Lung diseases often result in severe damage to the respiratory tract, and lead to a high risk of mortality within a short period of time. DL models based on ViT are considered to have promising advantages over CNN architectures in terms of computational efficiency, and accuracy when trained on large ImageNet datasets. In this study, we present a new DL approach based on the combination of CNN with ViT to improve the efficiency of pneumonia diagnosis using medical images. In the first stage, raw images are passed through a local filter to capture local relations on the inputs. The local filter block includes two convolutional layers with kernel 3 × 3. This local filtering method aims to enhance rich features before being fed into the patching layer of the ViT block. The proposed method is experimented on the benchmark chest X-ray dataset. The proposed method is evaluated and compared to some well-known models, which include ViT, VGG19, Resnet50, Densnet201. Experimental results demonstrated that the proposed approach based on CNN and ViT reaches higher efficiency with about 1% accuracy to the standard ViT model, and about 2% higher with VGG19, Resnet50, Densnet201 and smaller in model architecture.

The AcT (Action Transformer) model has shown promising results in action recognition tasks. However, achieving high accuracy in complex and dynamic action sequences remains a challenge. In this paper, we present an approach to improve the accuracy of the AcT model by increasing the model's training complexity, validated on the MPOSE2021 and MSR Action datasets. Our method enhances the AcT model by incorporating a multi-level feature fusion technique. We introduce additional convolutional and pooling layers to capture more detailed spatial and temporal information from the input data. This increases the model's ability to discriminate between subtle action variations and improves its accuracy in recognizing complex actions. We evaluate the effectiveness of our proposed approach through extensive experiments on the MPOSE2021 and MSR Action datasets. The results demonstrate that our enhanced AcT model achieves significantly improved accuracy compared to the baseline AcT model and outperforms existing state-of-the-art methods. Our method effectively captures the intricacies of complex actions and provides more accurate predictions.

In the realm of automobile insurance, the imperative of automating car damage evaluation has surged, offering streamlined assessment processes and heightened accuracy. Deep learning techniques have notably influenced vehicle damage assessment, reshaping insurance procedures. However, the primary challenge remains in crafting robust models for damage detection and segmentation. This study presents a novel contribution through the development of the Vehicle Damage Detection (VehiDE) dataset, specifically tailored for comprehensive car damage assessment. The dataset, encompassing 13,945 high-resolution images annotated across eight damage categories, serves as a foundational resource for advancing automated damage identification methodologies. Notably, VehiDE offers support for multiple tasks, including classification, object detection, instance segmentation, and salient object detection, thereby fostering versatile research avenues. Through extensive experimental analysis, including the evaluation of state-of-the-art methodologies on VehiDE, this study not only highlights the dataset's efficacy but also unveils new insights into the challenging nature of car damage assessment. Moreover, the study pioneers the exploration of salient object detection techniques in this domain, showcasing their potential in addressing irregular damage types. By offering VehiDE to the research community, we aim to catalyze advancements in the field of car damage assessment, paving the way for more accurate and efficient automated systems.

In the medical diagnosis, the accuracy of decision-making holds immense significance. Healthcare professionals heavily rely on diagnostic results to guide treatment plans, as these decisions have a substantial impact on patients' health outcomes. An incorrect diagnosis can lead to detrimental consequences, including treatment delays or the administration of inappropriate treatments, thereby posing serious harm to the patient's well-being. The utilization of artificial intelligence (AI) in medical image recognition plays a crucial role in assisting various disease diagnosis systems. Improving accuracy of brain tumor classification is an important diagnosis and effective treatment planning. In this paper, we present an approach to improve brain tumor classification accuracy based on vision transformer approach and customizing head classifier and hyperparameter optimization for training. ViT model are a new type of neural network that can achieve state-of-the-art accuracy on image classification tasks. Our learning architecture approach reaches high accuracy with more than 95%. The experimental results demonstrated that the ViT model achieves a higher validation accuracy. This approach also show that the ViT model is more robust to noise and can generalize better to new data.

In the field of dermatological diseases, especially for skin cancer, machine learning (ML) methods are used to classify melanoma and nevus using skin images. ML techniques result in high accuracy of diagnostic tasks since they are trained on balanced datasets. However, MLs working with imbalanced datasets produce erroneous results on precision, sensitivity, and specificity measured criteria. To deal with this problem, an augmentation approach combined with a category seesaw is used for the compensation factor. It increases the penalty for misclassified instances, thereby reducing the occurrence of false positives within the less common categories. This paper presents an approach to improve the efficiency of DCNN for classifying multi-class medical images on imbalanced datasets. The solution consists of three major contributions: (1) feature extraction based on some backbone models with customizing fully connected layers for classifier layers, (2) optimizing loss function (LF) and training parameters, (3) solving the problem of imbalanced samples using optimizing domination of weights between asymmetric classes with majority and minority categories. The method was evaluated and analyzed using the ISIC2018 benchmark and Chest X-ray dataset. Some well-known backbones were used for this study, e.g., EfficientNets, MobileNets, and DenseNets. The use of these backbones is to demonstrate that our methods are more efficient and stable in both light and heavy DCNN architectures. We also provide comparisons with existing methods that deal with the imbalance problem, e.g., data augmentation (AU), downsamples, customizing LF, and focal loss method (FL) for focusing on hard samples. Experimental results showed that these methods achieve good performance. However, there are several problems caused by generating new samples, and weighting samples, such as data overloading to train classifier models, a corrupt problem when applied to imbalanced data. Moreover, the FL method produced insufficient results on various DCNN backbones. Differently, our approach solves the imbalanced dataset based on boosting the sample weights of the minority and reducing the impact ratio of samples in majority categories. This strategy results in high precision and stable performance with various DCNN models without augmenting the dataset. Experiment results on ISIC2018 dataset demonstrated that our approach achieves more efficiency than other methods in some specific evaluation criteria as follows: higher than the FL method with 2.73% recall, 2.63% precision, 2.81% specificity, and 3.09% F1 using EfficientNet backbones; higher than AU method with 5.16% recall, 5.97% precision, 8.93% specificity, 6.16% F1 using DenseNet backbones.

In the world of auto insurance, automatic car damage identification has garnered a lot of interest. However, it is difficult for us to develop a workable model for car damage identification due to the absence of high-quality datasets that are accessible to the general public. In order to achieve this, the Vehicle Damage Detection (VehiDE) dataset, the large-scale dataset made available to the public for the purpose of segmenting and detecting visual automotive damage. This dataset comprises 13,945 high-resolution photos of damaged cars together with more than 32,000 occurrences of each damage category with detailed annotations. Statistical dataset analysis is provided together with a description of the image collecting, selection, and annotation procedures. In order to emphasize the expertise of automotive damage identification, extensive experiments on the VehiDE dataset are conducted using cutting-edge deep approaches for a variety of jobs and provide thorough analysis.

Over a decade, deep learning methods using convolutional neural network (CNN) architecture have achieved breakthroughs in the precision criterion, which compared to the traditional machine learning methods. However, those approaches still faced some limitations of processing time and precision when they are applied to large samples and hard datasets. Recently, some new methods based on the transformer learning approach have been applied to image processing. This direction approach has illustrated the promising results in the terms of accuracy and computational time. This paper presents a new approach, which combines a pre-processing technique of image filtering and vision transformer (ViT) learning for the problem of plant insect pests and diseases recognition. The proposed solution involves some stages: neural network-based image filtering, then passes results through a ViT module to extract feature map, and then fed to multiple head network for classification. The proposed method applies image filtering pre-processing to highlight features before passing results to the ViT processing stage instead of using ViT from raw input images. Furthermore, element-wise multiplication in the frequency domain reduces processing time instead of using convolutional processing in the spatial domain. Experimental results demonstrate that applying filtering preprocessing does not significantly increase the number of learning parameters and training time compared to using ViT directly and it leverages to improve accuracy to compare to well-known models based on deep CNN. The research results also illustrated that the ViT solution and the proposed method are reached more accurate than CNN-based deep learning methods.

In recent years, pedestrian detection and tracking have significant progress in both performance and latency. However, detecting and tracking pedestrian human-body in highly crowded environments is a complicated task in the computer vision field because pedestrians are partly or fully occluded by each other. That needs much human effort for annotation works and complex trackers to identify invisible pedestrians in spatial and temporal domains. To alleviate the aforementioned problems, previous methods tried to detect and track visible parts of pedestrians (e.g., heads, pedestrian visible-region), which achieved remarkable performances and can enlarge the scalability of tracking models and data sizes. Inspired by this purpose, this paper proposes simple but effective methods to detect and track pedestrian heads in crowded scenes, called PHDTT (Pedestrian Head Detection and Tracking with Transformer). Firstly, powerful encoder-decoder Transformer networks are integrated into the tracker, which learns relations between object queries and image global features to reason about detection results in each frame, and also matches object queries and track objects between adjacent frames to perform data association instead of further motion predictions, IoU-based methods, and Re-ID based methods. Both components are formed into single end-to-end networks that simplify the tracker to be more efficient and effective. Secondly, the proposed Transformer-based tracker is conducted and evaluated on the challenging benchmark dataset CroHD. Without bells and whistles, PHDTT achieves 60.6 MOTA, which outperforms the recent methods by a large margin. Testing videos are available at https://bit.ly/3eOPQ2d .

This paper presents a deep learning model to address the problem of recognition of animals and plants. The context of this work is to make an effort in protection of rare species that are seriously faced to the risk of extinction in Vietnam such as Panthera pardus, Dalbergia cochinchinensis, Macaca mulatta. The proposed approach exploits the advanced learning ability of convolutional neural networks and Inception residual structures to design a lightweight model for classification task. We also apply the transfer learning technique to fine-tune the two state-of-the-art methods, MobileNetV2 and InceptionV3, specific to our own dataset. Experimental results demonstrate the superiority of our object predictor (e.g., 95.8% accuracy) in comparison with other methods. In addition, the proposed model works very efficiently with the inference speed of around 113 FPS on a CPU machine, enabling it for deployment on mobile environment.

EfficientNet is an efficient architecture that is as small as possible while still having acceptable performance. On the ImageNet challenge, with a much fewer parameter calculation load, it could have its position among the state-of-the-arts. The base module of the original EfficientNet is Mobile Inverted Bottleneck Convolution which has an expansion convolution to have a much bigger number of channels for depthwise convolution. That’s why it is called inverted bottleneck. In this paper, we introduce a new module called Mobile Equal channel Convolution blocks which has the same number of channels for all convolution layers inside to make the module more balance. Moreover, we also adjust a little bit the EfficientNet architecture to make the new variant can have a similar number of parameters to the original. The experiments on ImageNet dataset show that the new variant can have higher accuracy with a lower number of parameters.

Many current automatic license plate (LP) recognition systems are designed to handle a fixed form of LPs. In the present work, we develop an effective system using deep convolutional neuron network (CNN) that can process LPs with different layouts (e.g., variable character lengths, diverse colors, square-like and rectangular shapes). Firstly, we make an attempt of gathering a sufficient large and diverse Vietnamese LP dataset and manually creating the annotations for images. Secondly, a CNN model is derived to detect the LPs in images and predict the LP’s shape (i.e., one-row or two-row form). Thirdly, we design an efficient and unified CNN model to predict the characters of an input LP image patch. The proposed system has been extensively validated on two datasets (Vietnamese and Chinese LPs), demonstrating promising accuracy (e.g., 95% – 99%) and real-time CPU inference in comparison with the state-of-the-art approaches.

Melanoma is the deadliest of many different types of skin cancer. Clinical screening is followed by dermoscopic analysis and histopathological examination in the diagnosis of melanoma. Melanoma is a type of skin cancer that is highly curable if caught early. A visual examination of the affected area of the skin is the first step in melanoma skin cancer diagnosis. Dermatologists use a high-speed camera to take dermatoscopic images of skin lesions, which have an accuracy of 65–80% in melanoma diagnosis without any additional technical support. This research shows how to classify skin cancer using skin lesion photos using an automated classification approach based on image processing techniques. By studying images of skin lesions, the classification system will be able to determine whether or not a patient has melanoma. The contribution of this paper includes testing many different backbones and input sizes on the CNN models to evaluate the accuracy of the model on the siim-isic dataset. The overall prediction rate of melanoma diagnosis was raised to 82–86% on Sensitivity.

Artificial intelligence (AI) based medical image recognition plays an role in assisting disease diagnosis systems. Recent research on deep learning techniques has provided fast and powerful approaches for image analysis and classification for pathologists in their diagnostic tasks. This paper presents a new approach to enhance the efficiency of the skin disease classification task. The method consists of three major parts as follows (1) Extracting skin damaged regions (region of interesting –ROI) based on the semantic segmentation technique, (2) morphological processing for refining the results of ROI extraction, (3) disease classification based on the deep convolution neural network (DCNN). In most of the previous approaches, disease diagnosis systems use image samples, which are taken from medical devices, producing very small, damaged regions in full images. In our approach, the segmentation task utilizes to extract only the damaged skin regions for disease diagnosis. Thusly, the advantage of the proposed method supports reducing the important data space and focuses on ROIs of the disease features for improving efficient diagnosis systems. The proposed method has been evaluated on the benchmark dataset of ISIC2018, which is already available online for both training and validation data. To ensure objectivity, some well-known backbones are used for feature extraction task such as DenseNet, MobileNet, EfficientNet models. The experimental results show that, by using the same feature extraction backbones, the proposed method outperforms the standard methods on some performed metrics (Recall, Accuracy, Precision, Specificity, and F1) with DenseNet (3.11%, 2.04%, 3.37%. 2.91%, and 3.49%), EfficientNet (1.04%, 0.71%, 0.05%, 0.77% and 1.10%), and MobileNet (4.15%, 1.66%, 6.10%, 1.34%, and 5.03%), respectively.

Rice disease prediction plays important task for automated rice disease recognition systems. Feature extraction and classification based on deep learning play important tasks in vision-based diseases recognition. The advancement of deep convolutional neural network using mage data illustrates the approach for identification of rice diseases using deep features with the expectation of high returns. Instead of fine-tuning task which concerns estimation of internal parameters of a model to adjust precisely with certain observations. This paper focuses on extrinsic parameters for model training, which utilities for improving precise of recognition system. Some pretrain models AlexNet, ResNet101 were implemented as mainstream of the convolutional neural network (CNN) architecture. Our approach directly estimates locations of features based on deep learning classification for of rice leaf diseases recognition. There are four kinds of the rice diseases investigated, such as rice blast, bacterial leaf blight, alum poisoning, and leaf folder. A large dataset resolution images from real scenes in the farm were collected for training and evaluation. In this study, the augmentation of image also applied for evaluation input images. The output prediction results of set samples are used for voting final decision. The experimental results show that the proposed approach with hyperparameters customization and data augmentation outperforms.

Melanoma, one of the most dangerous types of skin cancer, results in a very high mortality rate. Early detection and resection are two key points for a successful cure. Recent researches have used artificial intelligence to classify melanoma and nevus and to compare the assessment of these algorithms to that of dermatologists. However, training neural networks on an imbalanced dataset leads to imbalanced performance, the specificity is very high but the sensitivity is very low. This study proposes a method for improving melanoma prediction on an imbalanced dataset by reconstructed appropriate CNN architecture and optimized algorithms. The contributions involve three key features as custom loss function, custom mini-batch logic, and reformed fully connected layers. In the experiment, the training dataset is kept up to date including 17,302 images of melanoma and nevus which is the largest dataset by far. The model performance is compared to that of 157 dermatologists from 12 university hospitals in Germany based on the same dataset. The experimental results prove that our proposed approach outperforms all 157 dermatologists and achieves higher performance than the state-of-the-art approach with area under the curve of 94.4%, sensitivity of 85.0%, and specificity of 95.0%. Moreover, using the best threshold shows the most balanced measure compare to other researches, and is promisingly application to medical diagnosis, with sensitivity of 90.0% and specificity of 93.8%. To foster further research and allow for replicability, we made the source code and data splits of all our experiments publicly available.

Breast cancer subtypes, which play a significant role in breast cancer prognosis and targeted therapy selection, can be identified with gene expression profiling. It is also beneficial for personalized treatment to know bio-markers that impact the development of cancer cells from studying gene expression. Therefore, this study uses recursive feature elimination, support vector machine classifier with grid search cross-validation to prognosticate breast cancer subtypes and propose the cancer-related bio-markers. We experiment with 2682 samples of gene expression data collected from two different sources. Using the same dataset with the state-of-the-art solution, we achieve an accuracy of 89.40% and improve 5.44% accuracy. Besides, our solution suggests 16 bio-markers associated with cancer that have supporting evidence from the literature and propose 11 new genes potential for future research.

Nowadays, computer-aided diagnosis (CAD) systems play an important role in supporting doctors and improving the performance of thyroid nodule detection which able to reduce the physician’s examination burden. In addition, medical data have certain characteristics that make their application very challenging on the existing system. In this paper, we propose an ultrasound image-based of thyroid nodule detection method using state-of-the-art object detectors. We employed radiologist knowledge from clinical studies such as size and aspect ratio distributions of real nodules to generate better region proposals and improve the detection accuracy. Additionally, we contribute a new Vietnamese thyroid nodule dataset, which is collected from Vietnamese Hospitals and validated by experienced radiologists. With this dataset, we expect to provide an essential medical imaging research resource for CAD development and validation. Through experimental results, our detection network FasterIncResnet** that was optimized with radiologist’s knowledge yielded the highest performance compared with original-based methods. Our optimized detection sensitivity, specificity, and AUC were 0.87, 0.86, and 0.89, respectively, on the Vietnamese dataset. The overall performance improves ~2.4% compared with the original FasterIncResnet method. The result proved the effectiveness of the proposed method on a new Vietnamese dataset in order to build an accurate system for thyroid nodule detection.

Today, several studies have been concretized in the areas of robotics, self-driving cars, intelligent assistance systems, and so on. Developing an increasingly optimal neural network in terms of accuracy and processing speed for resource-limited systems has become a major research trend. Some research orientations include focusing on developing solutions to optimize machine learning models and learning parameters. In this study, we investigated an optimization solution for learning hyperparameters of adaptive learning systems for improving object recognition accuracy. The proposed method was developed from a framework searching a set of learning hyperparameters based on the evaluation of the previous CNN model with the collected dataset during the movement of advanced driver assistance systems (ADAS) equipment. The proposed solution consists of some major steps in a loop of adaptive learning system, such as (1) training an initial recognition model, (2) locating and receiving image data of different cases of the object during ADAS movement based on object tracking process, (3) finding optimal hyperparameters on the found dataset based on the previous recognition model, and (4) using the trained recognition model to update the current recognition model. The experimental results proved that the trained recognition model was capable of being more intelligent and displayed more diverse recognition than the previous model. The updated task for the recognition model was continuously repeated throughout the ADAS life. This approach supports and enables the recognition system to be self-adaptive and more intelligent in real life settings without manually processing.

Research on machine vision and some advanced sensors is important for improving the existing applications in product inspections, automation, intelligent transportation systems, human life assistance, and other intelligent systems. In the field of robotic systems, visual odometry (VO) is the process of analyzing the associated camera images to discover the location and orientation of a mobile robot. VO is the process of recovering the related position and orientation of a robot by analyzing the associated vision devices. It aims to recover the parameters of the equivalent odometry data using sequential camera images to estimate the motion of travel by the robot. In structure from motion technology, the general idea is that a camera captures the sequential images of the world scene in different positions. The core idea is finding the relations of images in order to extract the camera position and thus, the robot location.

Human pose estimation is a fundamental research topic in computer vision. This topic has been largely improved recently thanks to the development of the convolution neural network. This paper introduces an efficient human pose estimator based on Mask RCNN, a member of RCNN family. It uses MobileNetV3 as backbone and replaces the vanilla convolutions with the proposed expanded depthwise separable convolutions to reduce the model size, FLOPs and inference time. The model can run in realtime speed at 25 FPS with acceptable scores.