Machine learning (ML) is a subclass of artificial intelligence (AI) that presents systems with the capability of automatically learn and upgrade from the knowledge without programmed explicitly. ML concentrates on computer program development for acce...
Machine learning (ML) is a subclass of artificial intelligence (AI) that presents systems with the capability of automatically learn and upgrade from the knowledge without programmed explicitly. ML concentrates on computer program development for accessing data to utilize it for learning themselves. ML is performing a central role in medical image analysis. Several algorithms based on ML have been implemented in medical imaging for problems like segmentation, classification, and detection. With the broad application of deep learning (DL) methods, there has been a significant improvement in the performance of medical image analysis. DL methods are a set of algorithms in ML that attempt to automatically learn multiple levels of abstraction and representation for assisting to make sense of data. Wireless capsule endoscopy (WCE) is a method in which a patient swallows a camera-embedded pill-shaped device that moves through the gastrointestinal (GI) tract, captures and transmits images to an outside receiver. WCE devices produce over 60,000 images during their course of operation inside the GI tract. These images need to be examined by expert physicians who attempt to identify images that contain inflammation/disease. It can be hectic for a physician to go through such a large number of frames, therefore computer-aided detection methods are considered an efficient alternative. Numerous anomalies that can take place in the GI tract of a human being in which the most vital are tumors, polyp, and ulcers. Moreover, general abnormalities and bleeding inside the GI tract may be the symptoms of these diseases. Recently, there has been a lot of research in the field of WCE and medical image analysis. Segmentation, classification, and detection of the diseases mentioned above are thoroughly investigated using both conventional image processing and DL approaches. A detailed survey of all the attempts for WCE medical image analysis for the detection of tumor, polyp, and ulcer was an open challenge. The motivation behind this thesis is to provide a comprehensive review of the techniques adopted for the detection of tumors, polyps, and ulcers while considering purely WCE source. Additionally, generalized anomalies found in WCE images followed by bleeding/lesion detection are provided to avoid the insight limitations of research. The focal point is to provide a comparative investigation of the current techniques, thereby creating possibilities for future research in this specific domain. Detailing all the recent attempts that have been taken in the work, we have implemented a cascaded DL approach for the joint classification of tumor, polyp, and ulcer. The motivation behind using a cascaded approach is to consider that output generated from WCE is a compressed frame due to the limitation of battery life and storage capacity of WCE device resulting in the degradation of the quality of images. Furthermore, during the transmission of WCE images, channel noise such as an additive white Gaussian noise (AWGN) and compression artifacts can degrade the quality of WCE images. The proposed DL approach employs two deep neural networks (DNNs), denoising convolutional neural network (DnCNN) as a denoiser and CNN model for the joint classification of tumor, polyp, and ulcer found within the GI tract of human body. In addition to WCE, colonoscopy is another way for the examination of the colon/large intestine of a human body. A DL-based detection of the polyp within the colonoscopy images is proposed in this dissertation where a CNN model that utilizes lesser hidden layers resulting the model to be lighter for processing and effective at the same time is employed. We introduce a new activation function i.e., MISH for some hidden layers for improved propagation of information along with ReLU. A generalized intersection over union (GIoU) is adopted as new loss to optimized the non-overlapping bounding box considering the shape and orientation of polyp structure. Detailed performance comparison of the proposed deep CNN model is provided by benchmarking with other method showing efficient results.
The Internet revolution has caused the user’s for easier access to fast multimedia data exchange that compels intellectual property security. The communication of information across the Internet is prevalent quite often, and since the communication channel is not always secure; hence it is essential to protect data alteration to ensure the dependability of the data. An efficient watermarking scheme that uses low frequency components named as modified selective embedding in low Frequency (M-SELF) employing Fast Fourier Transform (FFT) is proposed. A coordinates calculation computes an optimal value of radial implementation for embedding the watermark at the encoder side. The decoder side follows the same process but in an inverse way to extract the watermark. The transmission of sensitive information in the form of a watermarked image is a crucial task to consider. Impairments in the form of noises such as additive white Gaussian noise (AWGN), etc., can corrupt the embedded watermark/bits and can eventually degrade the perceptual quality of an image. To avoid such situations a denoiser approach i.e., DnCNN is implemented in an integrated way for the first time after the encoder i.e., M-SELF technique. Performance evaluation of the proposed integrated M-SELF watermarking technique is done in terms of percentage of bits being retrieved as a success rate at each level of noise and perceived visual quality by using full-reference image quality assessment (IQA) metrics such as peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). Moreover, the performance of the proposed integrated M-SELF watermarking technique is benchmarked and analyzed with other DL approach and conventional filtering techniques.
The recent wave in the practice of multimedia services has highlighted the demand for constant monitoring and supervision of multimedia systems based on the users’ quality of experience (QoE). The fast and vigorous quality of the video’s measurements is essential to maximize the QoE and resource allocation control within networks. Degradation in video stream quality can occur due to compression and transmission through communication channel that error-prone. For the estimation of the perceptual quality of video stream, there are two approaches, i.e., subjective quality assessment (human observers) and objective video quality metrics (VQM). A typical video transmission system comprised of an encoder that encode a video, transmission channel, and decoder which decode the video. Giving to the network condition i.e., bandwidth, the encoder encodes the video stream at specific bitrate. The process of encoding generally results in compression leading into quality degradation of the video. Nowadays, most online video web sites and broadcasters produce video content at high-definition (HD) resolutions. The success of HD contents has led to the development of 4K ultra-high definition (UHD) contents regarded as the future standard in video applications. In perfect conditions, UHD is supposed to accommodate viewers with improved visual experience by a wide field of view (FOV) in both vertical and horizontal directions of the screen. Higher frame rates (HFR) have a direct relationship with the quality of the videos. The perceived quality of the human as a subjective analysis is more accurate when it comes to HFR and high-resolution video contents. The video encoding industry is evolving with each year passing; High Efficiency Video Coder (HEVC) H.265/HEVC, and open-source VP9 encoders providing 50% bandwidth saving when compared to previous encoder i.e., Advanced Video Coding (AVC) H.264. A detailed subjective analysis is conducted for UHD videos at frame rate of 30fps and 60fps that are compressed at five different quantization parameters (QP) to investigate the perceptual quality of the users. This thesis also includes a preliminary work that use three different encoders i.e., H.264, HEVC/H.265, and VP9 at five different QP levels for different frame rates making an extensive subjective analysis reflected as differential mean opinion score (DMOS). Furthermore, the encoding efficiency as the encoding time for each encoder and qualitative performance by employing full-reference (FR) quality metrics are presented. Moreover, a qualitative result for finding a correlation for the subjective quality assessment model and FR quality metrics are also provided.