MACHINE VISION AND IMAGE PROCESSING
Year:2020 | Volume:6 | Issue:2|Papers Count:12

One of the effective methods for visual recognition (including classification, object recognition and image semantic labeling) is to identify probable regions including an object that is known as object proposals. In this paper, an effective approach is proposed relying on identifying appropriate regions based on image segmentation which is called SDE (superpixel dual extension). The proposed approach comprises of two phases. In the first phase, using a bottom-up segmentation algorithm, an image is presented by some regions as superpixels. In the second phase each superpixel is then extended to adjacent regions, according to a set of predefined states and the 8-connectivity. The most important advantage of this extension is to generate regions that are able to completely surround an object. Using descriptors such as color, texture and keypoints for feature extraction resolves computer vision problems and enhances the performance. Here, a set of well-known metrics of image segmentation including overlap, recall, area under curve, and pair of pixels’ coherency are measured in order to precisely assess the proposed method. Furthermore, to more evaluate the effectiveness of the method a classification problem on MSRC dataset is carried out. The results are shown quality enhancement around 7% for graph-based segmentation and 14% for clustering-based segmentation. Moreover, 11% improvement in accuracy of image classification is also achieved.

This paper presents a new image watermarking method in Contourlet domain, where watermark is embedded through quantum chaotic map in the selected Contourlet coefficients of the cover image. The main idea of this method is data embedding in selected sub bands, providing higher resiliency through better spectrum spreading compared to other sub bands. The proposed method leads to higher imperceptibility and robustness against several common watermarking attacks such as compression, adding noise, filtering and scaling. It has been observed that, in comparison with other Contourlet based and wavelet based methods, the proposed method is thanks to its capability in directional selectivity.

There are challenges such as depth perception and self-occlusion, in the field of 3D human pose estimation and reconstruction which obstructs precise estimation of body joints. In this paper, we first extract human pose by focusing on 2D ground-truth using sparse coding and. In the second approach, we use a learning-based Convolutional Neural Networks using sparse coding and a model based rectifier to extract the estimated pose. Pose estimation by proposedmethod has reduced the mean error of the reconstruction in comparison with the state of the artworks.

Today, the detection and annotation of objects in images is one of the major challenges in some applications of machine vision. In recent years, the use of deep learning has attracted the attention of researchers. In this regard, this paper first introduces the newest deep networks and analyzes the strengths and weaknesses of these methods. An improved network of R-FCN network has been presented. The proposed method is based on the ResNet architecture and the fully-convolutional network. In this method, a new architecture is proposed based on region proposal deep network and a combined method based on the binary fuzzy SVM and the SVR for final detection and categorization of objects. Also, a new loss function called Cauchy-Schwartz Divergence loss, has been used. This function has shown better performance in terms of speed and accuracy. The proposed ResNet-101 architecture was tested on the SUN dataset for the detection and annotation of 36 objects, and the results indicate improved performance of this method compared to the basic R-FCN network method. The proposed method, In terms of Mean Average Precision, has 48. 38% performance and average duration for each image is 0. 13 Compared to the best method in this area, it performed about 2% in performance and 0. 04 seconds in better time.

Muck-pile size distribution is one of the most important parameters in open pit blasting that can affect mining and mineral processing efficiency. For evaluating fragmentation by blasting, digital image analysis is a fast and reliable indirect technique. In this study, based on neural network and visual feature extractions, an algorithm was developed to determine muck-piles size distribution using digital images. 26 test images of fragmented rockwere used to determine size distribution and the results were compared with the results of automatic and manual aged detection of Split-Desktop software. The results showeda general improvement in evaluating rockparticles size distribution. We obtained an improvement of 67%, 57% and 28%, respectively using Fourier transform, Gabor and wavelet methods. Fourier transform, Gabor and wavelet methods showed also an improvement of 52%, 40% and 21 %, respectively in evaluating of F10 to F50.

Graph based semi-supervised methods for automatic image annotation are mainly focused on single-label problems. However, most of the real world problems require multiple labels per image. As a hybrid semi-supervised approach, LGC+ML-KNN is proposed for multi-label image annotation. LGC is a graph based semi-supervised learning algorithm that annotates unlabeled samples. Subsequently, ML-KNN learns from many more labeled samples, as compared to the initial training set. Experiments on several datasets confirm that the proposed approach has better accuracy than available methods, especially when a very small portion of the training set are the labeled samples.

Watermarking is a method for embedding a watermark in a digital image to preserve its copyright. Watermark should have two contradictory properties of transparency and robustness. Locations of embedding watermark in the image and threshold values play an important role in improving these two properties. In this paper, a novel robust image watermarking scheme in hadamard transform domain is proposed which uses Hybrid Firefly Algorithm (HFA) for selecting suitable blocks and threshold values to balance transparency and robustness. HFA is the modified version of firefly algorithm which is proposed and used in this paper for first time. This algorithmgenerates distinct and discrete values for selecting blocks and simultaneously continues values for selecting threshold values. Hadamard transform applies on each selected block of the cover image and four watermark bits are embedded in one block using selected threshold values. Watermark detection is done blindly and objective function of the optimization algorithm is a combination of transparency and robustness. This scheme is applicable to color and gray scale images. Experimental results show that the proposed scheme is highly robust against different attacks and gives better results in terms of transparency compared to other similar methods.

Blurriness is one of the common distortions in images. This distortion is caused by spilling the pixel information overthe adjacent pixels. Blurriness has different types. The knowledge about the type of image blurriness is one of the important parameters which directly affects performance of de-blurring methods. In this paper, a method has been proposed to classify the fourtypes of global blurrinessin digital imagesin the spatial domain. These blurriness include the Gaussian blur, rectangular blur, motion blur and defocus blur. In the proposed method, the correlation concept is used to classify the type of image blurriness. The correlation concept depicts the relations between the image pixels. Also, the model and correlation of adjacent pixels are proportional to the type of blurriness. Appropriate features are extracted to detect the type of blurriness. The accuracy of the proposed method for detecting the type of blurriness is 90. 4%. This method has a better performance compared to other existing methods in terms of accuracy and computational cost.

The content based image retrieval searches digital images in a large image database and uses visual content of images instead of metadata. This approach has many usages in security and authentication for example scanning the iris, fingerprint or finger knuckle print. This paper contributes a new method for personal authentication using finger knuckle print based on a new local binary pattern and image segmentation. The capabilities of quantum science lead to take its advantage in different areas of image processing. The main idea is inspired by the theory of quantum estimation and is applied to the feature extraction phase, in addition, the quantum circuit of the proposed feature is also designed. In order to measure the efficiency and accuracy of the proposed method, the EER (Equal Error Rate) is calculated. After implementing the proposed algorithm on the POLYU dataset, which contains 7920 images, the EER = 0. 67 and accuracy =99% are obtained which indicate that the new method has more efficiency and accuracy than the similar approaches.

Facesketchsynthesis of example image plays an important role in both digital entertainment and law enforcement. In this paper, face sketch synthesis has two main processes. In the first process, neighbors are selected and in the second process, reconstruction weight representation is done. Running time and computation complexity depends on neighbor patches selection process. Face sketchgeneration with state-of-the-art methods perform neighbor selection process in a data-driven manner by K nearest neighbor searching. Hence, the running time for synthesisincreases. Also, for neighbor selection need to check the whole training dataset. As a result, the computational complexity increases with the scale of the training database and is limited scalability. In this paper, we proposed a simple but effective with encoding local binary pattern and random sampling in place of pixels. Then by extracting shape from resulting textures and determining state of surfaces, we represent facial sketch. Our experiments onpair of CUHK database imagesdemonstrate the proposed method in comparison to state-of-the-art methods has superiorityof view generated sketch quality and running time. Also, the proposed methodin front of face hallucinationproblemswhich cause heterogeneous transformation on facial sketch is resistant.

Texture and color are two important attributes for object recognition. Recently, quaternionic representation of color images have been used as an effective method for color image processing. Using such a representation, it is possible to consider the mutual interaction between different color channels. In the last decade, several quaternion operations like rotation, reflection, and Clifford translation have been developed. Such operators are able to extract shallow information from the color images. In this paper, we first propose a set of new quaternion operators called hybrid quaternionic operators, which can be produced by a cascade of several simple quaternionic operators. Such operators can extract deeper information from the color images. We then use such operators, and present a novel color texture classification method using the concept of sparse coding. Experimental results indicate that the proposed method outperforms several existing and popular methods.

In this paper, a new method for automated detection of a human region of interest is provided that makes use of camera surveillance in department stores. In this work, a region of interest is an area in the image where more people commute. For this purpose, first humans are distinguished from other objects in the image utilizing a histogram of oriented gradients (HOG) descriptors. Every detected individual is considered as an event in the image. Then, a non-parametric distribution based on Bayesian risk is applied to obtain the most interested regions from the position of detected humans. In the proposed distribution, a new high-efficiency kernel is provided. In Bayesian risk, a novel loss function is proposed that has a higher accuracy in compared with square loss function and performs better in finding peaks of a distribution function. For the evaluation, data from live surveillance cameras located in different parts of some stores are used. For the proposed kernel, on average, an accuracy of 85% and for the loss function, an accuracy of 93. 5% on artificial data and 90% on real data are acquired which are better results in compared with other similar works.