نشریه ماشین بینایی و پردازش تصویر
سال هفتم شماره 2 (پاییز و زمستان 1399)

In this paper, a two stage co-segmentation method,based on matrix decomposition,has been proposed. In the first step, each image is segmented into some super-pixels and then salient parts of each image are extracted via structured matrix decomposition (SMD) method. The low-rank matrix represents image background and the sparse matrix contains salient objects. In this step, the super-pixels that are partitioned as background with high confidence will be removed. In the second step, all remaining super-pixels are considered all together and the tree structure is rearranged and then the SMD method is applied again to this new data. Parts of the common salient object compose the low-rank matrix due to the large number of them in the remaining super-pixels. In other words, the proposed approach has embedded intra-image adjacency information and inter-images similarity information into the matrix decomposition method via proper weighting of the tree structure.iCoseg dataset has been used to evaluate its performance. The results demonstrate its effectiveness and superiority.

Nowadays, nondestructive evaluations (NDE) techniques for the diagnosis of defects in the industrial components follow three steps: detection, location, and Determination of defect profile. Despite the fact that the NDE techniques available in the industry have fairly acceptable results for defect detection and localization, but accurate diagnosis of the shape, dimensions, and depth of the defect still remained a challenging task. In this paper, a method for reliable estimation of defect profile in conductive materials is presented using an eddy current testing (ECT) based measurement system and a post-processing technique based on deep learning approach. Accordingly, a deep learning method is used for defect characterization in metallic structures through magnetic field C-scan images which have been obtained by an anisotropic magneto-resistive (AMR) sensor. In this regards, after modeling and regulating the deep convolutional neural network (DCNN) to apply to the obtained C-scan images, the performance of the proposed deep learning method is compared with the conventional artificial neural networks (ANNs) such as multi-layer perceptron (MLP) and Radial based function (RBF) on a number of specimens with different known defects. Results confirm the superiority of the proposed approach relative to other conventional methods for defect profile estimation.

Deep learning, if used by an appropriately large dataset, can result in considerable accuracy. However, preparing a large dataset with accurate annotation is a costly process. For this reason, weakly supervised learning algorithms have attracted many researchers in recent years. In a weakly supervised learning approach, a large dataset is gathered with weak annotation to meet both the need of large dataset for deep learning, and reduce the cost of annotations. In this paper, a weakly supervised learning algorithm is proposed to recognize vehicle license plates using deep convolutional neural networks. In the training phase, only the characters existed in the plate images are labeled, and the proposed algorithm can recognize the presence of each character in addition to their coordinates. Therefore, in the test phase, the license plate is fully recognizable. In order to evaluate the performance of the proposed algorithm, a dataset containing 1397 plate images has been collected. The results of the experiments show that 95.6% of the plates and 99.1% of the characters are correctly recognized.

In this paper, a method for enhancing the contrast of mammogram images based on the design of a nonlinear polynomial filter is presented. Mammograms suffer from low contrast and also basic contrast enhancing techniques such as Histogram Equalization and its modifications, which have good performance against natural images- are challenged when faced with mammograms. Because increasing the contrast of medical images such as mammograms leads to saturation in bright regions and weakens marginal pattern of tumor. In this work, the design of non-linear filter is done such that shape of the potential tumor -that plays an important role in diagnosing it- are preserved. In this paper, unlike other methods, the nonlinear filter coefficients are adjusted to each input image through a support of the well-known contrast enhancement algorithm, i.e. CLAHE. Quantitative and qualitative evaluations of this method in various experiments and on several images from MIAS dataset show that the proposed method has acceptable performance.

Mammography is the most common and effective screening method for breast cancer detection. In this paper a computer aided system for classification of benign and malignant tumors in digital mammogram is presented. First, a median filter is used for noise reduction, and then artifacts and pectoral muscle are removed to make the mammogram ready for segmentation. For segmentation of mammogram, a new contrast enhancement method is presented which employs the difference of two complement enhanced images and then a histogram based fuzzy C-means (HFCM) clustering are used for region-of-interest (ROI) extraction. Then, some geometrical and textural features are extracted, and finally linear support vector machine and decision tree classifier are used to classify the region of interest into benign and malignant classes. The proposed algorithm is validated on the MIAS and DDSM databases. The experimental results showed that the performance of the proposed method is promising compared to the other methods evaluated.

In this paper, we have proposed a new joint architecture using Deep Neural Network (DNN) and a traditional descriptor for feature extraction towards signature identification. The proposed approach is an extended version of ResNet-18, which is enhanced using our two paths architecture. In the first path, we explore features using a deep convolutional neural network, and in the second path, we discover global features using a traditional heuristic approach. Our traditional approach extracts global features that are stable with rotation and scaling. For evaluation, we performed extensive experiments on accessible datasets of CEDAR, UTsig, and GPDS through the proposed approach. Our results show that the proposed joint approach outperformed the baseline ResNet-18 and demonstrate our approach superiority. Also, the comparisons with the related works show that our approach results are better or in par with state of the art.

Endoscopy is a technique that facilitates the visualization of the internal parts of the patient's body with minimal non-invasive intervention. In recent years, many studies have focused on analyzing acquired signals from the endoscopic and surgical camera. In the present study, a comprehensive overview of the types of endoscopic image processing techniques will be presented and a general classification of these methods will be presented. This paper uses research from databases of IEEE, Science Direct, Pub Med, SPIE and Medical Physics. In order to create an outlook of this research, an introduction to endoscopy and a variety of endoscopic methods in medical image analysis is presented. According to the researches and studies investigated in this study, we categorize the perspectives in the process of endoscopic images into four general approaches: methods based image improvement and enhancement, real-time processing methods, image quality evaluation methods, and image-based management methods. Subcategories of these groups may overlap each other in some cases.

Synthetic Aperture Radar (SAR) images have many applications in different fields. But, the existence of multiplicative noise called speckle is a problem in processing SAR images. So, speckle suppression is an essential step in processing SAR images. In this paper, a novel Bayesian wavelet domain speckle suppression method is proposed that is based on joint modeling. In the proposed method, first logarithm and then wavelet transform are applied to the SAR images. Normal Inverse Gaussian (NIG) distribution is used for statistical modeling of wavelet coefficients. Due to the importance of precise statistical modeling and the existence of dependency between wavelet coefficients, proposed method captures the dependencies of wavelet coefficients using joint modeling. Then, based on using the proposed model, a minimum mean square error (MMSE) estimator is designed to denoise SAR images. Experimental results using synthetic and real SAR images demonstrate the efficiency of the proposed method.

In today’s modern medicine, with the spreading use of radiological imaging devices in medical centers, the need to accurate, reliable, and portable medi cal image analysis and understanding systems has been increasing constantly. Since usually the images are not accompanied by the required clinical anno tation, automatic tagging and captioning systems are among the most desired applications. This research proposes an automatic structured radiology report generation system that is based on annotation methods. Extracting useful and descriptive image features to model conceptual contents of the images is one of the main challenges in this regard. Considering the ability of deep neural networks in soliciting informative and effective features as well as lower reso urce requirements, MobileNets are employed as the main building block of the proposed system. Furthermore due to the lack of large labeled medical data for training the network and risk of over-fitting, a joint descriptor is induced from the deep features and local bina ry patterns. Experimental results confirm the efficiency of the proposed hybrid approach with accuracy 91.4%, as compared to the end-to-end deep networks and classic annotation methods.

Thickness evaluation and analysis of corneal layers are important for diagnosis and treatments considering corneal disease. Optical Coherence Tomography (OCT) can produce micron-scaled cross-sectional images in a non-invasive and non-contacting manner. Since manual segmentation and layer detection within such images are time-consuming, physicians prefer automatic/semi-automatic methods. This paper reviewed main and important methods of corneal layer segmentationsapplied to OCT images. The methods are compared and described in three categories: preprocessing, segmentation, and thickness mapping (layers’ topography). The purpose of preprocessing was to remove noise and artifacts from such OCT images. Studies show that methods based on Hough transform, which are consistent with the corneal arc structure when compared to graph and threshold methods, are able to extract accurate boundaries in a reasonable time. Meanwhile, artificial intelligence and the deep learning approach has opened new horizons in segmentation and analysis of such images. In studies, generally the aim was to extract and present OCT image information in a form that would help ophthalmologists better diagnose and treat corneal abnormalities; therefore it can be concluded thatlayer topography and its related issues that require automatic processing of a set of cross-sectional images, isan important output that is not addressed in many research.

One of the important factors in growth of microalgae is the amount of salt needed to feed them. In this study, the culture medium for microalgae, Nannochloropsis Oculata, was prepared at different concentrations of salt and the growth of active microalgae was investigated with the help of state machine technology on every day. The maximum and minimum concentration of microalgae on the seventh of the breed were obtained286.23 × 104 ± 0.38 × 105 cells per ml at 30 mg per liter and a minimum salinity of 100 mg per liter with the density treatments density 168.58 × 104 ± 0.48 × 105 cells per ml, respectively. In the growth system analysis, the simple linear regression algorithm (with the lowest error), linear regression, Multilayer Perceptron algorithms (MLP) and Gaussian process (with the greatest errors) with a resolution of 0.9095, 0.9039, 0.8623 and 0.7335 was achieved. Among the results, minimal error, 54.32, was belong to linear regression algorithm simple and the greatest error, 70.79, was related to MLP algorithm. In the assessment section by Artificial Neural Network (ANN) with a different number of neurons (4 to 20) was also studied. According to the evaluation data, it was concluded that image processing techniques to estimate the growth of Nannochloropsis at different salty levels were generally successful

Retina-based identification system is a stable and reliable biometric system that does not have problems like oblivion, being exposed, lost, and forgery. In this paper, a new method with high precision performance is presented which, unlike other retina-based methods, is resistant to rotation and displacement of the vessels. The system consists of three main stages of pre-processing, extraction of features and matching of features. Extraction of vessels is carried out using mixed continuous wavelet transform method. At the extraction stage, using local pattern detection operator, the characteristic of the special points of the vessels, such as the number of branches, the smallest angle and the type of points (branch and intersection), is identified. In the matching of features step, the number of special matching points and the degree of similarity of the topology of the graph resulting from the connection of the feature points of the input image and the reference image are calculated and ultimately the identity of the individual will be detected. The proposed method has been implemented on the DRIVE, VARIA, DIARET and STARE databases, which it’s accuracy was100% , 99.81%, 99.7% and 100% respectively.