Search

We propose a two-stage Convolutional Neural Network (CNN) based classification framework for detecting COVID-19 and Community-Acquired Pneumonia (CAP) using the chest Computed Tomography (CT) scan images. In the first stage, an infection - COVID-19 or CAP, is detected using a pre-trained DenseNet architecture. Then, in the second stage, a fine-grained three-way classification is done using EfficientNet architecture. The proposed COVID+CAP-CNN framework achieved a slice-level classification accuracy of over 94% at identifying COVID-19 and CAP. Further, the proposed framework has the potential to be an initial screening tool for differential diagnosis of COVID-19 and CAP, achieving a validation accuracy of over 89.3% at the finer three-way COVID-19, CAP, and healthy classification. Within the IEEE ICASSP 2021 Signal Processing Grand Challenge (SPGC) on COVID-19 Diagnosis, our proposed two-stage classification framework achieved an overall accuracy of 90% and sensitivity of .857, .9, and .942 at distinguishing COVID-19, CAP, and normal individuals respectively, to rank first in the evaluation. Code and model weights are available at https://github.com/shubhamchaudhary2015/ct_covid19_cap_cnn, Comment: Top Ranked Model Paper at the ICASSP 2021 COVID-19 Grand Challenge

In the literature, many 3D-Synthesized Image Quality Assessment (IQA) algorithms are proposed, which are based on predicting the geometric and structural distortions present in the synthesized datasets. With the exponential growth of accurate inpainting algorithms, certain types of distortions, such as Blackholes, has become obsolete. Unfortunately, the existing IQA algorithms are mainly concentrating on efficiently identifying these black holes and subsequently predicting the perceptual quality of 3D synthesized views. The performance of these algorithms is quite weak in the recently proposed IETR dataset. Towards this end, we propose a new completely blind IQA algorithm, which is based on the following key observations: 1. Distortions such as blurriness, blockiness (compression artifacts), and fast fading (object shifting) primarily affect the perceptual quality of 3D-synthesized views. 2. The perceptual characteristics of natural and synthetic synthesized views are quite different; distortions in natural views are perceptually more sensitive than the former. 3. Human Visual System's (HVS) ability to access the perceptual quality of an image also depends on some other properties of the images, such as contrast. All these observations are integrated into the proposed algorithm named Distortion-Specific Contrast-Based (DSCB) IQA. Various experiments validate that the proposed DSCB IQA efficiently competes with human perception and exhibits substantially better results (at least 17% gain in terms of PLCC) when compared to the existing NR IQAs.

Color-difference interpolation (CDI) has been a widely used technique for various color demosaicking methods. CDI-based methods perform interpolation in the color-difference domain assuming that the color-difference signal is a low-pass signal. Recently, a residual interpolation (RI) algorithm, which conducts interpolation in the residual domain, has been developed, and it assumes that the residual domain is flatter or smoother than the channel-difference domain. In this paper, we comprehensively show a frequency domain analysis of these assumptions and observe that it is image dependent and creates artifacts in the interpolated image. With this view, we propose an algorithm that uses the inter-color correlation as well as the residual smoothness among the different channel much better than the existing algorithms. Experimental results emphasize that the proposed algorithm atribute better performances the existing algorithms in terms of both visual and objective quality.

To blindly evaluate the visual quality of image is of great importance in many image processing and computer vision applications. In this paper, we develop a novel training-free no-reference (NR) quality metric (QM) based on a unified brain theory, namely, free energy principle. The free energy principle tells that there always exists a difference between an input true visual signal and its processed one by human brain. The difference encompasses the “surprising” information between the real and processed signals. This difference has been found to be highly related to visual quality and attention. More specifically, given a distorted image signal, we first compute the aforesaid difference to approximate its visual quality and saliency via a semi-parametric method that is constructed by combining bilateral filter and auto-regression model. Afterwards, the computed visual saliency and a new natural scene statistic (NSS) model are used for modification to infer the final visual quality score. Extensive experiments are conducted on popular natural scene image databases and a recently released screen content image database for performance comparison. Results have proved the effectiveness of the proposed blind quality measure compared with classical and state-of-the-art full- and no-reference QMs.

Multispectral demosaicking, which is an extension of color demosaicking, is a challenging problem because each band is significantly undersampled and thus precise reconstruction is needed for the restoration of high-frequency components, such as edges, textures etc. In general, existing algorithms borrow high-frequency information either from different bands via inter-color correlation or from within the bands, and produces artifact in the reconstructed image. To meet this inherent shortcoming, we propose to incorporate two different high-frequency components and integrate them optimally in the linear minimum mean square sense (LMMSE) for the precise reconstruction of undersampled components. Experimental results demonstrate that the proposed algorithm based on the optimized high-frequency achieves superior performance compared to existing algorithms both in terms of objective and subjective quality.

Recommending content to users involves understanding a) what to present and b) how to present them, so as to increase quality of experience (QoE) and thereby, content consumption. This work attempts to address the question of how to present contents in a way so that the user finds it easy to get to desired content. While the process of User Interface (UI) design is dependent on several human factors, there are basic design components and their combination that have to be common to any recommender system user interface. Personalization of the UI design process involves picking the right components and their combination, and presenting a UI to suit the usage behavior of an individual user, so as to enhance the QoE. This work proposes a system that learns from a user's content consumption patterns and makes some recommendations regarding how to present the content for the user (in the context of Video-On-Demand/Live-TV services on Computer displays), so as to enhance the QoE of the recommender system.

Recently, a lot of bilateral and non-local means based reconstruction algorithms are proposed in literature. The success of these filters lies in finding similar patches in the neighbourhood. However, in complex regions and in presence of noise, finding similar patches based on mean square error (MSE) is not reliable. This results into blurred edges and visible patches in the reconstructed image. To address this issue, we propose a new Observation model based and Perceptually Motivated Bilateral Filter (OPBIF) for Image Reconstruction. In which image quality assessment (IQA) matrices are used to find the similarity among the patches. From experimental results, it is validated that the proposed algorithm has the capability of reconstructing sharp edges, as compared to existing non-linear filtering algorithms.

Prediction based algorithms reported in the literature are not able to integrate lossy and near-lossless/lossless coding and uses only causal pixels (non-symmetrical predictor structure) for prediction. A non-symmetrical predictor structure, however, is not able to efficiently adapt near the intensity varying areas, which results into poor prediction. Hence, we propose a novel two-stage algorithm for lossy, near lossless/lossless compression using a symmetrical predictor structure is proposed. In the first stage, the proposed algorithm encodes and decodes the given image using the JPEG-2000 standard algorithm (lossy coding). This JPEG-2000 decoded image in the first stage, enables us to use the symmetrical predictor (using both causal and non-causal pixels) for prediction in the second stage. A performance evaluation shows that our algorithm is significantly better in terms of compression performance as compared to some of the computationally complex methods.

Recently, a few bilateral filter based interpolation and intraframe deinterlacing algorithms have been proposed, but these algorithms only use prior information (bilateral filter). In this paper, we propose an efficient and fast intra-frame deinterlacing algorithm using an observation model based bilateral filter (using both likelihood and prior information). Our proposed algorithm is also able to use approximated horizontal pixels for the deinterlacing, which results into the better prediction of the edges. From extensive experiments, it is observed that the proposed algorithm has the capability of provide satisfactory results in terms of both objective and subjective quality.

High-Efficiency Video Coding (HEVC) is the newest video coding standard which can significantly reduce the bit rate by 50% compared with existing standards. One new efficient tool is sample adaptive offset (SAO), which classifies reconstructed samples into different categories, and reduces the distortion by adding an offset to samples of each category. Two SAO types are adopted in HEVC: edge offset (EO) and band offset (BO). Four 1-D directional edge patterns are used in edge offset type, and only one is selected for each CTB. However, single directional pattern cannot remove artifacts effectively for the CTBs, which contain edges in different directions. Therefore, we analyze the performance of each edge pattern applied on this kind of CTB, and propose to take advantage of existing edge classes and combine some of the them as a new edge offset class, which can adapt to multiple edge directions. All the combinations are tested, and the results show that for Low Delay P condition, they can achieve 0.2% to 0.5% bit rate reduction.

Many image interpolation algorithms have been developed in the recent past aiming for high prediction accuracy. But these algorithms are focused only towards better predictor design. In this paper, we propose a generic two phase image interpolation algorithm based upon error feedback mechanism. In the first phase, we learn error pattern occurred during interpolation of down sampled version of original Low Resolution (LR) image. It is assumed that similar error pattern also occurrs during the interpolation of original LR image. Hence, error pattern learnt in first phase, is employed during the interpolation of original LR image (second phase). From extensive experiments, we found that our algorithm gives a significant improvement in prediction accuracy of existing interpolation algorithms. In particular, our algorithm plays a significant role in improving prediction accuracy of those algorithms which have inherently poor prediction capability for certain types of images.

In this paper, we propose a new reversible watermarking algorithm based on additive prediction-error expansion which can recover original image after extracting the hidden data. Embedding capacity of such algorithms depend on the prediction accuracy of the predictor. We observed that the performance of a predictor based on full context prediction is preciser as compared to that of partial context prediction. In view of this observation, we propose an efficient adaptive prediction (EAP) method based on full context, that exploits local characteristics of neighboring pixels much effectively than other prediction methods reported in literature. Experimental results demonstrate that the proposed algorithm has a better embedding capacity and also gives better Peak Signal to Noise Ratio (PSNR) as compared to state-of-the-art reversible watermarking schemes.

Merge/skip mode is one of the most important inter prediction tools adopted in the High Efficiency Video Coding (HEVC) standard which is the state-of-the-art video coding standard. It is very efficient in reducing the side information for the blocks within the same object. However, it is difficult for parallel encoding and decoding due to the data dependency problem between neighboring prediction units (PU). Furthermore, different shapes and positions of PUs would result in different definition of the merge/skip candidate list (MCL), which would lead to potentially extra hardware cost and is not easy to be efficiently implemented by the hardware. To deal with this problem, two reconfigurable hardware-friendly MCL construction schemes are proposed in this paper. The first scheme which is called unified MCL (UMCL) uses one candidate list for all PUs inside the motion estimation region (MER), which is regarded as the basic parallel processing unit for the hardware realization. The second scheme which is named boundary MCL (BMCL) allows different candidate lists for the PUs on the boundary of MER. Both of the two schemes can have flexible parallel degree based on the requirement specification. Experimental results show that UMCL reduces the hardware complexity significantly with little coding performance degradation and BMCL achieves significant coding gain while maintaining the hardware complexity.

In this paper, we are proposing a simple lossless prediction based coding method for videos. Our algorithm works on 3 modes of operations and selection of the mode is done on a pixel-by-pixel basis. Selection of the mode is based on intensity variation of pixels in temporal direction. If there is a large intensity value variation, intraframe prediction mode is chosen. Otherwise interframe prediction mode is activated. Intraframe prediction uses Gradient Adaptive Predictor [5] whereas Interframe prediction switches between two algorithms. In interframe case, one prediction mode takes care of local characteristics of pixels of current and motion compensated frames while the other incorporates global characteristics. The proposed method is computationally very simple and results into better performance as compared to competitive but complex methods reported in literature.

Bit-depth expansion is an art of converting low bit-depth image into high bit-depth image. Bit-depth of an image represents the number of bits required to represent an intensity value of the image. Bit-depth expansion is an important field since it directly affects the display quality. In this paper, we propose a novel method for bit-depth expansion which uses Minimum Risk Based Classification to create high bit-depth image. Blurring and other annoying artifacts are lowered in this method. Our method gives better objective (PSNR) and superior visual quality as compared to recently developed bit-depth expansion algorithms.

Predictor based algorithms reported in literature uses only causal pixels and hence a non-symmetrical predictor structure for prediction. We observed that the performance of predictor is highly dependent on the predictor structure used. In view of this, we propose a novel interpolation based prediction scheme that enables us to use symmetrical predictor structure. In this sense, we have also used non causal pixels in our scheme. Also, from various interpolation algorithms available, we selected a simple one to ensure decoder simplicity, without any significant loss in performance. From performance evaluation, we found that our algorithm is significantly better in terms of compression performance as compared to some of the computationally complex methods.

In this paper we present two computationally simple algorithms that can be used for prediction of pixels of images. In one of the algorithms, prediction is made by estimating intensity value variations in four directions and their reciprocals are used to make prediction of unknown pixel. This algorithm captures local characteristics of the unknown pixel well as it uses only a small number of neighborhood pixels. The other algorithm finds slope as the relative intensity-value variations and classifies image pixels in fourteen bins by classifying the slope in the same number of bins. LS based predictors are estimated for pixels belonging to each of the bins and hence the they represent global characteristics of these pixels. Since one algorithm takes care of local characteristics while the other one represents global feature, we propose a switching method for these two algorithms that takes advance of both the algorithms. Switching is done on a pixel-by-pixel basis and the same gives approximately 0.10 bpp better performance as compared to some of the computationally complex methods reported in literature at a lower computational complexity.

Recently a lot of interpolation algorithms are proposed, but these interpolation algorithms are highly computationally expensive. Hence these algorithms cannot be implemented and used in real time applications. In view of real time applications we have proposed a computationally simple interpolation algorithm. In our proposed algorithm the unknown pixels are categorized into various bins depending upon the characteristic of the neighboring pixels (activity level) and for each bin fixed prediction parameters are used for prediction. We have presented different set of fixed predictors for both smooth type and edgy type of images. We have also proposed a modified algorithm in which selection of prediction parameter is done on block by block basis instead of image basis. Our proposed algorithm gives much better qualitative and quantitative performance as compared to other computationally simple interpolation algorithms.

In this paper, we propose a prediction algorithm that uses adaptive predictor structures for loss less video coding. The proposed encoder finds cross-correlation coefficient between current frame and motion compensated previous frame and classifies the coefficient value into a small number of bins. For general videos, we propose four bins and associate different predictor structures with each of the bins. Similarly for medical videos, numbers of bins are three and each of these bins is associated with different predictor structures. Performance of our method is compared with the fixed predictor structure used in practice and found that the proposed method not only gives better prediction performance but it is computational efficient as compared to fixed predictor structure.

In this paper we proposed a new computationally efficient interpolation algorithm for natural images in which unknown pixels are divided into few bins. The categorization of these unknown pixels into bins is based upon the characteristics of the neighboring pixels. These characteristics are obtained by taking difference of two slopes which are in orthogonal direction and these slopes are calculated from a set of neighboring pixels. We used the Least-Squares (LS) based approach to find optimal predictors for pixels belonging to various slope bins. We also presented a simplified proposed algorithm in which we used bilinear interpolation algorithm instead of estimating LS based predictor for some bins and it results into further reduction in computational complexity without sacrificing the much performance. Our proposed algorithm gives better interpolation quality with significantly lower computational complexity as compared to recently reported interpolation algorithms.

In this paper, we propose selective estimation of least square based predictor algorithm and efficient overhead management scheme for lossless compression of digital mammograms. We exploit the characteristics of mammograms that most of the mammograms contain large number of blocks with constant gray level pixels, so a block based selective least square estimation algorithm is proposed. In our proposed algorithm if all the pixels have same intensity value in any block, then we represents those blocks by a single (‘1‘) bit otherwise the block is decorrelated using the feed forward type of autoregressive modeling. We exploit the relationship between autoregression parameters which saves around 25% overhead burden. We have also empirically found that the AR parameters of the neighboring blocks are highly correlated and to get the best decorrelation among these parameters, median edge detector (MED) is used which gives us around 40% more saving in overhead burden. So, our proposed lossless compression algorithm for digital mammograms gives better entropy and minimum overhead burden then most of the algorithms reported in literature.

This paper presents a novel algorithm for enhancement of spatial resolution of images. The proposed algorithm estimates a Least square based predictor of lower order and interpolates the coefficients of higher order predictor. We have reduced the predictor order form p to (p−1) that results into a saving of computational power. The proposed algorithm is generic that can be used with most of the LS based interpolation algorithms reported in literature. We have shown that use of interpolated prediction coefficient causes insignificant loss in subjective as well as objective (PSNR) quality of the higher resolution (HR) image as compared with the PSNR obtained by the actual prediction coefficient and there is around 40% to 50% reduction in computational complexity.