Pre-trained Convolutional Network Research Articles

Towards Detecting Dementia via Deep Learning

Dementia is a brain disorder that causes loss of memory leading to disruption in the normal course of life of an individual. It is emerging as a global health problem in adults with age 65 years or above. Early diagnosis of dementia has gone forth as a key research zone with the aim of early identification for hindering the advancement. Deep learning provides path-breaking applications in medical imaging. This study provides a detailed summary of different implementation approaches of deep learning for detecting the disease. Transfer learning for multi-class classification has also been explored for detecting dementia. The pre-trained convolutional network, AlexNet is used with 3 optimizers, SGDM, ADAM, RMSProp. A Dataset of 60 MRI images is taken from the OASIS dataset. Accuracy of the methods has been compared and the best parameters including classifier, learning rate, and a batch size of the model have been identified. SGDM classifier with a learning rate 10-4 and a mini-batch size of 10 have shown the best performance in a reasonable time.

Сравнительные результаты точности классификации радиолокационных изображений объектов набора MSTAR сверточными нейронными сетями с различными архитектурами.

The results of MSTAR objects ten-classes classification using a VGG-type deep convolutional neural network with eight convolutional layers are presented. The maximum accuracy achieved by the network was 97.91%. In addition, the results of the MobileNetV1, Xception, InceptionV3, ResNet50, InceptionResNetV2, DenseNet121 networks, prepared using the transfer learning technique, are presented. It is shown that in the problem under consideration, the use of the listed pretrained convolutional networks did not improve the classification accuracy, which ranged from 93.79% to 97.36%. It has been established that even visually unobservable local features of the terrain background near each type of object are capable of providing a classification accuracy of about 51% (and not the expected 10% for a ten-alternative classification) even in the absence of object and their shadows. The procedure for preparing training data is described, which ensures the elimination of the influence of the terrain background on the result of neural network classification.

Deep learning towards detecting dementia

Dementia is a brain disorder that causes loss of memory leading to disruption in the normal course of life of an individual. It is emerging as a global health problem in adults with age 65 years or above. Early diagnosis of dementia has gone forth as a key research zone with the aim of early identification for hindering the advancement. Deep learning provides path-breaking applications in medical imaging. This study provides a detailed summary of different implementation approaches of deep learning for detecting the disease. Transfer learning for multi-class classification has also been explored for detecting dementia. The pre-trained convolutional network, AlexNet is used with 3 optimizers, SGDM, ADAM, RMSProp. A Dataset of 60 MRI images is taken from the OASIS dataset. Accuracy of the methods has been compared and the best parameters including classifier, learning rate, and a batch size of the model have been identified. SGDM classifier with a learning rate 10-4 and a mini-batch size of 10 have shown the best performance in a reasonable time.

Deep learning towards detecting dementia

Dementia is a brain disorder that causes loss of memory leading to disruption in the normal course of life of an individual. It is emerging as a global health problem in adults with age 65 years or above. Early diagnosis of dementia has gone forth as a key research zone with the aim of early identification for hindering the advancement. Deep learning provides path-breaking applications in medical imaging. This study provides a detailed summary of different implementation approaches of deep learning for detecting the disease. Transfer learning for multi-class classification has also been explored for detecting dementia. The pre-trained convolutional network, AlexNet is used with 3 optimizers, SGDM, ADAM, RMSProp. A Dataset of 60 MRI images is taken from the OASIS dataset. Accuracy of the methods has been compared and the best parameters including classifier, learning rate, and a batch size of the model have been identified. SGDM classifier with a learning rate 10-4 and a mini-batch size of 10 have shown the best performance in a reasonable time.

Bacterial Disease Detection for Pepper Plant by Utilizing Deep Features Acquired from DarkNet-19 CNN Model

In recent years, computer-aided agriculture applications have been developing rapidly as a prominent research area. In parallel with the developments in technology, the use of automatic systems, sensor fusion, the internet of things, and artificial intelligence-based systems is becoming widespread in agriculture. The use of these systems allows for safer, faster, and more cost-effective operations based on human factors in agricultural applications. Among these applications, there are artificial intelligence applications developed based on image processing and machine learning. Plant disease detection systems are also among these artificial intelligence studies. Within the scope of this study: I. It has been ensured that the leaf images of the pepper plant have been segmented and their features have been extracted from the pre-trained convolutional neural network. II. These obtained features have been classified through the classifier methods in order to detect bacterial disease. In the study, a total of 2475 images of pepper leaves with 1478 healthy and 997 bacterial diseases, which are among the PlantVillage data sets, have been used. To extract the features, the DarkNet-19 network model has been used as a pre-trained convolutional network. The SoftMax classifier in the last layer of the convolutional network model has been removed from the network and SVM, KNN, and Decision-Tree-based classifiers are used instead of it. According to the results, the level of performance achieved using the DarkNet-19 network and SVM classifier is quite satisfactory.

Using Non-Additive Entropy to Enhance Convolutional Neural Features for Texture Recognition.

Here we present a study on the use of non-additive entropy to improve the performance of convolutional neural networks for texture description. More precisely, we introduce the use of a local transform that associates each pixel with a measure of local entropy and use such alternative representation as the input to a pretrained convolutional network that performs feature extraction. We compare the performance of our approach in texture recognition over well-established benchmark databases and on a practical task of identifying Brazilian plant species based on the scanned image of the leaf surface. In both cases, our method achieved interesting performance, outperforming several methods from the state-of-the-art in texture analysis. Among the interesting results we have an accuracy of 84.4% in the classification of KTH-TIPS-2b database and 77.7% in FMD. In the identification of plant species we also achieve a promising accuracy of 88.5%. Considering the challenges posed by these tasks and results of other approaches in the literature, our method managed to demonstrate the potential of computing deep learning features over an entropy representation.

Micro-optics assembly for fast axis collimation by means of convolutional neural network.

The manufacture of high-power diode laser systems highly depends on the quality of the collimation of semiconductor laser diodes. The collimation process is conducted by precise alignment of micro-optics, where the first and the most critical step is the placement of the fast axis collimator (FAC) lens. The sub-micron positioning of the FAC lens is conventionally conducted with an active alignment strategy by monitoring changes in the laser beam profile with an external sensitive camera. The optimization of the beam profile characteristics is controlled by a specifically programmed motorized robotic aligner. However, active alignment, a very accurate method, often results in a higher cycle time than the passive approach, where the lens is placed in a pre-measured position. Here, we developed a new active approach, without closed loop control to position the micro-optics, that relies on the use of a pretrained convolutional network (CNN). We trained and evaluated three CNNs that can predict the optimal lens position using the single camera image of the laser beam. We predict that implementation of the best performing CNN-based model would lead to a decrease in alignment time from tens of seconds to hundreds of milliseconds and will be broadly applied in a high-volume manufacturing environment.

Application and comparison of feature-based classification models for multistable impact motions of percussive drilling

Dynamics of the bit-rock interaction under percussive drilling often encounter multistability that produces coexisting impact motions for a wide range of drilling conditions. Some of them may be detrimental to its performance as it cuts through the inhomogeneous rock layers. A necessary mitigation is the ability to distinguish between coexisting impact motions in order to maintain a high-performance drilling. For this purpose, dynamical responses of a vibro-impact system mimicking the bit-rock interaction of percussive drilling were explored in this study by using machine learning techniques. As a fundamental approach of improving machine learning, hand-crafted and automatic feature extractions were carried out. Results from the simulated data show that extracting appropriate features and using a suitable network are essential for characterising the vibro-impact motions. Extracting statistical, histogram of gradients, continuous wavelet transform and pre-trained convolutional network features are effective and less computationally intensive. With their high accuracies, they become the first point of consideration when designing the classification model for multistable vibro-impact motions of percussive drilling.

Long short-term memory improved Siamese network for robust target tracking

Visual target tracking is an important function in real-time video monitoring application, whose performance determines the implementation of many advanced tasks. At present, Siamese-network trackers based on template matching show great potential. It has the advantage of balance between accuracy and speed, due to the pre-trained convolutional network to extract deep features for target representation and off-line tracking of each frame. During tracking, however, the target template feature is only obtained from the first frame of the video in the existing algorithms. The tracking performance is completely depending on the framework of template matching, resulting in the independence of frames and ignoring the feature of inter-frame connection of video sequence. Therefore, the existing algorithms do not perform well in the face of large deformation and severe occlusion. We propose a long short-term memory (LSTM) improved Siamese network (LSiam) model, which takes advantages of both time-domain regression capability of the LSTM and the balanced ability in tracking accuracy and speed of Siamese network. It focus on the temporal and spatial correlation information between video sequences to improve the traditional Siamese-network trackers with an LSTM prediction module. In addition, an improved template updating module is constructed to combine the original template with the changed appearance. The proposed model is verified in two types of difficult scenarios: deformation challenge and occlusion challenge. Experimental results show that our proposed approach can get better performance in terms of tracking accuracy.

Classification of Facial Expression of Post-Surgical Pain in Children

There are certain difficulties in differentiating between children's facial expression related to pain and other stimuli. In addition, the limited communication ability of children in the preverbal stage leads to misdiagnosis when the child feels pain, for example, post-surgical conditions. In this article, a classification approach of facial expression of child pain is presented based on models of pre-trained convolutional neuronal networks from the study carried out in a Colombian hospital of level 4 (Hospital Universitario San Vicente Fundación), in the recovery areas of child surgery services. AlexNet and VGG (16, 19 and Face) networks are evaluated in the own dataset using the FLACC scale and their performances are compared in three experiments. The results show that the VGG-19 model achieves the best performance (92.9%) compared to the other networks. The effectiveness of the model and transfer learning for the classification of facial expression of child pain shows a promising solution for the assessment of post-surgical pain.

Convolutional neuronal networks combined with X-ray phase-contrast imaging for a fast and observer-independent discrimination of cartilage and liver diseases stages

We applied transfer learning using Convolutional Neuronal Networks to high resolution X-ray phase contrast computed tomography datasets and tested the potential of the systems to accurately classify Computed Tomography images of different stages of two diseases, i.e. osteoarthritis and liver fibrosis. The purpose is to identify a time-effective and observer-independent methodology to identify pathological conditions. Propagation-based X-ray phase contrast imaging WAS used with polychromatic X-rays to obtain a 3D visualization of 4 human cartilage plugs and 6 rat liver samples with a voxel size of 0.7 × 0.7 × 0.7 µm3 and 2.2 × 2.2 × 2.2 µm3, respectively. Images with a size of 224 × 224 pixels are used to train three pre-trained convolutional neuronal networks for data classification, which are the VGG16, the Inception V3, and the Xception networks. We evaluated the performance of the three systems in terms of classification accuracy and studied the effect of the variation of the number of inputs, training images and of iterations. The VGG16 network provides the highest classification accuracy when the training and the validation-test of the network are performed using data from the same samples for both the cartilage (99.8%) and the liver (95.5%) datasets. The Inception V3 and Xception networks achieve an accuracy of 84.7% (43.1%) and of 72.6% (53.7%), respectively, for the cartilage (liver) images. By using data from different samples for the training and validation-test processes, the Xception network provided the highest test accuracy for the cartilage dataset (75.7%), while for the liver dataset the VGG16 network gave the best results (75.4%). By using convolutional neuronal networks we show that it is possible to classify large datasets of biomedical images in less than 25 min on a 8 CPU processor machine providing a precise, robust, fast and observer-independent method for the discrimination/classification of different stages of osteoarthritis and liver diseases.

ALGORITHM OF SKELETON-BASED STATIC HAND GESTURE RECOGNITION

We present the monocamera approach of static hand gestures recognition based on skeletonization. The problem of creating skeleton of the human’s hand, as well as body, became solvable a few years ago after inventing so called convolutional pose machines – the novel architecture of artificial neural network. Our solution uses such kind of pretrained convolutional artificial network for extracting hand joints keypoints with further skeleton reconstruction. In this work we also propose special skeleton descriptor with proving its stability and distinguishability in terms of classification. We considered a few widespread machine learning algorithms to build and verify different classifiers. The quality of the classifier’s recognition is estimated using the wellknown Accuracy metric, which identified that classical SVM (Support Vector Machines) with radial basis kernel gives the best results. The testing of the whole system was conducted using public databases containing about 3000 of test images for more than 10 types of gestures. The results of a comparative analysis of the proposed system with existing approaches are demonstrated. It is shown that our gesture recognition system provides better quality in comparison with existing solutions. The performance of the proposed system was estimated for two configurations of standard personal computer: with CPU (Central Processing Unit) only and with GPU (Graphics Processing Unit) in addition where the latest one provides realtime processing with up to 60 frames per second. Thus we demonstrate that the proposed approach can find an application in the practice.

Optimization of Feature Loss for Image Enhancement

Image-transformation problem is a problem in which an input image is transformed to an output image. In most of the recent methods, a feed-forward neural network is defined which utilizes per-pixel loss between the output image and the ground-truth image. In this paper we have showcased that high-quality images can be generated by defining a feature-loss function which is based on high-level perceptual features extracted from pre-trained convolutional networks. We have combined both the approaches that have been formerly mentioned and have proposed a feature-loss function for training a feed-forward neural network capable of image transformation tasks. We have compared out method with that of an optimization based approach, similar to the one utilized in Generative Adversarial Networks (GANs) and our method produced visually appealing results whilst fully capturing the intricate details of the object in the image.

Towards weeds identification assistance through transfer learning

Reducing the use of pesticides through selective spraying is an important component towards a more sustainable computer-assisted agriculture. Weed identification at early growth stage contributes to reduced herbicide rates. However, while computer vision alongside deep learning have overcome the performance of approaches that use hand-crafted features, there are still some open challenges in the development of a reliable automatic plant identification system. These type of systems have to take into account different sources of variability, such as growth stages and soil conditions, with the added constraint of the limited size of usual datasets. This study proposes a novel crop/weed identification system that relies on a combination of fine-tuning pre-trained convolutional networks (Xception, Inception-Resnet, VGNets, Mobilenet and Densenet) with the “traditional” machine learning classifiers (Support Vector Machines, XGBoost and Logistic Regression) trained with the previously deep extracted features. The aim of this approach was to avoid overfitting and to obtain a robust and consistent performance. To evaluate this approach, an open access dataset of two crop [tomato (Solanum lycopersicum L.) and cotton (Gossypium hirsutum L.)] and two weed species [black nightshade (Solanum nigrum L.) and velvetleaf (Abutilon theophrasti Medik.)] was generated. The pictures were taken by different production sites across Greece under natural variable light conditions from RGB cameras. The results revealed that a combination of fine-tuned Densenet and Support Vector Machine achieved a micro F1 score of 99.29% with a very low performance difference between train and test sets. Other evaluated approaches also obtained repeatedly more than 95% F1 score. Additionally, our results analysis provides some heuristics for designing transfer-learning based systems to avoid overfitting without decreasing performance.

A Multi-Scale Approach for Remote Sensing Scene Classification Based on Feature Maps Selection and Region Representation

Scene classification is one of the bases for automatic remote sensing image interpretation. Recently, deep convolutional neural networks have presented promising performance in high-resolution remote sensing scene classification research. In general, most researchers directly use raw deep features extracted from the convolutional networks to classify scenes. However, this strategy only considers single scale features, which cannot describe both the local and global features of images. In fact, the dissimilarity of scene targets in the same category may result in convolutional features being unable to classify them into the same category. Besides, the similarity of the global features in different categories may also lead to failure of fully connected layer features to distinguish them. To address these issues, we propose a scene classification method based on multi-scale deep feature representation (MDFR), which mainly includes two contributions: (1) region-based features selection and representation; and (2) multi-scale features fusion. Initially, the proposed method filters the multi-scale deep features extracted from pre-trained convolutional networks. Subsequently, these features are fused via two efficient fusion methods. Our method utilizes the complementarity between local features and global features by effectively exploiting the features of different scales and discarding the redundant information in features. Experimental results on three benchmark high-resolution remote sensing image datasets indicate that the proposed method is comparable to some state-of-the-art algorithms.

Automatic Grading of Ankle Osteoarthritis Based on Takakura Staging System: A Deep Learning- Based Approach

Category:Ankle, Ankle ArthritisIntroduction/Purpose:The Takakura staging system has been used for the stratification in ankle osteoarthritis(OA). Patient’s OA stage is determined by visual examination on the status of talar and distal tibia in anteroposterior ankle radiograph. Clinical decisions about whether to treat conservatively or to treat with operation such as supra-malleolar osteotomy or arthroplasty may depend on this grading system. However, this is not completely reproducible between examiners and it makes a debating situation about different treatment method to a same ankle radiograph. If highly reproducible measurement method may be suggested this debating may no longer need. Therefore, the purpose of this study was to suggest a deep learning-based algorithm that automatically grades ankle osteoarthritis and to present feasibility of the provided automatic grading system.Methods:2529 AP both-ankle radiographs were collected and graded for OA by a radiologist and orthopedic surgeon. We converted Takakura staging system into 3 grades(Grade1: stageI, Grade2: II-IIIa, Grade3: IIIb-IV) according to treatment plan. To confine the region of interest(ROI), a rectangle encompassing ankle portion was automatically generated using an object detection model(YOLOv2). The data oversampling was done to overcome small data and class imbalance. Four pre-trained convolutional networks(One Inception-v3 and three ResNet models) were fine-tuned using augmented data. We tried two different ensemble methods: voting ensemble and gradient boosting. Voting ensemble adjusts the decision through selecting majority votes among trained models. Gradient boosting(XGboost model) trains new classification model to focus on the cases that previous model mis- classified. The evaluation of trained models and ensemble model were performed in terms of average classification accuracy.Gradient-class activation map(CAM) method was utilized to present CAM highlighting the location where highly affected the network for the decision.Results:A total of 3836 original ROIs were obtained and as follows: grade 1, 1382; grade 2, 1927; grade 3, 527. The number of oversampled ROIs was 16398 like follows: grade 1, 5528; grade 2, 7708; grade 3, 3162. The performance of each classifier was ranged 71.0% ˜ 77.3% in terms of average classification accuracy. Ensemble methods yielded average classification accuracies of 78.1% and 79.2% for voting ensemble and XGboost, respectively.Conclusion:Deep learning-based algorithm application for automatic grading of ankle osteoarthritis based on Takakura staging system is feasible. This approach is expected to be applied to various staging system for arthritis assessment through radiographs.

A baseline regularization scheme for transfer learning with convolutional neural networks

In inductive transfer learning, fine-tuning pre-trained convolutional networks substantially outperforms training from scratch. When using fine-tuning, the underlying assumption is that the pre-trained model extracts generic features, which are at least partially relevant for solving the target task, but would be difficult to extract from the limited amount of data available on the target task. However, besides the initialization with the pre-trained model and the early stopping, there is no mechanism in fine-tuning for retaining the features learned on the source task. In this paper, we investigate several regularization schemes that explicitly promote the similarity of the final solution with the initial model. We show the benefit of having an explicit inductive bias towards the initial model. We eventually recommend that the baseline protocol for transfer learning should rely on a simple L2 penalty using the pre-trained model as a reference.

Transfer Learning Assisted Classification and Detection of Alzheimer's Disease Stages Using 3D MRI Scans.

Alzheimer’s disease effects human brain cells and results in dementia. The gradual deterioration of the brain cells results in disability of performing daily routine tasks. The treatment for this disease is still not mature enough. However, its early diagnosis may allow restraining the spread of disease. For early detection of Alzheimer’s through brain Magnetic Resonance Imaging (MRI), an automated detection and classification system needs to be developed that can detect and classify the subject having dementia. These systems also need not only to classify dementia patients but to also identify the four progressing stages of dementia. The proposed system works on an efficient technique of utilizing transfer learning to classify the images by fine-tuning a pre-trained convolutional network, AlexNet. The architecture is trained and tested over the pre-processed segmented (Grey Matter, White Matter, and Cerebral Spinal Fluid) and un-segmented images for both binary and multi-class classification. The performance of the proposed system is evaluated over Open Access Series of Imaging Studies (OASIS) dataset. The algorithm showed promising results by giving the best overall accuracy of 92.85% for multi-class classification of un-segmented images.

Transfer learning approach to map urban slums using high and medium resolution satellite imagery

Slums provide cheaper workforce and informal services which contribute substantially towards GDP. However, such areas, due to the high population density, sub-standard housing and lack of essential services are urban risks. The socio-physical development of such settlements has often been neglected due to poor laws and provisions in urban management and policies. One of the primary reasons for negligence has been the unavailability of slum maps to study the evolution of slums and to actively manage and contain them. Various remote sensing techniques have been utilized to answer the problem but have not produced universal solutions. In recent years, Deep Learning (DL) techniques with remote sensing have been found beneficial in comprehending the underlying structure of physical features present in the satellite imageries. This study deals with one of the Deep Learning techniques which use pre-trained convolutional networks for slum detection in Very High Resolution (VHR) and Medium Resolution (MR) satellite imagery. We created a training dataset which comprises of four classes including slums, built, green and water. We further trained the model to detect these classes in the entire city. Classification performance was evaluated for Very high and Medium Resolution imagery with the help of manually delineated slum boundaries gathered from urban local authorities of Mumbai. The Overall accuracy of 94.2 and 90.2 and kappa of 0.70 and 0.55 is obtained from VHR and MR imagery respectively. We provide a comprehensive technique for the detection of informal settlements which can be tailored and applied to any city to detect various landforms.

Selective clustering for representative paintings selection

Selective classification (or rejection based classification) has been proved useful in many applications. In this paper we describe a selective clustering framework with reject option to carry out large-scale digital arts analysis. With the help of deep learning techniques, we extract content-style features from a pre-trained convolutional network for the paintings. By proposing a rejection mechanism under Bayesian framework, we focus on selecting style-oriented representative paintings of an artist, which is an interesting and challenging cultural heritage application. Two kinds of samples are rejected during the rejection based robust continuous clustering process. Representative paintings are selected during the selective clustering phase. Visual qualitative analysis on small painting set and large scale quantitative experiments on a subset of Wikiart show that the proposed rejection based selective clustering approach outperforms the standard clustering methods.