Srdjan Sladojevic

Deep Neural Networks Based Recognition of Plant Diseases by Leaf Image Classification

The latest generation of convolutional neural networks (CNNs) has achieved impressive results in the field of image classification. This paper is concerned with a new approach to the development of plant disease recognition model, based on leaf image classification, by the use of deep convolutional networks. Novel way of training and the methodology used facilitate a quick and easy system implementation in practice. The developed model is able to recognize 13 different types of plant diseases out of healthy leaves, with the ability to distinguish plant leaves from their surroundings. According to our knowledge, this method for plant disease recognition has been proposed for the first time. All essential steps required for implementing this disease recognition model are fully described throughout the paper, starting from gathering images in order to create a database, assessed by agricultural experts. Caffe, a deep learning framework developed by Berkley Vision and Learning Centre, was used to perform the deep CNN training. The experimental results on the developed model achieved precision between 91% and 98%, for separate class tests, on average 96.3%.

Logging real packet reception patterns for end-to-end quality of experience assessment in wireless multimedia transmission

While a fairly large number of databases exist that provide impaired video sequences designed for the development and evaluation of Quality of Experience (QoE) approaches, the impairments due to transmission errors resulting in packet loss in these databases are based on simulation of small number of scenarios and not representative of real transmission scenarios that arise in end-to-end transmission of multimedia to mobile devices. This paper proposes a solution to this problem in the form of a framework for recording real packet drops as they occur in different situations of wireless multimedia transmission to mobile devices. The logs can be used to generate realistic impairments, as well as design new highly-efficient quality assessment approaches based on monitoring network performance in real time. An Android-based mobile device is used to receive streamed H.264 videos and real Real-time Transport Protocol (RTP) packet reception sequences (traces) are recorded. To evaluate the approach and the impact of different transmission scenarios on perceived quality, a study of the quality experienced by the users in 3 different real wireless transmission scenarios was conducted and the results are presented in the paper, showing that they diverge significantly from the packet-loss sequences usually considered in quality of experience studies.

QoE-aware Rate-Conservative dynamic HTTP streaming over mobile cellular networks

In this paper we investigate the design of DASH stream-switching solution suitable for mobile cellular video streaming consumers. Our goal is to design a solution that reduces the amount of downloaded content while adhering to user preferences regarding received video quality. We firstly present a graphical (trellis-like) representation of a stream-switching process, and then investigate different approaches to search for an optimal path through a “stream-switching graph” using suitably defined utility function. We target our search towards the path that will: i) respect underlying wireless-link rate limitations, ii) maintain segment selection that is efficient in terms of the quality vs the amount of the downloaded data, and iii) reduce disturbing quality variations. The simulation results demonstrate the performance of the proposed QoE-aware Rate-Conservative (QaRC) DASH as compared to the recently proposed baseline stream-switching solution.

Comparison of compression performance of 10-bit vs. 8-bit depth, under H.264 Hi422 profile

H.264 is one of the first video coding standard incorporating coding formats with a bit-depth of above 8 bits. This paper presents the results of compression comparison tests for the H.264 “High 422” profile, between 10-bit and 8-bit sample depths. The simulations were run on five 720p and 1080i high definition sequences. PSNR and SSIM metrics were used to evaluate the objective quality performance of both bit-depths, and with the aim of enabling a fair comparison, both metrics were computed with 10-bit precision, up-scaling the 8-bit decoded sequences to 10-bits. Some works have been published in this field based on the evaluation of commercial 10-bit H.264 implementations. In this work, we carry out a neutral evaluation of H.264 standard performance using the official H.264 Reference Software. Unlike the expected 10-bit coding gains, the results show unnoticeable differences between both sample depths in terms of objective quality, lower 0.1dB for PSNR and 0.002 for SSIM, with a 5% bit rate saving that can be achieved for the luminance component, especially for the 720p format, and negligible quality improvement for U and V chroma components.

Local binary patterns and neural networks for no-reference image and video quality assessment

In the modern world, where multimedia is predicted to form 86% of traffic transmitted over the telecommunication networks in the near future, content providers are looking to shift towards Quality of Experience, rather than Quality of Service in multimedia delivery. Thus, no-reference image quality assessment and the related video quality assessment remaining open research problem, with significant market potential. In this paper we describe a study focused on evaluating the applicability of Local Binary Patterns (LBP) as features and neural networks as estimators for image quality assessment. We focus on blockiness artifacts, as a prominent effect in all block-based coding approaches and the dominant artifact in occurring in videos coded with state-of-the-art video codecs (MPEG-4, H.264, HVEC). In this initial study we show how an LBP-inspired approach, tuned to this particular effect, can be efficiently used to predict the MOS of JPEG coded images. The proposed approach is evaluated on a well-known public database and against widely-used features. The results presented in the paper show that the approach achieves superior performance, which forms a sound basis for future research aimed at video quality assessment and precise blocking artifact detection with sub-frame precision.

Data-driven approach to dynamic visual attention modelling

Visual attention deployment mechanisms allow the Human Visual System to cope with an overwhelming amount of visual data by dedicating most of the processing power to objects of interest. The ability to automatically detect areas of the visual scene that will be attended to by humans is of interest for a large number of applications, from video coding, video quality assessment to scene understanding. Due to this fact, visual saliency (bottom-up attention) models have generated significant scientific interest in recent years. Most recent work in this area deals with dynamic models of attention that deal with moving stimuli (videos) instead of traditionally used still images. Visual saliency models are usually evaluated against ground-truth eye-tracking data collected from human subjects. However, there are precious few recently published approaches that try to learn saliency from eyetracking data and, to the best of our knowledge, no approaches that try to do so when dynamic saliency is concerned. The paper attempts to fill this gap and describes an approach to data-driven dynamic saliency model learning. A framework is proposed that enables the use of eye-tracking data to train an arbitrary machine learning algorithm, using arbitrary features derived from the scene. We evaluate the methodology using features from a state-of-the art dynamic saliency model and show how simple machine learning algorithms can be trained to distinguish between visually salient and non-salient parts of the scene.

Deep Learning for Plant Diseases: Detection and Saliency Map Visualisation

Recently, many researchers have been inspired by the success of deep learning in computer vision to improve the performance of detection systems for plant diseases. Unfortunately, most of these studies did not leverage recent deep architectures and were based essentially on AlexNet, GoogleNet or similar architectures. Moreover, the research did not take advantage of deep learning visualisation methods which qualifies these deep classifiers as black boxes as they are not transparent. In this chapter, we have tested multiple state-of-the-art Convolutional Neural Network (CNN) architectures using three learning strategies on a public dataset for plant diseases classification. These new architectures outperform the state-of-the-art results of plant diseases classification with an accuracy reaching 99.76%. Furthermore, we have proposed the use of saliency maps as a visualisation method to understand and interpret the CNN classification mechanism. This visualisation method increases the transparency of deep learning models and gives more insight into the symptoms of plant diseases.

Suppression of Metal Artefacts in CT Using Virtual Singorams and Corresponding MR Images

Medical imaging is invaluable when it comes to gaining insight into the human body. As is well known, medical images need to deal with artefacts. This article presents a modern procedure for metal artifact reduction in computed tomography, which relies on additional information extracted from corresponding magnetic resonance images. We conducted a simulation study so as to compare the resulting images with those corrected, using the baseline linear interpolation method. The outcome indicates that the proposed method incomparably outperforms the baseline and reduces metal artefacts, improving the quality of images, which can be later used in a clinical setting.

Video Quality Assessment for Mobile Devices on Mobile Devices

Pervasiveness of mobile devices and ubiquitous broadband Internet access have laid foundations for video content to be consumed increasingly on smart phones or tablets. As over 85% of the global consumer traffic by 2016 is estimated to be generated by streaming video content, video quality as perceived by end-users of such devices is becoming an important issue. Most of the studies concerned with Video Quality Assessment (VQA) for mobile devices have been carried out in a carefully controlled environment, thus potentially failing to take into account variables or effects present in real-world conditions. In this paper, we compare the results of traditional approach to VQA for mobile devices to those obtained in real-world conditions by using a physical mobile device, for the same video test-set. Results indicate that a difference in perceived video quality between the two settings exists, thus laying foundations for further research to explain the reasons behind it.

Echocardiographic Parameters as Predictors of In-Hospital Mortality in Patients with Acute Coronary Syndrome Undergoing Percutaneous Coronary Intervention

Different ways have been used to stratify risk in acute coronary syndrome (ACS) patients. The aim of the study was to examine the usefulness of echocardiographic parameters as predictors of in-hospital outcome in patients with ACS after percutaneous coronary intervention (PCI). A data of 2030 patients with diagnosis of ACS hospitalized from December 2008 to December 2011 was used to develop a risk model based on echocardiographic parameters using the binary logistic regression. This model was independently evaluated in validation cohort prospectively (954 patients admitted during 2012). In-hospital mortality in derivation cohort was 7.73%, and 6.28% in validation cohort. Developed model has been designed with 4 independent echocardiographic predictors of in-hospital mortality: left ventricular ejection fraction (LVEF RR = 0.892; 95%CI = 0.854–0.932, P < 0.0005), aortic leaflet separation diameter (AOvs RR = 0.131; 95%CI = 0.027–0.627, P = 0.011), right ventricle diameter (RV RR = 2.675; 95%CI = 1.109–6.448, P = 0.028) and right ventricle systolic pressure (RVSP RR = 1.036; 95%CI = 1.000–1.074, P = 0.048). Model has good prognostic accuracy (AUROC = 0.84) and it retains good (AUROC = 0.78) when testing on the validation cohort. Risks for in-hospital mortality after PCI in ACS patients using echocardiographic measurements could be accurately predicted in contemporary practice. Incorporation of such developed model should facilitate research, clinical decisions, and optimizing treatment strategy in selected high risk ACS patients.