International Journal of Image, Graphics and Signal Processing @ijigsp
Journal articles - International Journal of Image, Graphics and Signal Processing
All articles: 1220
Bio-chip design using multi-rate system for EEG signal on FPGA
Scientific article
Digital Signal Processing (DSP) is one of the fastest growing techniques in the electronics industry. The signal-rate system in digital signal processing has evolved the key of fastest speed in digital signal processor. Field Programmable Gate Array (FPGA) offers good solution for addressing the needs of high performance DSP systems. The focus of this paper is on the basic DSP functions, namely filtering signals to remove unwanted frequency bands. Multi-rate Digital Filters (MDFs) are the main theme to build bio-chip design in this paper. For different purposes DSP systems need to change the sampling rate of the signal to achieve some applications. This can be done using multi-rate system where designers can increase or decrease the operating sampling rate. This bio-chip has attractive features like, low requirement of the coefficient word lengths, significant saving in computation time and storage which results in a reduction in its dynamic power consumption. This paper introduces an efficient FPGA realization of multi-rate digital filter with narrow pass-band and narrow transition band to reduce noises and changing the frequency sampling rate by factor which is required according to application. This bio-chip works on bio-signals like EEG signal.
Free
Bit Serial Architecture for Variable Block Size Motion Estimation
Scientific article
H.264/AVC is the latest video coding standard adopting variable block size, quarter-pixel accuracy and motion vector prediction and multi-reference frames for motion estimations. These new features result in higher computation requirements than that for previous coding standards.The computational complexity of motion estimation is about 60% in the H.264/AVC encoder. In this paper most significant bit (MSB first) arithmetic based bit serial Variable Block Size Motion Estimation (VBSME) hardware architecture is proposed. MSB first bit serial architecture main feature is, its early termination SAD computation compared to normal bit serial architectures. With this early termination technique, number computations are reduced drastically. Hence power consumption is also less compared to parallel architectures. An efficient bit serial processing element is proposed and developed 2D architecture for processing of 4x4 block in parallel .Inter connect structure is developed in such way that data reusability is achieved between PEs. Two types of adder trees are employed for variable block size SAD calculation with less number of adders. The proposed architecture can generate up to 41 motion vectors (MVs) for each macroblock. The inter connection complexity between PEs reduced drastically compared to parallel architectures. The architecture supports processing of SDTV (640x480) with 30fps at 172.8 MHz for search range [+8, -7]. We could reduce 14% of computations by using early termination technique.
Free
Scientific article
This paper introduces BlinkFusion, a real-time, interpretable, and detector-independent pipeline for analyzing blink events. The system finds eye ROIs using YOLOv5 as the main detector and a Haar cascade as a backup that has been calibrated. It then stabilizes the detections using lightweight tracking. A small landmark regressor inside each ROI gives six points to calculate the Eye Aspect Ratio (EAR), which keeps the geometric meaning. An uncertainty-weighted smoother combines pose, landmark, and detector confidence. An online hysteresis state machine with minimum-duration and derivative gates makes blink onsets and offsets. Platt-calibrated and fused detector confidences make strong arbitration possible at a reasonable cost. Performance is assessed on EyeBlinkDB (RGB ≥25 fps, ~720p), using subject-independent 10-fold splits and temporal-IoU matching. The model achieves Precision 0.928 ± 0.008, Recall 0.945 ± 0.008, F1 0.936 ± 0.008, AP@tIoU=0.3 = 0.962 ± 0.006, with timing precision of 28.1 ± 1.7 ms onset MAE and 37.1 ± 2.3 ms offset MAE. Condition stratification validates robustness: The values for Bright/Dim F1 are 0.947/0.922, for Frontal/±15°/≥30° yaw F1 they are 0.952/0.936/0.903, for Glasses (No/Yes) they are 0.946/0.925, and for Occlusion (No/Yes) they are 0.948/0.906. With adaptive scheduling (YOLO invocation rate ρ≈0.045), pipeline runs at about 32 FPS on the CPU (about 90 FPS on the observed frame loop) with a fixed 3-frame delay, which is fast enough for real-time use on cheap hardware. Ablations demonstrate progressive improvements resulting from calibrated fusion, tracking, quality-weighted EMA, and hysteresis (F1: 0.903 → 0.952). The approach is modular (you may switch the detector), doesn't need bounding boxes, and makes judgments that can be explained using EAR traces and thresholds. So, BlinkFusion is a useful, ready-to-use solution for HCI and clinical contexts that need clear, accurate, and quick blink analytics.
Free
Block Texture Pattern Detection Based on Smoothness and Complexity of Neighborhood Pixels
Scientific article
In this paper, a novel method for detecting Block Texture Patterns (BTP), based on two measures: smoothness and complexity of neighborhood pixels is proposed. With these two measures, a new classification for texture detection is defined. Texture detection with these measures can be used in many image processing and computer vision applications. As an example, the applicability of BTP on data hiding algorithms is discussed, and the advantages of this classification on these algorithms are shown.
Free
Block-Based Compressive Sensed Thermal Image Reconstruction using Greedy Algorithms
Scientific article
This paper implements a block based compressive sensing technique for thermal image reconstruction using greedy algorithms. A total of fourteen different sensing patterns were tested for data acquisition. Orthogonal Matching Pursuit (OMP) and Regularized Orthogonal Matching Pursuit (ROMP) with two different thresholds were implemented for image reconstruction with OMP having an edge over ROMP in terms of error and PSNR. ROMP was faster in terms of iterations needed for reconstruction. As the threshold for ROMP was increased the number of iterations needed decreased. Gaussian, Bernoulli and Hadamard patterns were the best for reconstruction. Hadamard matrix, Bernoulli matrix with +/-1 entries and Bernoulli matrix with 0/1 entries have the added advantage of being more conducive for hardware implementation. This paper used Discrete Cosine Transform as the sparsifying basis for reconstruction.
Free
Block-based Local Binary Patterns for Distant Iris Recognition Using Various Distance Metrics
Scientific article
Nowadays iris recognition has become a promising biometric for human identification and authentication. In this case, feature extraction from near-infrared (NIR) iris images under less-constraint environments is rather challenging to identify an individual accurately. This paper extends a texture descriptor to represent the local spatial patterns. The iris texture is first divided into several blocks from which the shape and appearance of intrinsic iris patterns are extracted with the help of block-based Local Binary Patterns (LBPb). The concepts of uniform, rotation, and invariant patterns are employed to reduce the length of feature space. Additionally, the simplicity of the image descriptor allows for very fast feature extraction. The recognition is performed using a supervised machine learning classifier with various distance metrics in the extracted feature space as a dissimilarity measure. The proposed approach effectively deals with lighting variations, blur focuses on misaligned images and elastic deformation of iris textures. Extensive experiments are conducted on the largest and most publicly accessible CASIA-v4 distance image database. Some statistical measures are computed as performance indicators for the validation of classification outcomes. The area under the Receiver Operating Characteristic (ROC) curves is illustrated to compare the diagnostic ability of the classifier for the LBP and its extensions. The experimental results suggest that the LBPb is more effective than other rotation invariants and uniform rotation invariants in local binary patterns for distant iris recognition. The Braycurtis distance metric provides the highest possible accuracy compared to other distance metrics and competitive methods.
Free
Blur Classification using Ridgelet Transform and Feed Forward Neural Network
Scientific article
The objective of image restoration approach is to recover a true image from a degraded version. This problem can be stated as blind or non-blind depending upon whether blur parameters are known prior to the restoration process. Blind restoration deals with parameter identification before deconvolution. Though there exists multiple blind restorations techniques but blur type recognition is extremely desirable before application of any blur parameters estimation approach. In this paper, we develop a blur classification approach that deploys a feed forward neural network to categories motion, defocus and combined blur types. The features deployed for designing of classification system include mean and standard deviation of ridgelet energies. Our simulation results show the preciseness of proposed method.
Free
Brain Tissue Classification from Multispectral MRI by Wavelet based Principal Component Analysis
Scientific article
In this paper, we propose a multispectral analysis system using wavelet based Principal Component Analysis (PCA), to improve the brain tissue classification from MRI images. Global transforms like PCA often neglects significant small abnormality details, while dealing with a massive amount of multispectral data. In order to resolve this issue, input dataset is expanded by detail coefficients from multisignal wavelet analysis. Then, PCA is applied on the new dataset to perform feature analysis. Finally, an unsupervised classification with Fuzzy C-Means clustering algorithm is used to measure the improvement in reproducibility and accuracy of the results. A detailed comparative analysis of classified tissues with those from conventional PCA is also carried out. Proposed method yielded good improvement in classification of small abnormalities with high sensitivity/accuracy values, 98.9/98.3, for clinical analysis. Experimental results from synthetic and clinical data recommend the new method as a promising approach in brain tissue analysis.
Free
Scientific article
Brain MRI is mainly affected by noise and intensity inhomogeneity (IIH) during its acquisition. Brain tissue segmentation plays an important role in biomedical research and clinical applications. Brain tissue segmentation is essential for physicians for the proper diagnosis and right treatment of brain-related disorders. Fuzzy C-means (FCM) clustering is one of the widely used algorithms for brain tissue segmentation. Traditional FCM has the limitations of misclassification of pixels that leads to inaccurate cluster centers. Due to this, it is unable to address the issues of noise and IIH. In FCM there exists uncertainty in controlling the fuzziness of the clusters as the fuzzifier is fixed. This paper proposed a novel linguistic fuzzifier-based FCM (LFFCM) to overcome the limitations of traditional FCM during brain tissue segmentation from the MR images. In this method, a linguistic fuzzifier is used instead of a fixed fuzzifier. The spatial information incorporated in the membership function can reduce the misclassification of pixels. The proposed LFFCM can handle IIH, due to having highly accurate cluster centers. The inclusion of the adaptive weights in the membership function results in accurate cluster centers. Various brain MR images are used to evaluate the proposed technique and the results are compared with some state-of-the-art techniques. The results reveal that the proposed method performed better than the other.
Free
Scientific article
Tumor boundary detection is one of the challenging tasks in the medical diagnosis field. The proposed work constructed brain tumor boundary using bi-modal fuzzy histogram thresholding and edge indication map (EIM). The proposed work has two major steps. Initially step 1 is aimed to enhance the contrast in order to make the sharp edges. An intensity transformation is used for contrast enhancement with automatic threshold value produced by bimodal fuzzy histogram thresholding technique. Next in step 2 the EIM is generated by hybrid approach with the results of existing edge operators and maximum voting scheme. The edge indication map produces continuous tumor boundary along with brain border and substructures (cerebrospinal fluid (CSF), sulcal CSF (SCSF) and interhemispheric fissure) to reach the tumor location easily. The experimental results compared with gold standard using several evaluation parameters. The results showed better values and quality to proposed method than the traditional edge detection techniques. The 3D volume construction using edge indication map is very useful to analysis the brain tumor location during the surgical planning process.
Free
Brain Tumor Classification Using Back Propagation Neural Network
Scientific article
The conventional method for medical resonance brain images classification and tumors detection is by human inspection. Operator-assisted classification methods are impractical for large amounts of data and are also non-reproducible. Hence, this paper presents Neural Network techniques for the classification of the magnetic resonance human brain images. The proposed Neural Network technique consists of the following stages namely, feature extraction, dimensionality reduction, and classification. The features extracted from the magnetic resonance images (MRI) have been reduced using principles component analysis (PCA) to the more essential features such as mean, median, variance, correlation, values of maximum and minimum intensity. In the classification stage, classifier based on Back-Propagation Neural Network has been developed. This classifier has been used to classify subjects as normal, benign and malignant brain tumor images. The results show that BPN classifier gives fast and accurate classification than the other neural networks and can be effectively used for classifying brain tumor with high level of accuracy.
Free
Breast Cancer Classification from Ultrasound Images using VGG16 Model based Transfer Learning
Scientific article
Ultrasound based breast screening is gaining attention recently especially for dense breast. The technological advancement, cancer awareness, and cost-safety-availability benefits lead rapid rise of breast ultrasound market. The irregular shape, intensity variation, and additional blood vessels of malignant cancer are distinguishable in ultrasound images from the benign phase. However, classification of breast cancer using ultrasound images is a difficult process owing to speckle noise and complex textures of breast. In this paper, a breast cancer classification method is presented using VGG16 model based transfer learning approach. We have used median filter to despeckle the images. The layers for convolution process of the pretrained VGG16 model along with the maxpooling layers have been used as feature extractor and a proposed fully connected two layers deep neural network has been designed as classifier. Adam optimizer is used with learning rate of 0.001 and binary cross-entropy is chosen as the loss function for model optimization. Dropout of hidden layers is used to avoid overfitting. Breast Ultrasound images from two databases (total 897 images) have been combined to train, validate and test the performance and generalization strength of the classifier. Experimental results showed the training accuracy as 98.2% and testing accuracy as 91% for blind testing data with a reduced of computational complexity. Gradient class activation mapping (Grad-CAM) technique has been used to visualize and check the targeted regions localization effort at the final convolutional layer and found as noteworthy. The outcomes of this work might be useful for the clinical applications of breast cancer diagnosis.
Free
Breast lesion segmentation and area calculation for MR images
Scientific article
In this paper, our goal is to determine the boundaries of lesion and then calculate the area of existing lesion in breast magnetic resonance (MR) images to provide a useful information to the radiologists. For this purpose, at first stage region growing (RG) method and active contour model (Snake) is applied to the images to make the boundaries of lesion visible. RG method is one of the simplest approaches for image segmentation and provides accurate results with lower computation time due to its seed point initialization step. Snake method molds a closed contour to the boundary of a region in an image and is also popular in medical image segmentation studies. In the presented study, both of these methods are utilized to determine the lesion boundaries. After determining the boundaries of lesion accurately in the second stage of the study, bit-quad method is applied to the segmented images. Bit quad method is used to compute the area and perimeter of binary lesion images based on matching the logical state of regions of image to binary patterns. Finally, to evaluate the performance of the proposed study, computer simulations are performed. It is demonstrated via computer simulations that the lesion area and parameter values are very close to real values. By means of this study it is aimed to support radiologists during diagnosis and assessment of breast lesions.
Free
Scientific article
While the use of artificial intelligence in ophthalmology has increased over the past few years, inflammatory diseases like choroiditis remain widely unexplored. A delayed diagnosis in such cases can cause severe complications that can lead to visual disability, and hence, automated diagnosis systems can be of great importance. In this study, we present a systematic, evidence-driven methodology for designing a lightweight choroiditis classifier using a real-world small dataset with a constraint of low resource availability. This hybrid framework consists of two components. The first feature extraction component extracts features using the MobileNet-V3-Small model. The second component, namely the classification component, utilises the Support Vector Machine as the binary classifier. This optimal combination was identified through systematic comparative experiments. Statistical testing confirms the robustness of the classifier selection. The model gives a cross-validation accuracy of 94% and a held-out test accuracy of 97.06% with a training time of approximately 3 minutes for the end-to-end pipeline on a carefully collected and previously introduced choroiditis dataset. Being lightweight and computationally efficient, this model is a suitable candidate for future development into a reliable computer-aided diagnosis tool that could assist experts in reviewing images, providing telemedicine care, and prioritising patient appointments based on the initial results of the automated systems.
Free
Scientific article
The growing quality and safety concern about fresh agricultural produce among consumers have led to the development of non-destructive quality assessment and testing techniques of fruits and vegetables. Humans judge the quality of fruits based on sensory attributes like taste, aroma etc. The shape, size, color, presence of defects which are external to fruits also influence the degree of consumer acceptability of produce. The traditional time consuming, manual fruit quality inspection is replaced by automated, fast, consistent, non-destructive techniques using computer vision in combination with learning algorithms. But the lack of benchmark datasets for agricultural produce has made an objective comparison of the proposed methods difficult. Hence, the proposed work aims to build a medium scale dataset for mango fruits of “Alphonso” cultivar with three classes: chilling injury, defective and non-defective. The reliability of the proposed dataset consisting of 2279 color images of mango fruits with 736 samples in chilling injury class, 632 samples in defective class and 911 samples in non-defective class, was established using a novel approach of developing a predictive model based on discriminant function analysis (DFA) which assigns group membership to each sample of the dataset. Extensive benchmarking analysis is established on the validated dataset using statistical and deep learning algorithms like support vector machine (SVM) and convolutional neural network (CNN), respectively. SVM achieved significant disease classification accuracy of 95% and 91.52% accuracy was achieved by custom CNN. The results of the proposed work indicate that the proposed color image dataset of mango fruits can be used as a benchmark dataset by other researchers for objective comparison in quality evaluation of mango fruits.
Free
Scientific article
Federated learning enables industrial sites to train shared condition-monitoring models without exchanging raw vibration, acoustic, or sensor data, but its openness admits Byzantine clients that submit corrupted updates. Existing robust aggregation rules—Krum, trimmed mean, median, and their variants—operate in Euclidean parameter space and implicitly assume that a malicious update must be far from honest updates to do harm. We show that this assumption fails for models with Kolmogorov–Arnold (KAN) spline heads: an adversary can hold the spline coefficients close to the honest consensus in the L2 sense while distorting the functional shape the spline encodes, an attack we term functional spline-poisoning. We propose ByzaSpline-Fed, a Byzantine-robust aggregation rule that scores clients by Gram-weighted functional B-spline distance on a consensus knot grid and rejects functional outliers. The rule reduces exactly to spline-aware federated averaging under honest participation, admits a breakdown bound of f < K/2 under the standard honest-majority assumption, and supports optional differential privacy. We evaluate across four industrial benchmarks spanning classification, unsupervised acoustic anomaly detection, and remaining-useful-life regression—CWRU, Paderborn, MIMII, and C-MAPSS—reporting task-correct metrics under clean and adversarial conditions. Under thirty percent functional poisoning, Euclidean defenses lose up to thirty-four accuracy points while ByzaSpline-Fed remains within roughly three points of its clean performance. All reported results are averaged over eight independent random seeds, and the improvements over the strongest competing defense are statistically significant under the Wilcoxon signed-rank test (p < 0.05).
Free
Scientific article
Colon cancer is a growth of cells that begins in a part of the large intestine called the colon. Colon cancer happens when cells in the colon develop changes in their DNA. Consequently, fewer infections and fatalities may result from early identification of this cancer. Histological analysis is used for a final diagnosis of colon cancer. Histopathology, or the microscopic examination of damaged tissue, is crucial for both cancer diagnosis and treatment. This work suggests a novel deep learning technique for colon cancer detection effectively. Histopathology images are collected from various type of sources. To enhance the quality of raw images, pre-processed techniques such as image scaling, colour map improved image sharpening, and image restoration are used. Resize the image's dimensions in image resizing to minimize the processing time. A colour map enhances the sharpness of an image by combining two techniques: The contrast adjustment technique is used to alter the image's contrast first. The resultant image is then enhanced by applying the image sharpening process and scaling it using a weighting fraction. As using the final image has increased quality, blur and undesirable noise are removed using image restoration. Next, the pre-data are used in the Attention U-Net segmentation procedure, which segments the region of the pre-data. To extract features from this segmented image to perform an accurate diagnosis, efficientnetB0 is used. In data extraction, the Bidirectional GRU model is used to process the data further in order to develop predictions. When processing input sequences in both directions with the BiGRU model, it is feasible to gather contextual information to increase accuracy and predict colon cancer effectively. In the proposed model colon disease prediction classifier offer 97% accuracy, 96% specificity and 95.49% F1_score. Thus, the proposed model effectively predicts colon cancer and improves accuracy.
Free
CNN and GAN Based Stroke Detection Using CT Scan Images
Scientific article
The objective of the research work is to detect stroke using CT scan images. In the research work an analysis of 3D CNN method for stroke detection is presented. The work also presents a new method of stroke detection using semi-supervised Adversarial Networks (SGAN).3D CNN is the traditional approach to any type of image classification problem. But being data-hungry, it becomes difficult to use them when data is scarce. High-quality medical data is difficult to find and hence alternative approaches seem worth approaching. The relatively new GANs can generate images like the training images, and its SGAN variant can use these generated images for training the classifier. We investigate the usefulness of SGANs comparatively with CNNs in this paper. The proposed SGAN method is compared with state of art methods in literature using accuracy, sensitivity and specificity. The SGAN method demonstrates an accuracy of 93%, Sensitivity of 100% and Specificity of 90%. For small data sets in medical imaging the proposed SGAN method exhibit an encouraging performance as compared to other methods using large datasets. In the research paper, we propose methodologies for detecting strokes by using 2 approaches: 3D CNNs and SGANs. The relatively new GANs can generate images like the training images, and its SGAN variant can use these generated images for training the classifier. We investigate the usefulness of SGANs comparatively with CNNs in this paper.
Free
COVID-19 Automatic Detection from CT Images through Transfer Learning
Scientific article
Identification of COVID-19 may help the community and patient to prevent the disease containment and plan to attend disease in right time. Deep neural network models widely used to analyze the medical images of COVID-19 for automatic detection and give the decision support for radiologists to summarize the accurate remarks. This paper proposed deep transfer learning for chest CT scan images to detection and diagnosis of COVID-19. VGG19, InceptionRestNetV3, InceptionV3 and DenseNet201 neural network used for automatic detection of COVID-19 disease form CT scan images (SARS-CoV-2 CT scan Dataset). Four deep transfer learning models were developed, tested and compared. The main objective of this paper is to use pre-trained features and converge pre-trained features with targeted features to improve the classification accuracy. It is observed that DenseNet201 noted the best performance and the classification accuracy is 99.98% for 300 epochs. The findings of the experiments show that the deeper networks struggle to train adequately and give less consistency when there is limited data. The DenseNet201 model adopted for COVID-19 identification from lung CT scans has been intensively optimized with optimal hyper parameters and performs at noteworthy levels with precision 99.2%, recall 100%, specificity 99.2%, and F1 score 99.2%.
Free
Scientific article
This work introduces the novelty as an application of histogram-based bins approach with statistical moments for detecting and classifying malaria using blood smear images into parasitized and uninfected cell images and the rising disease of COVID-19/Normal lung images. Proposed algorithms greatly vary as compared to the previous work. This work aims to improve accuracy in detection and classification and reduce feature vector dimensionality. It focuses on detailed image contents extracted into 8 bins by considering the significance of the R, G, and B color component relationship in the formation of each pixel. The texture features are represented by the first four moments for each of the three colors separately. This leads to the generation of 12 features vectors, each of size 8 components for each image in the database. Feature dimensionality reduction is achieved by applying different feature selection techniques to obtain desired optimum feature space. The comprehensive feature analysis presented here identifies many useful findings in order to validate the contribution of each image content uniquely in detection and classification. The proposed approach experimented with two image datasets: the malaria dataset obtained from the National Library of Medicine (NLM) and the lung image dataset acquired from the Radiography Database from Kaggle. The performance of work presented here is evaluated and compared with previous work with the same set of parameters, namely precision, recall, F1 score, and the AUC. We have achieved and improved the performances compared to previous work and also achieved better results even for the COVID-19 dataset.
Free
