MKMahmoud I. Khalil

Mahmoud I. Khalil

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Mahmoud I. Khalil has a total of 60 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Causal Reversible Color Transform

    by Ayman W. Mohsen, Mahmoud I. Khalil, Hazem M. Abbas

    Lossless image compression algorithms perform cross-channel decorrelation transformations, which are called reversible color transforms (RCTs), usually using integer lifting steps, which either require bit width inflated by one bit for the chroma channels or the use modular arithmetic that introduces wraparound artifacts, resulting from integer overflows. In this work we propose a fundamental technique that solves these problems entirely, by fusing the RCT with spatial prediction and adding causality to color domain, by predicting image data in a permuted RGB space. This method entirely avoids overflows, maintains reversibility, produces residuals with the same word size as the input, supports any kind of scan-line spatial predictors, and can be generalized to any number of channels. The proposed transformation is even guaranteed to improve coding efficiency, compared to all known fixed RCTs, without increasing computational cost or memory usage.

  • L1-Based Prediction for Lossless Image Compression

    by Ayman W. Mohsen, M. Watheq El-Kharashi, Mahmoud I. Khalil, Hazem M. Abbas

    In this work, we present a lightweight and highly efficient prediction scheme for lossless image compression based on online gradient descent with L1 loss. Our predictor maintains adaptive weights for a group of estimators using only the basic integer operations (multiply, add, and shift). Thus, the entire codec can be implemented using integer SIMD instructions in an interleaved manner, making it suitable for streaming, low-power, Internet of Things (IoT), and embedded applications. We present the technique in a fully operational codec achieving 2.863 bits per dimension (BPD) on the DIV2K dataset with symmetric encoding and decoding speeds exceeding 200 MB/s single-threaded on a modest, old laptop CPU: Intel i5-1145G7. Compared to the state-of-the-art libjxl 0.11.1 (the latest as of writing time), our encoder and decoder are faster by over 70 and 13 times, respectively, at comparable compression efficiency, single threaded.


  • Studying the Effect of Preprocessing on Simultaneous Localization and Mapping in Low-Light Conditions

    by Mohamed Mohamed, Mahmoud Khalil, Cherif Salama

    Low light conditions make computer vision practices difficult. Although the current simultaneous localization and mapping methods are mature, it is difficult to have accurate maps with challenging low light conditions like the ones encountered when dealing with underwater images. This paper studies the effect of using preprocessing to improve the performance and accuracy of simultaneous localization and mapping that focuses on low-light scenes, especially those taken underwater. More specifically, the paper compares the effect of classical and deep learning preprocessing approaches. After hyperparameter tuning, the classical contrast limited adaptive histogram equalization approach is found to achieve the best results with a 20.74% increase in accuracy on the Aqualoc underwater dataset.

  • A Genetic Algorithm Approach to Automate Architecture Design for Acoustic Scene Classification

    by Noha W. Hasan, Ali S. Saudi, Mahmoud I. Khalil, Hazem M. Abbas

    Convolutional neural networks (CNNs) have been widely used with remarkable success in the acoustic scene classification (ASC) task. However, the performance of these CNNs highly relies on their architectures, requiring a lot of effort and expertise to design CNNs suitable for the investigated problem. In this work, we propose an efficient genetic algorithm (GA) that aims to find optimized CNN architectures for the ASC task. The proposed algorithm uses frequency-dimension splitting of the input spectrograms in order to explore the architecture search space in sub-CNN models in addition to classical single-path CNNs. Specifically, this algorithm aims to find the best number of sub-CNNs in addition to their architectures to better capture the distinct features of the input spectrograms. The proposed GA is specifically designed for sound classification to suit the ASC task than many other GAs that optimize conventional single-path CNN architectures. Experimental results on three benchmark datasets demonstrate the effectiveness of the proposed method. Specifically, the proposed algorithm has achieved around 17.8%, 16%, and 17.2%, relative improvement in accuracy with respect to the baseline systems on the development datasets of DCASE2018-Task1A, DCASE2019-Task1A, and DCASE2020-Task1A, respectively.


  • Open Access

    Amharic Abstractive Text Summarization

    by Amr M. Zaki, Mahmoud I. Khalil, Hazem M. Abbas

    Text Summarization is the task of condensing long text into just a handful of sentences. Many approaches have been proposed for this task, some of the very first were building statistical models (Extractive Methods) capable of selecting important words and copying them to the output, however these models lacked the ability to paraphrase sentences, as they simply select important words without actually understanding their contexts nor understanding their meaning, here comes the use of Deep Learning based architectures (Abstractive Methods), which effectively tries to understand the meaning of sentences to build meaningful summaries. In this work we discuss one of these new novel approaches which combines curriculum learning with Deep Learning, this model is called Scheduled Sampling. We apply this work to one of the most widely spoken African languages which is the Amharic Language, as we try to enrich the African NLP community with top-notch Deep Learning architectures.


  • Open Access

    Viral Sequence Identification in Metagenomes using Natural Language Processing Techniques

    by Aly O. Abdelkareem, Mahmoud I. Khalil, Ali H. A. Elbehery, Hazem M. Abbas

    Viral reads identification is one of the important steps in metagenomic data analysis. It shows up the diversity of the microbial communities and the functional characteristics of microorganisms. There are various tools that can identify viral reads in mixed metagenomic data using similarity and statistical tools. However, the lack of available genome diversity is a serious limitation to the existing techniques. In this work, we applied natural language processing approaches for document classification in analyzing metagenomic sequences. Text featurization is presented by treating DNA similar to natural language. These techniques reveal the importance of using the text feature extraction pipeline in sequence identification by transforming DNA base pairs into a set of characters with a term frequency and inverse document frequency techniques. Various machine learning classification algorithms are applied to viral identification tasks such as logistic regression and multi-layer perceptron. Moreover, we compared classical machine learning algorithms with VirFinder and VirNet, our deep attention model for viral reads identification on generated fragments of viruses and bacteria for benchmarking viral reads identification tools. Then, as a verification of our tool, It was applied to a simulated microbiome and virome data for tool verification and real metagenomic data of Roche 454 and Illumina for a case study.


  • Emotion Recognition In Videos For Low-Memory Systems Using Deep-Learning

    by Ahmed F. Hagar, Hazem M. Abbas, Mahmoud I. Khalil

    This paper explores a deep learning model for emotion recognition in videos, suitable for systems with limited memory like robots and embedded-systems. The proposed model is a Mini-xception+LSTM architecure with around 80k parameters. This model got a classification accuracy of 93% in dinstinction between Anger and Amusement emotions using the BioVidEmo dataset, compared to 70% accuracy that a recent work got for the same two emotions, and got 86 % and 90 % classification accuracy using the CK+dataset for seven and six emotions, respectively.