB SUVR imaging from FBP data. The average imputed PiB image from all three views was used for performance evaluation. Correlation was evaluated between the imputed vs. real PiB SUVR image voxel‐wise, and between the virtual PiB mcSUVR derived from imputed PiB vs. the real PiB mcSUVR. The trained RIED‐Net model was also applied to an independent dataset with 46 subjects from www.gaain.org/centiloid‐project, and the same metric was used to assess performance.ResultThe imputed PiB SUVR images were visually more similar to real PiB SUVR images than FBP. Voxel‐wise correlation improved from 0.89 between PiB and FBP to 0.95 between the synthetic and real PiB SUVR image (p&lt;0.0001) in the cross‐validation. The agreement of mcSUVR improved from r=0.91 to r=0.96 (p&lt;0.0001) in the cross‐validation dataset and from r=0.92 to r=0.96 (p&lt;0.001) in the independent dataset.ConclusionWe proposed a novel encoder‐decoder based deep model for synthetic imaging. The model discovered the voxel‐wise nonlinear associations between the input images and the output images which significantly improved agreements of amyloid burden measurements from different tracers. The result was further confirmed in an independent dataset demonstrating the generalizability.