Image Classification Results Research Articles

Deep Learning-Based Pulmonary Nodule Screening: A Narrative Review

AbstractGiven its capacity to generate three-dimensional pictures, computed tomography is the most effective means of detecting lung nodules with more excellent resolution of detected nodules. Small lung nodules can easily be overlooked on chest X-rays, making interpretation difficult. Artificial intelligence algorithms have recently demonstrated remarkable progress in medical imaging, especially with deep learning techniques such as convolutional neural networks (CNNs). CNN produces excellent results in natural image recognition and classification using abundant available data and the computational abilities of modern computers. It further reduces false-positive pulmonary nodules in medical image processing. This review article provides a detailed and inclusive review of recent advances, challenges, performance comparisons, and possible future directions for the problem of pulmonary nodule screening using deep learning methods.

Mosaicking based optimal threshold image enhancement for violence detection with deep quadratic attention mechanism

Violence is a prevalent societal issue that poses significant threats to individuals and communities. To address this challenge, researchers have developed various machine learning models to detect and prevent violent incidents. The violence detection problem statement is a critical challenge that requires innovative and collaborative solutions. By leveraging advancements in artificial intelligence, machine learning, and data analytics, systems are developed to effectively detect violence, prevent its occurrence, and protect vulnerable individuals and communities. Therefore proposed framework utilizes Differential Evolution Algorithm (DEA) along with mosaicking process which helps in enhancing the quality of the image in order to detect the presence of violence. Hockey fight dataset consist of 40 video clips out of which 20 are labelled as fight videos and remaining 20 is labelled as non-fight videos. The image is split as 30:20 ratio and the image will undergo either brightness enhancement or contrast enhancement according to the threshold value obtained by the fitness function. In mosaicking process, Ruled KNN based SURF algorithm is employed in the proposed model as it works faster and helps in providing high accuracy which assist in detecting the violence present in the images. The classification technique employed in the proposed method is Deep Quadratic Attention Mechanism (DQAM). It helps in providing better results for image classification, reducing the convergence rate and helps in improving the normalization. The performance of the proposed method is evaluated using different performance metrics which includes PSNR, SSIM and proposed method is compared with previous phase model in comparative analysis for comparing the efficiency of the proposed and existing model.

Hybrid Deep Learning Framework for Melanoma Diagnosis Using Dermoscopic Medical Images

Background: Melanoma, or skin cancer, is a dangerous form of cancer that is the major cause of the demise of thousands of people around the world. Methods: In recent years, deep learning has become more popular for analyzing and detecting these medical issues. In this paper, a hybrid deep learning approach has been proposed based on U-Net for image segmentation, Inception-ResNet-v2 for feature extraction, and the Vision Transformer model with a self-attention mechanism for refining the features for early and accurate diagnosis and classification of skin cancer. Furthermore, in the proposed approach, hyperparameter tuning helps to obtain more accurate and optimized results for image classification. Results: Dermoscopic shots gathered by the worldwide skin imaging collaboration (ISIC2020) challenge dataset are used in the proposed research work and achieved 98.65% accuracy, 99.20% sensitivity, and 98.03% specificity, which outperforms the other existing approaches for skin cancer classification. Furthermore, the HAM10000 dataset is used for ablation studies to compare and validate the performance of the proposed approach. Conclusions: The achieved outcome suggests that the proposed approach would be able to serve as a valuable tool for assisting dermatologists in the early detection of melanoma.

Adaptive classification of artistic images using multi-scale convolutional neural networks

The current art image classification methods have low recall and accuracy rate issues . To improve the classification performance of art images, a new adaptive classification method is designed employing multi-scale convolutional neural networks (CNNs). Firstly, the multi-scale Retinex algorithm with color recovery is used to complete the enhancement processing of art images. Then the extreme pixel ratio is utilized to evaluate the image quality and obtain the art image that can be analyzed. Afterward, edge detection technology is implemented to extract the key features in the image and use them as initial values of the item to be trained in the classification model. Finally, a multi-scale convolutional neural network (CNN) is constructed by using extended convolutions, and the characteristics of each level network are set. The decision fusion method based on maximum output probability is employed to calculate different subclassifies’ probabilities and determine the final category of an input image to realize the art image adaptive classification. The experimental results show that the proposed method can effectively improve the recall rate and precision rate of art images and obtain reliable image classification results.

A Genetic Programming Approach With Building Block Evolving and Reusing to Image Classification

Genetic programming (GP) has shown promising results in image classification in the last decade. However, most existing GP-based image classification methods often have a complex tree/program structure and a large search space, which may lead to poor performance. To address this, this paper develops a two-stage-based GP approach to automatically evolving solutions/ensembles for image classification. In the new approach, the process of constructing an image classification solution is divided into two stages, i.e., evolving small building blocks for feature extraction and evolving ensembles of classifiers by reusing these blocks. Accordingly, at each stage, a simple tree structure can be designed to facilitate the search. In the first stage, a simple block representation and a new search mechanism including a population updating strategy are developed to evolve diverse and effective blocks. In the second stage, a small set of diverse blocks are selected and transformed into primitives, which produces a new tree representation to evolve ensembles of classifiers for image classification. The new designs allow the proposed approach to construct sufficiently but not over complex solutions for difficult tasks by using/searching small trees. The new approach outperforms most GP-based and non-GP-based comparison methods on five image datasets including CIFAR10, Fashion_MNIST and SVHN. Deep analysis is conducted to provide more insights into the proposed approach.

Efficient visual transformer transferring from neural ODE perspective

AbstractRecently, the Visual Image Transformer (ViT) has revolutionized various domains in computer vision. The transfer of pre‐trained ViT models on large‐scale datasets has proven to be a promising method for downstream tasks. However, traditional transfer methods introduce numerous additional parameters in transformer blocks, posing new challenges in learning downstream tasks. This article proposes an efficient transfer method from the perspective of neural Ordinary Differential Equations (ODEs) to address this issue. On the one hand, the residual connections in the transformer layers can be interpreted as the numerical integration of differential equations. Therefore, the transformer block can be described as two explicit Euler method equations. By dynamically learning the step size in the explicit Euler equation, a highly lightweight method for transferring the transformer block is obtained. On the other hand, a new learnable neural memory ODE block is proposed by taking inspiration from the self‐inhibition mechanism in neural systems. It increases the diversity of dynamical behaviours of the neurons to transfer the head block efficiently and enhances non‐linearity simultaneously. Experimental results in image classification demonstrate that the proposed approach can effectively transfer ViT models and outperform state‐of‐the‐art methods.

A diffusion model multi-scale feature fusion network for imbalanced medical image classification research

Background and objectiveMedicine image classification are important methods of traditional medical image analysis, but the trainable data in medical image classification is highly imbalanced and the accuracy of medical image classification models is low. In view of the above two common problems in medical image classification. This study aims to: (i) effectively solve the problem of poor training effect caused by the imbalance of class imbalanced data sets. (ii) propose a network framework suitable for improving medical image classification results, which needs to be superior to existing methods. MethodsIn this paper, we put in the diffusion model multi-scale feature fusion network (DMSFF), which mainly uses the diffusion generation model to overcome imbalanced classes (DMOIC) on highly imbalanced medical image datasets. At the same time, it is processed according to the cropped image augmentation strategy through cropping (IASTC). Based on this, we use the new dataset to design a multi-scale feature fusion network (MSFF) that can fully utilize multiple hierarchical features. The DMSFF network can effectively solve the problems of small and imbalanced samples and low accuracy in medical image classification. ResultsWe evaluated the performance of the DMSFF network on highly imbalanced medical image classification datasets APTOS2019 and ISIC2018. Compared with other classification models, our proposed DMSFF network achieved significant improvements in classification accuracy and F1 score on two datasets, reaching 0.872, 0.731, and 0.906, 0.836, respectively. ConclusionsOur newly proposed DMSFF architecture outperforms existing methods on two datasets, and verifies the effectiveness of generative model inverse balance for imbalance class datasets and feature enhancement by multi-scale feature fusion. Further, the method can be applied to other class imbalanced data sets where the results will be improved.

Dune Morphology Classification and Dataset Construction Method Based on Unmanned Aerial Vehicle Orthoimagery.

Dunes are the primary geomorphological type in deserts, and the distribution of dune morphologies is of significant importance for studying regional characteristics, formation mechanisms, and evolutionary processes. Traditional dune morphology classification methods rely on visual interpretation by humans, which is not only time-consuming and inefficient but also subjective in classification judgment. These issues have impeded the intelligent development of dune morphology classification. However, convolutional neural network (CNN) models exhibit robust feature representation capabilities for images and have achieved excellent results in image classification, providing a new method for studying dune morphology classification. Therefore, this paper summarizes five typical dune morphologies in the deserts of western Inner Mongolia, which can be used to define and describe most of the dune types in Chinese deserts. Subsequently, field surveys and the experimental collection of unmanned aerial vehicle (UAV) orthoimages for different dune types were conducted. Five different types of dune morphology datasets were constructed through manual segmentation, automatic rule segmentation, random screening, and data augmentation. Finally, the classification of dune morphologies and the exploration of dataset construction methods were conducted using the VGG16 and VGG19 CNN models. The classification results of dune morphologies were comprehensively analyzed using different evaluation metrics. The experimental results indicate that when the regular segmentation scale of UAV orthoimages is 1024 × 1024 pixels with an overlap of 100 pixels, the classification accuracy, precision, recall, and F1-Score of the VGG16 model reached 97.05%, 96.91%, 96.76%, and 96.82%, respectively. The method for constructing a dune morphology dataset from automatically segmented UAV orthoimages provides a reference value for the study of large-scale dune morphology classification.

Design of image classification system for fabric inspection process using Raspberry Pi

This research is designed as a prototype of defect inspection system on fabric production using machine learning-based image processing technology using the open source Google teachable machine application integrated with Raspberry Pi-3B. The prototype of fabric defect inspection system is built by utilizing two rollers that function as a fabric roll house before and after the inspection process. On both rollers, a fabric is stretched to be inspected, so that from a roll of fabric with a certain length, it can be seen how many defects occur on the fabric. The inspection system is carried out using a web camera with a certain level of lighting connected to a raspberry pi as a control device. Raspberry Pi functions as an image processing device and fabric rolling motor controller. In addition to the category of fabric in good condition, this system classifies into two categories of defects, namely slap defects and sparse defects. The test results show that this system has an average frame per second (FPS) of 4.85, an average inference time of 181.1 ms, with an accuracy of image classification results of 98.4 %.

Artificial Intelligence for Computational Remote Sensing: Quantifying Patterns of Land Cover Types around Cheetham Wetlands, Port Phillip Bay, Australia

This paper evaluates the potential of using artificial intelligence (AI) and machine learning (ML) approaches for classification of Landsat satellite imagery for environmental coastal mapping. The aim is to identify changes in patterns of land cover types in a coastal area around Cheetham Wetlands, Port Phillip Bay, Australia. The scripting approach of the Geographic Resources Analysis Support System (GRASS) geographic information system (GIS) uses AI-based methods of image analysis to accurately discriminate land cover types. Four ML algorithms are applied, tested and compared for supervised classification. Technical approaches are based on using the ‘r.learn.train’ module, which employs the scikit-learn library of Python. The methodology includes the following algorithms: (1) random forest (RF), (2) support vector machine (SVM), (3) an ANN-based approach using a multi-layer perceptron (MLP) classifier, and (4) a decision tree classifier (DTC). The tested methods using AI demonstrated robust results for image classification, with the highest overall accuracy exceeding 98% and reached by the SVM and RF models. The presented scripting approach for GRASS GIS accurately detected changes in land cover types in southern Victoria over the period of 2013–2024. From our findings, the use of AI and ML algorithms offers effective solutions for coastal monitoring by analysis of change detection using multi-temporal RS data. The demonstrated methods have potential applications in coastal and wetland monitoring, environmental analysis and urban planning based on Earth observation data.

Hyperspectral image classification based on deep separable residual attention network

Hyperspectral image have rich spatial and spectral information, and how to fully extract and utilize the features of these two dimensions is a research hotspot in hyperspectral classification methods. At present, the unique convolutional operation and deep feature extraction structure of convolutional neural network enable them to have stronger feature representation capabilities and achieve good results in hyperspectral image classification. However, CNN methods do not assign different weights based on the importance of features in the feature extraction process, making it difficult to effectively utilize key features, and most importantly, using fixed shaped convolution kernel can easily overlook the differences between hyperspectral image features. A hyperspectral image classification method based on deep separable residual attention network is proposed to address the above issues. Firstly, to reduce the correlation between hyperspectral image data and minimize the interference of redundant information, principal component analysis is used to reduce the dimensionality of hyperspectral image. Secondly, a shallow feature extraction module is constructed, which can dynamically adjust the size of the receptive field according to the actual situation of the image, adaptively extract shallow features, and reduce the loss of original image features. Then, a depthwise separable residual attention mechanism module is proposed, based on which features are extracted. Starting from global and local features, contextual information on image features in channel and spatial domains is extracted. Finally, use a multi-scale feature fusion module to fully integrate feature maps at different scales. Using Indian Pines, Pavia University and Botswana as experimental datasets, the overall classification accuracy of this paper's method is 98.47 %, 98.70 %, 98.83 % with only 50, 50, 30 training samples per class. The Kappa coefficient is 98.25 %, 98.27 %, and 98.73 %, respectively. Compared with advanced methods, this method not only has higher classification accuracy, but also fully utilizes key features at various network levels.

2bit Nonlinear Diffractive Deep Neural Network (2bit ND2NN): A quantized nonlinear all-optical diffractive deep neural network implementation

To alleviate the problems of long training time and computational resource consumption in diffractive deep neural networks (D2NNs), we present the 2bit nonlinear diffractive deep neural network (2bit ND2NN) model, which employs the concept of quantized neural networks. 2bit ND2NN converts the phase of each diffraction layer pixel from continuous to discrete values, i.e., the five regions of 0, π/2, π, 3π/2 and 2π. We use the formulas for phase and relative thickness to determine the thickness values of the pixels. Furthermore, the phase and amplitude of neurons in the 2bit ND2NN model are both considered as learnable parameters. We use a revised formula to determine the size of neurons in the model, and we determine the diffraction grating spacing, the number of grating layers, the pixel size and the pixel counts through ablation experiments. Experimental results for image classification indicate that the highest accuracy achieved by 2bit ND2NN on the MNIST and Fashion-MNIST datasets is 97.88 % and 89.28 %, respectively.

Adversarial attacks and adversarial training for burn image segmentation based on deep learning.

Deep learning has been widely applied in the fields of image classification and segmentation, while adversarial attacks can impact the model's results in image segmentation and classification. Especially in medical images, due to constraints from factors like shooting angles, environmental lighting, and diverse photography devices, medical images typically contain various forms of noise. In order to address the impact of these physically meaningful disturbances on existing deep learning models in the application of burn image segmentation, we simulate attack methods inspired by natural phenomena and propose an adversarial training approach specifically designed for burn image segmentation. The method is tested on our burn dataset. Through the defensive training using our approach, the segmentation accuracy of adversarial samples, initially at 54%, is elevated to 82.19%, exhibiting a 1.97% improvement compared to conventional adversarial training methods, while substantially reducing the training time. Ablation experiments validate the effectiveness of individual losses, and we assess and compare training results with different adversarial samples using various metrics.

Enhancing surface drainage mapping in eastern Canada with deep learning applied to LiDAR-derived elevation data.

Agricultural dykelands in Nova Scotia rely heavily on a surface drainage technique called land forming, which is used to alter the topography of fields to improve drainage. The presence of land-formed fields provides useful information to better understand land utilization on these lands vulnerable to rising sea levels. Current field boundaries delineation and classification methods, such as manual digitalization and traditional segmentation techniques, are labour-intensive and often require manual and time-consuming parameter selection. In recent years, deep learning (DL) techniques, including convolutional neural networks and Mask R-CNN, have shown promising results in object recognition, image classification, and segmentation tasks. However, there is a gap in applying these techniques to detecting surface drainage patterns on agricultural fields. This paper develops and tests a Mask R-CNN model for detecting land-formed fields on agricultural dykelands using LiDAR-derived elevation data. Specifically, our approach focuses on identifying groups of pixels as cohesive objects within the imagery, a method that represents a significant advancement over pixel-by-pixel classification techniques. The DL model developed in this study demonstrated a strong overall performance, with a mean Average Precision (mAP) of 0.89 across Intersection over Union (IoU) thresholds from 0.5 to 0.95, indicating its effectiveness in detecting land-formed fields. Results also revealed that 53% of Nova Scotia's dykelands are being used for agricultural purposes and approximately 75% (6924 hectares) of these fields were land-formed. By applying deep learning techniques to LiDAR-derived elevation data, this study offers novel insights into surface drainage mapping, enhancing the capability for precise and efficient agricultural land management in regions vulnerable to environmental changes.

Efficient tick-shape networks of full-residual point-depth-point blocks for image classification

Light-weight convolutional neural networks (CNNs) are crucial for deploying computer vision applications in mobile devices thanks to their compact models, small computational complexity, and energy efficiency. However, such models ordinarily have steady-increased channels in their backbone leading to a sharp increase in the model size; while the deficiency of identity mappings in their residual mechanism can lead to modest performance in feature extraction. To mitigate those issues, we propose light-weight networks based on three novel concepts as follows. Firstly, an efficient perceptron is presented to encapsulate point-depth-point (PDP) features extracted by light-weight convolutions along with a full-residual (FR) mechanism through the architecture of a network. This full-residual connection is proposed to deal with the shortcoming of the existing light-weight models whose architecture has incompletely exploited identity mappings. It is due to the variability of spatial dimension caused by several strides of their convolutional operations. Secondly, a tick-shape backbone is then introduced by designing its structure in accordance with the channel elasticity of the FR-PDP perceptron subject to the shape of a check mark. Thirdly, taking advantage of the channel elasticity concept, three tick-shape networks (TickNets) are constructed in a light-weight architecture by hooking one or more tick-shape backbones. Experimental results for image classification on benchmark datasets have clearly corroborated the efficacy of the proposed methods. The code implementation of tick-shape models is available at https://github.com/nttbdrk25/TickNets.

Monolayer directional metasurface for all-optical image classifier doublet.

Diffractive deep neural networks, known for their passivity, high scalability, and high efficiency, offer great potential in holographic imaging, target recognition, and object classification. However, previous endeavors have been hampered by spatial size and alignment. To address these issues, this study introduces a monolayer directional metasurface, aimed at reducing spatial constraints and mitigating alignment issues. Utilizing this methodology, we use MNIST datasets to train diffractive deep neural networks and realize digital classification, revealing that the metasurface can achieve excellent digital image classification results, and the classification accuracy of ideal phase mask plates and metasurface for phase-only modulation can reach 84.73% and 84.85%, respectively. Despite a certain loss of degrees of freedom compared to multi-layer phase mask plates, the single-layer metasurface is easier to fabricate and align, thereby improving spatial utilization efficiency.

An efficient densenet-based deep learning model for Big-4 snake species classification

Snakebite poses a significant health threat in numerous tropical and subtropical nations, with around 5.4 million cases reported annually, which results in 1.8–2.7 million instances of envenomation, underscoring its critical impact on public health. The 'BIG FOUR' group comprises the primary committers responsible for most snake bites in India. Effective management of snakebite victims is essential for prognosis, emphasizing the need for preventive measures to limit snakebite-related deaths.The proposed initiative seeks to develop a transfer learning-based image classification algorithm using DenseNet to identify venomous and non-venomous snakes automatically. The study comprehensively evaluates the image classification results, employing accuracy, F1-score, Recall, and Precision metrics. DenseNet emerges as a potent tool for multiclass snake image classification, achieving a notable accuracy rate of 86%. The proposed algorithm intends to be incorporated into an AI-based snake-trapping device with artificial prey made with tungsten wire and vibration motors to mimic heat and vibration signatures, enhancing its appeal to snakes.The proposed algorithm in this research holds promise as a primary tool for preventing snake bites globally, offering a path toward automated snake capture without human intervention. These findings are significant in preventing snake bites and advancing snakebite mitigation strategies.

Multitask Adversarial Networks Based on Extensive Nonlinear Spiking Neuron Models.

Deep learning technology has been successfully used in Chest X-ray (CXR) images of COVID-19 patients. However, due to the characteristics of COVID-19 pneumonia and X-ray imaging, the deep learning methods still face many challenges, such as lower imaging quality, fewer training samples, complex radiological features and irregular shapes. To address these challenges, this study first introduces an extensive NSNP-like neuron model, and then proposes a multitask adversarial network architecture based on ENSNP-like neurons for chest X-ray images of COVID-19, called MAE-Net. The MAE-Net serves two tasks: (i) converting low-quality CXR images to high-quality images; (ii) classifying CXR images of COVID-19. The adversarial architecture of MAE-Net uses two generators and two discriminators, and two new loss functions have been introduced to guide the optimization of the network. The MAE-Net is tested on four benchmark COVID-19 CXR image datasets and compared them with eight deep learning models. The experimental results show that the proposed MAE-Net can enhance the conversion quality and the accuracy of image classification results.

Improved Landsat Operational Land Imager (OLI) Cloud and Shadow Detection with the Learning Attention Network Algorithm (LANA)

Landsat cloud and cloud shadow detection has a long heritage based on the application of empirical spectral tests to single image pixels, including the Landsat product Fmask algorithm, which uses spectral tests applied to optical and thermal bands to detect clouds and uses the sun-sensor-cloud geometry to detect shadows. Since the Fmask was developed, convolutional neural network (CNN) algorithms, and in particular U-Net algorithms (a type of CNN with a U-shaped network structure), have been developed and are applied to pixels in square patches to take advantage of both spatial and spectral information. The purpose of this study was to develop and assess a new U-Net algorithm that classifies Landsat 8/9 Operational Land Imager (OLI) pixels with higher accuracy than the Fmask algorithm. The algorithm, termed the Learning Attention Network Algorithm (LANA), is a form of U-Net but with an additional attention mechanism (a type of network structure) that, unlike conventional U-Net, uses more spatial pixel information across each image patch. The LANA was trained using 16,861 512 × 512 30 m pixel annotated Landsat 8 OLI patches extracted from 27 images and 69 image subsets that are publicly available and have been used by others for cloud mask algorithm development and assessment. The annotated data were manually refined to improve the annotation and were supplemented with another four annotated images selected to include clear, completely cloudy, and developed land images. The LANA classifies image pixels as either clear, thin cloud, cloud, or cloud shadow. To evaluate the classification accuracy, five annotated Landsat 8 OLI images (composed of >205 million 30 m pixels) were classified, and the results compared with the Fmask and a publicly available U-Net model (U-Net Wieland). The LANA had a 78% overall classification accuracy considering cloud, thin cloud, cloud shadow, and clear classes. As the LANA, Fmask, and U-Net Wieland algorithms have different class legends, their classification results were harmonized to the same three common classes: cloud, cloud shadow, and clear. Considering these three classes, the LANA had the highest (89%) overall accuracy, followed by Fmask (86%), and then U-Net Wieland (85%). The LANA had the highest F1-scores for cloud (0.92), cloud shadow (0.57), and clear (0.89), and the other two algorithms had lower F1-scores, particularly for cloud (Fmask 0.90, U-Net Wieland 0.88) and cloud shadow (Fmask 0.45, U-Net Wieland 0.52). In addition, a time-series evaluation was undertaken to examine the prevalence of undetected clouds and cloud shadows (i.e., omission errors). The band-specific temporal smoothness index (TSIλ) was applied to a year of Landsat 8 OLI surface reflectance observations after discarding pixel observations labelled as cloud or cloud shadow. This was undertaken independently at each gridded pixel location in four 5000 × 5000 30 m pixel Landsat analysis-ready data (ARD) tiles. The TSIλ results broadly reflected the classification accuracy results and indicated that the LANA had the smallest cloud and cloud shadow omission errors, whereas the Fmask had the greatest cloud omission error and the second greatest cloud shadow omission error. Detailed visual examination, true color image examples and classification results are included and confirm these findings. The TSIλ results also highlight the need for algorithm developers to undertake product quality assessment in addition to accuracy assessment. The LANA model, training and evaluation data, and application codes are publicly available for other researchers.

FFT-Based Dynamic Token Mixer for Vision

Multi-head-self-attention (MHSA)-equipped models have achieved notable performance in computer vision. Their computational complexity is proportional to quadratic numbers of pixels in input feature maps, resulting in slow processing, especially when dealing with high-resolution images. New types of token-mixer are proposed as an alternative to MHSA to circumvent this problem: an FFT-based token-mixer involves global operations similar to MHSA but with lower computational complexity. However, despite its attractive properties, the FFT-based token-mixer has not been carefully examined in terms of its compatibility with the rapidly evolving MetaFormer architecture. Here, we propose a novel token-mixer called Dynamic Filter and novel image recognition models, DFFormer and CDFFormer, to close the gaps above. The results of image classification and downstream tasks, analysis, and visualization show that our models are helpful. Notably, their throughput and memory efficiency when dealing with high-resolution image recognition is remarkable. Our results indicate that Dynamic Filter is one of the token-mixer options that should be seriously considered. The code is available at https://github.com/okojoalg/dfformer