نشریه ماشین بینایی و پردازش تصویر
سال هفتم شماره 1 (بهار و تابستان 1399)

Nowadays, a huge amount of images are produced by digital cameras. However, various reasons such as weakness in the design of the camera lens, lead to the creation of noisy images.The image enhancement methods sharp some componentsof images such as the edges in order to increase the resolution of the input images.But, sharpening of the images' edges results in increasing the noise. Hence, in order to reduce the blurriness of images, employing of sharpening techniques should be applied under controlled conditions in order to prevent loosing images details. In this paper, a new method is proposed to reduce the blurriness of digital images. The proposed method is combination of Relative Total Variation filter (RTVf) and Rolling Guidance filter (RGf) in HSV color space. In the proposed method, the image structure is extracted by using RTVfand then the images' edges are retrieved by using RGf. Then, the image details are extracted by subtracting the V channel of input image from the result image of RGf. During a repetition process, image's details are added to image V channel.  In this method, the intensity of the image pixels does not change to a similar ratio, which results in better display of the image details and prevents increasing noises.The proposed method has been tested on benchmark images.The achieved results show that the proposed method achieved to 47% of reducing blurriness, 85% of noise controlling, and 83% of saving naturalness of input images in compared to rival methods.

The aim of digital image quality assessment is to provide a model predicting the human judgments of viewing scenes. In this paper, a color image quality measure is presented in which a combination of three components including gradient magnitude, phase congruency and visual saliency is utilized for a better prediction. This combination is designed based on the fact that the human visual system extracts the low-level image features. The phase congruency and visual saliency provide robust and contrast invariant structural information of the viewing scene. On the other hand, gradient magnitude captures all changes including the local contrast. Combing these three features can enhance the assessment of local quality. In order to obtain a single measure, the visual saliency is employed in the pooling phase. The experimental results demonstrate that the proposed algorithm can effectively evaluate natural images quality in a consistent manner with the human visual perception.

Human visual system can recognize object accurately, swiftly, and effortlessly even when objects are under challenging conditions. Many research groups try to model this ability; however, these computational models could not achieve human performance. Convolutional neural networks (CNN’s) are the state-of-the-art successful computational vision models that try to implement feedforward path of human visual system. However, evidence shows that human visual system uses top-down expectation signals to increase accuracy and speed of object recognition under dificult conditons. In this study, we extend a well-known model using top-down expectation signals. In this regard, Alexnet network is considered as feedforward path. We used a pre-trained network on ImageNet dataset for object recognition and a pre-trained network on Places dataset for scene recognition. The pre-trained network on places was used to provide top-down feedback signals based on scene information. The feedback signals contain occurrence frequency information of the objects in the scene. These signals are integrated with information from feedforward path. To evaluate the proposed model several experiments were done on different image sets. The results showed that integrating the feedback information with the feedforward information significantly improve object recognition accuracy in comparison to the base model. This support the idea that content information facilitates object recognition ability, specifically when objects are under challenging conditions.

In this paper, an effective and shadow resistant method is provided to automatically estimate the speed and dimensions of vehicles using video received from a surveillance camera. In this method, at first by examining a few initial frames and considering the motion of vehicles, the vanishing points and focal length of the camera are obtained. Then, by identifying the foreground and removing the shadow, the precise boundary of each vehicle is determined and the 3D bounding box is created for each vehicle. After projecting car on a hypothetical road and eliminating the perspective, the metric coefficient (pixel to meter) is calculated according to the actual dimensions of the dominant car. Removing the perspective and using the metric coefficient allows estimating the speed and dimensions of cars in each frame. But to reduce the error, by tracking the cars, histograms are made for the speed and dimensions of each vehicle. Then the maximum of these histograms is reported as the speed and dimensions of each vehicle. Experiments show better results compared with previous works.The maximum error for the test sets in the speed estimationis 1.17 km/h and in the dimension estimation it equals to 2.6%.

In camera-based motion tracking, the target structure has a significant impact on the tracking accuracy. In the present research, a new method is proposed for the optimization of the location of five markers on a visible target for the camera. The proposed objective function for positioning of markers is equal to the total distance of each marker from all of the planes formed from the combination of triads made of other markers. To avoid the symmetry of the target structure, which makes the labeling of markers on the image impossible, a constraint is applied that guarantees a minimum difference in the distance between the pairs of markers. The genetic algorithm is exploited for maximizing the proposed objective function with the mentioned constraint. The experimental results obtained from the positioning of markers on the target using the proposed method (with or without the application of the constraint) and other methods are evaluated. These methods are compared in terms of error in estimating 3D pose and tracking speed in the presence of noise. It is notable that the results confirm the applicability of the proposed method in the positioning of markers.

Classification of brain tumors using MRI images along with medical knowledge can lead to proper decision-making on the patient's condition. Also, classification of benign or malignant tumors is one of the challenging issues due to the need for detailed analysis of tumor tissue. Therefore, addressing this field using image processing techniques can be very important. In this paper, various types of texture-based and statistical-based features are used to determine the type of brain tumor and different types of features are applied in this classification procedure. Sparse non-negative matrix factorization algorithm is used to learn the over-complete models based on the characteristics of each data category. Also, sparse structured principal component analysis algorithm is applied to reduce the dimension of training data. The classification process is carried out based on the calculated energy of the sparse coefficients. Also, the results of this categorization are compared with the results of the classification based on the neural network and support vector machine. The simulation results show that the proposed method based on the selected combinational features and learning the over-complete dictionaries can be able to classify the types of brain tumors precisely.

Sparse correntropy model is a face recognition model on the bases of sparse representation which is robust to noise and occlusion. In this mode, a linear combination of training data is determined such that, on the basis correntropy criterion, is as similar as possible to the test data, and L1-norm of coefficient vector of the linear combination is minimum. L1-norm is not differentiable. Therefore, efficient gradient-based methods can not be used to solve the problem. Thus, to simplify the model to be solved fast, the coefficients were considered to be non-negative. The non-negativity constraint is restrictive which can decrease the accuracy of the model. In this paper, to fix this difficulty, L2-norm instead of L1-norm of the linear combination is minimized. Then, a fast algorithm is proposed to solve the novel model. Experimental results confirm that the runtime and accuracy of our proposed method is better than that of sparse correntropy model with non-negative coefficients.

Self-learning super-resolution is an approach for enhancing single-image resolution. In this approach, instead of using the external database for learning the relation between low and high resolution image patches, only relation between patches in the input image pyramid are used for learning. In this paper, a novel self-learning single image super-resolution method by focusing on the organization of the low and the corresponding high-resolution information has been presented. In order to provide training data the low-resolution and the corresponding highresolution images are created by down-sampling and up-sampling of the input image in two image pyramids. In this paper, unlike most prior super-resolution methods, the images in the low-resolution pyramid are segmented and then used for the process of super-resolution. Another remarkable point in this paper is dividing all the images of different levels of the pyramid into the same numbers and similar regions. This is done by segmenting the image at the lowest level of the pyramid and generalizing its regions to the higher-level of the pyramid images. Due to the different number of regions in each input image, the number of training models of the proposed method is different for each image and depends on the content of the input image. The result of the experiments shows that the proposed method is quantitatively and qualitatively improved the previous methods.

In this paper, we are going to classify each pixel of a hyperspectral image. For this purpose, we group the spectral bands to sub-bands and try to decompose the corresponding sub-tensors to the endmember and abundance matrices. Abundance matrices obtained through tensor factorization methods contain spatial information in contrast to the ones acquired by matrix factorization. Therefore, the 2D abundance maps achieved by tensor decomposition methods, construct discriminant features for the classifier. A 3D CNN architecture is proposed for classification which utilizes the abundance maps of the individual sub-bands as input features. This way, we jointly exploit spectral and spatial information of the image. The experiments are performed on well-known hyperspectral data and reveal the effectiveness of the proposed sub-band tensor decomposition methods compared to matrix factorization approaches.

Wide-baseline image matching with significant viewpoint differences plays a fundamental role in many computer vision and photogrammetry applications, such as 3D reconstruction and image registration. One of the main problems of matching these images is the existence of a relatively large number of mismatches. Generally, a geometric consistency check process based on various geometrical constraints and robust estimator methods such as the epipolar line and RANSAC algorithm is used for mismatch elimination. However, conventional geometry filtering methods in wide-baseline images will fail if the number of outliers is very high. In addition, these methods have high computational complexity. In this paper, a novel mismatch elimination approach in wide-baseline images with significant viewpoint differences is presented. First, initial elliptical features are extracted using improved MSER (maximally stable extremal regions) detector in both images. Then, a distinctive DAISY descriptor is generated for each extracted feature. In the next step, the initial feature correspondence process is established using Euclidean distance between feature descriptors. Then, a novel mismatch elimination approach based on features shape matrix, named MESM (mismatch elimination based on shape matrix), is applied. Finally, the few remained blunders are removed by using a geometric constraint. The proposed image matching and mismatch elimination algorithms were successfully applied to match eight close-range image pairs with significant viewpoint differences, and the results demonstrate its capability to improve matching performance.

Semantic image segmentation based on Convolutional Neural Networks (CNNs) is one of the main approaches in computer vision area. In convolutional neural network-based approaches, a pre-trained CNN which is trained on the large image classification datasets is generally used as a backend to extract features (image descriptors) from the images. Whereas, the special size of output features from CNN backends are smaller than the input images, by stacking multiple deconvolutional layers to the last layer of backend network, the dimension of output will be the same as the input image. Segmentation using local image descriptors without involving relationships between these local descriptors yield weak and uneven segmentation results. Inspired by these observations, in this research we propose Context-Aware Features (CAF) unit. CAF unit generate image-level features using local-image descriptors. This unit can be integrated into different semantic image segmentation architectures. In this study, by adding the proposed CAF unit to the Fully Convolutional Network (FCN) and DeepLab-v3-plus base architectures, the FCN-CAF and DeepLab-v3-plus-CAF architectures are proposed respectively. PASCAL VOC2012 datasets have been used to train the proposed architectures. Experimental results show that the proposed architectures have 2.7% and 1.81% accuracy improvement (mIoU) compared to the related basic architectures, respectively.

The Scale Invariant Feature Transform (SIFT) algorithm is one of the most widely used algorithms in the machine vision field on which researchers have extensively studied and improved. SIFT is one of the common local detectors used in image registration, image mosaicking, copy-move image forgery, and etc. In this review paper, along with introducing the SIFT algorithm, the applications, pros and cons, modifications, categories and new research approaches in this algorithm are discussed. In addition, via four experiments, different aspects of this algorithm have been evaluated. This paper can help image processing researchers when utilizing the SIFT algorithm. The aim of the authors was to explore all the aspects of this algorithm.