Networks For Image Classification Research Articles

Physics-informed interactive network for hyperspectral image classification

In hyperspectral image (HSI) classification, convolutional neural networks (CNNs) have made great progress. In recent CNN-based HSI classification methods, spatial attention mechanism has been widely used to emphasize more salient features and suppress less useful ones. However, these methods tend to utilize plug-and-play attention modules which ignore the physical meaning of HSI pixels. To reflect the physical construction of HSI pixels in the attention mask, we propose a physics-informed interactive network (PI2Net) integrating a spatial guided spectral component stream (SGSC-stream) and a spectral constrained spatial attention stream (SCSA-stream). The SGSC-stream can obtain more expressive abundance features of the input HSI cubes through a component analysis block. The SCSA-stream completes the prediction of hyperspectral data in three-stage features extraction cells and a classification cell. To connect and interact between these two streams, a component-aware attention module (CAAM) and a class-wise endmember activation gate (CWEAG) are designed. The CAAM measures the similarity of abundance features to generate a component-aware attention mask, so as to learn the component-based relationships between pixels. Then, the multiscale attention masks are embedded into different stages of the SCSA-stream. The CWEAG is proposed to iteratively generate more accurate abundance features with the classification result of SCSA-stream. Experimental results based on three common hyperspectral data sets demonstrate that PI2Net outperforms several state-of-the-art methods in terms of qualitative and quantitative analysis.

Radiological image analysis using effective channel extension and fusion network based on COVID CT images

Attention mechanisms demonstrate significant capabilities in improving the performance of COVID CT image classification networks. Existing strategies often employ global pooling to acquire feature vectors, overlooking local feature information. Complex attention modules are then applied to enhance network performance, inevitably leading to increased model complexity. To address these issues, this paper introduces an Effective Channel Expansion and Fusion (ECEF) module specifically designed to enhance the performance of medical COVID CT image classification. The module represents a simple yet effective convolutional neural network module. In the channel direction of the given feature map, the module performs segmentation, conducts global pooling operations separately on the segmented and overall feature maps to obtain corresponding attention vectors, and merges them to produce high-dimensional attention vectors. Channel attention feature vectors are derived along the two pooling operation directions and dynamically fused. The resulting attention vectors are multiplied by the input feature map for adaptive feature refinement. Through these operations, the ECEF module adeptly captures local features in COVID CT images, enhancing the classification performance of medical COVID CT images. As a lightweight and versatile module, ECEF seamlessly integrates into any CNN architecture with negligible overhead. It can undergo end-to-end training alongside basic CNNs. ResNet is selected as the backbone network for empirical studies conducted on COVID CT images, including the COVID-CT dataset. Experimental results indicate that, compared to current advanced channel attention mechanisms, the ECEF module exhibits superior performance in COVID CT image classification tasks. The successful application of the ECEF module establishes a foundation for its widespread use in areas such as COVID CT medical image segmentation and object detection.

An improved dung beetle optimization with recurrent convolutional neural networks for efficient detection and classification of undersea water object images

An improved dung beetle optimization with recurrent convolutional neural networks for efficient detection and classification of undersea water object images

Applicability of pre-trained CNNs in temperate deforestation detection

ABSTRACT Technological advancements have opened up new possibilities for accurately identifying deforested areas and developing effective data collection strategies. Our paper proposes a practical approach to improve deforestation monitoring using satellite images and Convolutional Neural Networks (CNNs) for image classification. To train the CNNs, we used labeled data primarily sourced from the Amazon basin, while images from the Romanian Carpathian mountains were reserved for testing. Building on this foundation, we evaluated various pre-trained CNN architectures, including AlexNet, VGGNet, ResNet, DenseNet, EfficientNet, GoogLeNet, and Swin Transformer, comparing their performance with a typical CNN architecture. In addition, we refined the performance by testing four ensemble learning methods. We found a decrease of only 10% in accuracy for using the models on data from the temperate area. Our findings contribute to this interdisciplinary field by providing insights into the effectiveness of pre-trained CNNs in classifying satellite images from temperate regions, trained with knowledge from tropical deforestation data. This contribution may open ways for the development of a quasi real-time deforestation monitoring system and serve as a reference for future research in temperate areas.

A stochastic gradient method with variance control and variable learning rate for Deep Learning

In this paper we study a stochastic gradient algorithm which rules the increase of the mini-batch size in a predefined fashion and automatically adjusts the learning rate by means of a monotone or non-monotone line search procedure. The mini-batch size is incremented at a suitable a priori rate throughout the iterative process in order that the variance of the stochastic gradients is progressively reduced. The a priori rate is not subject to restrictive assumptions, allowing for the possibility of a slow increase in the mini-batch size. On the other hand, the learning rate can vary non-monotonically throughout the iterations, as long as it is appropriately bounded. Convergence results for the proposed method are provided for both convex and non-convex objective functions. Moreover it can be proved that the algorithm enjoys a global linear rate of convergence on strongly convex functions. The low per-iteration cost, the limited memory requirements and the robustness against the hyperparameters setting make the suggested approach well-suited for implementation within the deep learning framework, also for GPGPU-equipped architectures. Numerical results on training deep neural networks for multi-class image classification show a promising behaviour of the proposed scheme with respect to similar state of the art competitors.

Transformer-enhanced two-stream complementary convolutional neural network for hyperspectral image classification

Convolutional neural networks (CNNs) and Transformers have been widely employed in hyperspectral image (HSI) classification, due to their advantages in extracting local-range and global-range features respectively. However, some hybrid CNNs and Transformer-based methods still face some shortcomings in effectively exploiting local-global features in HSI. In this article, a Transformer-enhanced two-stream complementary convolutional neural network is proposed for HSI classification. It mainly comprises the spectral feature extraction stream, spatial feature extraction stream, and a spectral-spatial weight feature complementary module. Firstly, in spectral feature extraction stream, a hybrid convolution block is proposed to extract high-resolution spectral features. And then an attention mechanism, which is based on a Transformer encoder, is used to learn key informative features in spectral channel. The spatial feature extraction stream is designed for spatial feature extraction, and it has a similar architecture to the spectral feature extraction stream. Secondly, a spectral-spatial weight feature complementary module is developed to fully exploit the features which are based on the feature weight in spectral tokens and spatial tokens, they are from spectral and spatial feature extraction streams separately, and then the spectral-spatial tokens with different semantic features are derived. Finally, the spectral-spatial tokens are fed into the linear layer to produce classification outcomes. Furthermore, simplified convolution kernels are employed in our proposed method to decrease the total parameters. Results from experiments on four extensively used HSI datasets highlight the superior performance of our proposed method over other cutting-edge methods.

Digitally filtered resonant arguments for deep learning classification of asteroids in secular resonances

ABSTRACT Node secular resonances, or s-type secular resonances, occur when the precession frequencies of the node of an asteroid and some planets are in commensurability. They are important for changing the proper inclination of asteroids interacting with them. Traditionally, identifying the asteroid resonant status was mostly performed by visual inspection of plots of the time series of the asteroid resonant argument to check for oscillations around an equilibrium point. Recently, deep learning methods based on convolutional neural networks (CNNs) for the automatic classification of images have become more popular for these kinds of tasks, allowing for the classification of thousands of orbits in a few minutes. In this work, we study 11 s-type resonances in the asteroid main belt and in the Hungaria region and focus on the four most diffusive ones. Two secular resonances in the Hungaria region, the 2 · s − s4 − s6 and the s − 2 · s6 + s7 − g6 + g8 overlap, but this has negligible effects in terms of chaotic dynamics. Here, we obtained filtered images of the resonant arguments by filtering out all low-frequency signals with a Butterworth filter. A simple method based on amplitudes and periods of librations can perform a preliminary selection of asteroids in librating orbits. Our results show that CNN models applied to filtered images are much more effective in terms of metrics like accuracy, Precision, Recall, and F1-score than those that use images of osculating resonant arguments. Filtered resonant arguments should be preferentially used to identify asteroids interacting with secular resonances.

KLSANet: Key local semantic alignment Network for few-shot image classification

Few-shot image classification involves recognizing new classes with a limited number of labeled samples. Current local descriptor-based methods, while leveraging consistent low-level features across visible and invisible classes, face challenges including redundant adjacent information, irrelevant partial representation, and limited interpretability. This paper proposes KLSANet, a few-shot image classification approach based on key local semantic alignment network, which aligns key local semantics for accurate classification. Furthermore, we introduce a key local screening module to mitigate the influence of semantically irrelevant image parts on classification. KLSANet demonstrates superior performance on three benchmark datasets (CUB, Stanford Dogs, Stanford Cars), outperforming state-of-the-art methods in 1-shot and 5-shot settings with average improvements of 3.95% and 2.56% respectively. Visualization experiments demonstrate the interpretability of KLSANet predictions. Code is available at: https://github.com/ZitZhengWang/KLSANet.

Training of Convolutional Neural Networks for Image Classification with Fully Decoupled Extended Kalman Filter

Training of Convolutional Neural Networks for Image Classification with Fully Decoupled Extended Kalman Filter

A novel automatic Siamese neural network for hyperspectral image classification

Hyperspectral image is gradually becoming the main medium in satellite observations. It provides rich spectral information of different channels to help explore various gas states. Its classification problem is among the most significant tasks on the hyperspectral image processing and application. But the classification often faces the difficulty of imbalanced data and differentiated intra-class/inter-class difference. A lot of deep learning methods have attempted to solve this problem by designing different kinds of networks, but they suffer from the query of whether the network is optimal. This paper provides an automatically framework to deal with the above problem. First of all, we introduced siamese framework to reduce the impact of imbalanced data and various intra-class/inter-class difference. Then, we designed a multi-branch search space with several convolutional operators, especially a gradient-based updating rules to search the architecture. After using ”time [Formula: see text] performance” criterion, we obtain the model on two public hyperspectral image dataset. The experimental results outperformed all other compared algorithms.

Unsupervised meta-learning with domain adaptation based on a multi-task reconstruction-classification network for few-shot hyperspectral image classification

Although the deep-learning method has achieved great success for hyperspectral image (HSI) classification, the few-shot HSI classification deserves sufficient study because it is difficult and expensive to acquire labeled samples. In fact, the meta-learning methods can improve the performance for few-shot HSI classification effectively. However, most of the existing meta-learning methods for HSI classification are supervised, which still heavily rely on the labeled data for meta-training. Moreover, there are many cross-scene classification tasks in the real world, and domain adaptation of unsupervised meta-learning has been ignored for HSI classification so far. To address the above issues, this paper proposes an unsupervised meta-learning method with domain adaptation based on a multi-task reconstruction-classification network (MRCN) for few-shot HSI Classification. MRCN does not need any labeled data for meta-training, where the pseudo labels are generated by multiple spectral random sampling and data augmentation. The meta-training of MRCN jointly learns a shared encoding representation for two tasks and domains. On the one hand, we design an encoder-classifier to learn the classification task on the source-domain data. On the other hand, we devise an encoder-decoder to learn the reconstruction task on the target-domain data. The experimental results on four HSI datasets demonstrate that MRCN preforms better than several state-of-the-art methods with only two to five labeled samples per class. To the best of our knowledge, the proposed method is the first unsupervised meta-learning method that considers the domain adaptation for few-shot HSI classification.

MPSA: Multi-Position Supervised Soft Attention-based convolutional neural network for histopathological image classification

In recent years, significant achievements have been made in the field of histopathological image analysis using convolutional neural networks (CNNs). However, existing CNNs fail to fully capture the important local structures and regional information in histopathological images due to the complex tissue structures and variable pathological features present in these images. They often treat all regions equally, which further exacerbates the challenge of accurately analyzing such images. Current network model can’t extract deep layer features efficiently without guiding. To alleviate this problem, we propose a novel network model called Multi-Position Supervised Soft Attention (MPSA). MPSA adds regions of interest (RoI) labels at multiple feature layers for deep supervision, and then uses the supervised layers as soft attention to guide the learning of the classification network, enabling the network to accurately extract features of the lesion target. Additionally, we design a Multi-level Attention Feature Enhancement Module (MAFEM), which combines multiple levels of attention mechanisms to enhance the performance of the convolutional neural network in histopathological image classification. MAFEM includes spatial attention, soft attention of the main branch, and our proposed soft attention for multi-branch feature fusion. The proposed soft attention for multi-branch feature fusion aims to enhance the predictive performance of the classification model by activating relevant neurons in the diagnostic area in a highly activated state, while effectively avoiding noise activation. This innovative approach ensures that the model can focus on the most pertinent information, leading to improved classification accuracy. We conducted classification experiments on the liver cancer histopathological images dataset and the results showed that our method achieved a classification accuracy of 95.79%, indicating that it is very effective in the analysis of liver histopathological images. Our proposed network architecture has also demonstrated good generalization ability in other medical datasets, achieving a classification accuracy of 84.41% on the ultrasound carotid plaque dataset.

Dual-stream spectral-spatial convolutional neural network for hyperspectral image classification and optimal band selection

Along with the high spectral rich information it provides, one of the difficulties in processing a hyperspectral image is the need for expert knowledge and high-spec hardware to process very high-dimensional data. 3D convolutional neural network(3D CNN), which uses spectral and spatial features together, enables a powerful solution for HSI classification. This study proposes an efficient dual-stream 3D CNN for accurate HSI classification.The proposed method offers effective classification using spectral-spatial features without relying on pre-processing or post-processing. A comparative study of how CNN classification performance is affected by hyperspectral band selection based on deep reinforcement learning (DRL) is presented. Using the most relevant bands in the hyperspectral image is decisive in deep CNN networks without losing information and accuracy. The proposed method was compared with 3D CNN, 3D + 1D CNN, Multiscale 3D deep convolutional neural network (M3D-DCNN), and InceptionV3 algorithms using Indian Pines (IP), Salinas, Pavia Center (PaviaC), Houston 2013 and QUH-Tangdaowan datasets. It achieved 92.43 % overall accuracy (OA) in IP, 95.06 % OA in Salinas dataset, 99.00 % OA in PaviaC dataset, 91.25 % OA in Houston 2013 and 94.87 % OA in QUH-Tangdaowan. Codes are released at: https://github.com/lapistlazuli/DS-3DCNN.

MSTSENet: Multiscale Spectral–Spatial Transformer with Squeeze and Excitation network for hyperspectral image classification

Hyperspectral image (HSI) classification pertains to the task of assigning a single label to each pixel by analyzing its spectral–spatial characteristics. Convolutional Neural Networks (CNNs) have garnered significant attention on account of their remarkable performance in feature representation. However, these approaches possess a restricted capacity to acquire highly similar spectral features. Recently, Attention-based approaches have been devised as a means to surmount the limitations inherent in CNNs. In this paper, we address a novel architecture for spectral–spatial feature extraction and HSI classification. The framework efficiently amalgamates the robustness and efficacy of three modern backbone networks, i.e., Multiscale CNN, Squeeze and Excitation (SE), and Transformer named MSTSENet. Initially, the multiscale convolution module integrates three branches of 3D convolution layers, each employing distinct kernel sizes, facilitating multiscale and multi-resolution feature extraction. Subsequently, the SE module enhances the inter-channel relationship by adaptively recalibrating the feature weight. The strategic fusion of the SE module with multiscale CNN strengthens deep feature extraction while introducing a minimal increment in overall parameters. Lastly, the Transformer module orchestrates information aggregation across different spectral–spatial regions, accurately modeling long-range contextual dependencies and capturing dominant feature representations. To evaluate the model performance, comprehensive experiments have been carried out on widely used benchmark HSI datasets. The overall accuracy of 98.74%, 99.83%, and 99.98% is achieved over the Indian Pines, Pavia University, and Salinas Valley datasets, respectively. The code of this work will be accessible at https://github.com/irfan01000/MSTSENet.

Medtransnet: advanced gating transformer network for medical image classification

Medtransnet: advanced gating transformer network for medical image classification

LCTCS: Low-Cost and Two-Channel Sparse Network for Hyperspectral Image Classification

Using convolutional neural networks (CNNs) in classifying hyperspectral images (HSIs) has achieved quite good results in recent years. It is widely used in agricultural remote sensing, geological exploration, environmental monitoring, and marine remote sensing. Unfortunately, the complexity of network structures used for hyperspectral image classification challenges the efficient delivery of HSI data extremely, and existing methods suffer from a large amount of redundancy in the network weight parameters during training, as they either require huge computational resources or make inefficient use of storage space when designing the network structure, and many of the parameters that waste computational resources contribute less to the rich spectral and spatial information transfer in HSI. So we introduce LCTCS, a better low-memory and less-parametric network approach. LCTCS aims to improve the efficiency of computational resource utilization with advanced classification performance and lower levels of computational resources. Unlike the conventional 2D and 3D convolution used previously, we use simple and efficient 3D grouped convolution as a vehicle to convey the semantic features of HSIs. More specifically, we design a novel two-channel sparse network to classify HSIs since grouped 3D convolution conveys the properties of hyperspectral data well in the time and space domains.We have compared LCTCS with eight widely used network methods on four publicly available hyperspectral datasets for learning HSI information. A series of experiments shows that the model architecture designed has 65.89%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$65.89 \\%$$\\end{document} less storage space than the DBDA method, consumes 67.36%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$67.36 \\%$$\\end{document} fewer computational resources than the SSRN method on the IP dataset, and accomplishes a highly accurate classification task with the number of parameters accounting for only 1.99%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$1.99 \\%$$\\end{document} that of the DBMA method.

Deep Learning-Based Computer-Aided Diagnosis of Osteochondritis Dissecans of the Humeral Capitellum Using Ultrasound Images.

Ultrasonography is used to diagnose osteochondritis dissecans (OCD) of the humerus; however, its reliability depends on the technical proficiency of the examiner. Recently, computer-aided diagnosis (CAD) using deep learning has been applied in the field of medical science, and high diagnostic accuracy has been reported. We aimed to develop a deep learning-based CAD system for OCD detection on ultrasound images and to evaluate the accuracy of OCD detection using the CAD system. The CAD process comprises 2 steps: humeral capitellum detection using an object-detection algorithm and OCD classification using an image classification network. Four-directional ultrasound images of the elbow of the throwing arm of 196 baseball players (mean age, 11.2 years), including 104 players with normal findings and 92 with OCD, were used for training and validation. An external dataset of 20 baseball players (10 with normal findings and 10 with OCD) was used to evaluate the accuracy of the CAD system. A confusion matrix and the area under the receiver operating characteristic curve (AUC) were used to evaluate the system. Clinical evaluation using the external dataset resulted in high AUCs in all 4 directions: 0.969 for the anterior long axis, 0.966 for the anterior short axis, 0.996 for the posterior long axis, and 0.993 for the posterior short axis. The accuracy of OCD detection thus exceeded 0.9 in all 4 directions. We propose a deep learning-based CAD system to detect OCD lesions on ultrasound images. The CAD system achieved high accuracy in all 4 directions of the elbow. This CAD system with a deep learning model may be useful for OCD screening during medical checkups to reduce the probability of missing an OCD lesion. Diagnostic Level II. See Instructions for Authors for a complete description of levels of evidence.

Cross-modal attention network for retinal disease classification based on multi-modal images.

Multi-modal eye disease screening improves diagnostic accuracy by providing lesion information from different sources. However, existing multi-modal automatic diagnosis methods tend to focus on the specificity of modalities and ignore the spatial correlation of images. This paper proposes a novel cross-modal retinal disease diagnosis network (CRD-Net) that digs out the relevant features from modal images aided for multiple retinal disease diagnosis. Specifically, our model introduces a cross-modal attention (CMA) module to query and adaptively pay attention to the relevant features of the lesion in the different modal images. In addition, we also propose multiple loss functions to fuse features with modality correlation and train a multi-modal retinal image classification network to achieve a more accurate diagnosis. Experimental evaluation on three publicly available datasets shows that our CRD-Net outperforms existing single-modal and multi-modal methods, demonstrating its superior performance.

Quantum convolutional neural networks for multiclass image classification

Quantum convolutional neural networks for multiclass image classification

INSURANCE AMOUNT PREDICTION BASED ON ACCIDENTAL CAR DAMAGE LEVEL USING CNN

There is a paradigm shift in the automotive industry with the advancement of technology, especially in the area of ​​insurance claims handling. In this study, we propose a new method for predicting insurance claims based on car damage levels, utilizing advanced artificial intelligence algorithms for image analysis, particularly employing Convolutional Neural Networks (CNN) for image processing and classification. Our strategy is to provide car damage estimates through picture analysis, making insurance claims in the automobile sector simpler. In order to facilitate user interaction with Flask and increase the predictability of our models, we have also implemented enhancements to the user interface. By combining image analysis with a user-friendly interface, our method offers a practical solution for insurance companies to expedite claim processing, reduce fraudulent activities, and ensure fair settlements for policyholders. This research contributes to the ongoing transformation of insurance claim processing within the automotive industry, highlighting the potential of merging image analysis methodologies with user interfaces for enhanced efficiency and usability. Keywords: Insurance claim prediction, Convolutional Neural Networks (CNNs), Image analysis, User interface, Flask, Automotive industry.