MACHINE VISION AND IMAGE PROCESSING
Year:2018 | Volume:5 | Issue:1|Papers Count:10

In this paper, a new approach to weakly supervised semantic segmentation is proposed. The main goal in semantic segmentation is to assign a semantic label to each pixel. In weakly supervised setting, each training image is only labeled by the classes they contain, not by their locations. The main contribution of this paper is to simultaneously incorporate the object level and context level information in assigning class label to each pixel of the image. To do this, regions in each image are grouped such that groups of regions in images with the same semantic label have the same appearance and context. To do this, an iterative move-making algorithm is proposed. At first, each pixel of the image is initially labeled and then model of appearance and context for each class label is learned. Then, semantic label of each pixel is updated such that the regions with the same sematic label have the same appearance and context in the set of images. In the next step, appearance and context models for each semantic class are updated. It is repeated until in the two consecutive epochs, labels of the pixels are not changed. To evaluate our proposed approach, it is applied on the MSRC dataset. The obtained results show that our approach outperforms comparable state-of-the-art approaches.

Electromyographs are used for electromyography signal extraction from neurologically activated muscle cells. These signals are investigated to extract discriminating patterns to be categorized in the classification stage of myoelectric control systems (MCSs) designed for various applications. Feature extraction is a fundamental step in EMG signal processing which affects the overall performance of MCSs. To improve classification accuracy of MCSs, this paper proposes a novel approach for feature extraction from time-frequency images of EMG signals using local binary patterns and gray level co-occurrence matrices. In contrast to time alone and frequency alone approaches, by textural analysis of EMG signal spectrogram, time-frequency patterns of these signals are revealed, simultaneously. Furthermore, LBP and GLCM expose relational properties of time-frequency patterns which areexploited as the main features for classification. EMG physical action dataset is utilized in this study to evaluate the proposed method. In the classification stage, support vector machine classifiers are used in two segmented and holistic modes. The best classification accuracy of 98. 75% is obtained by segmented approach which is superior to the results provided by state of the art methods.

In the field of diagnosis and classification of animals there are always many problems which prevents the development of rapid and effective progress in this area. In recent years, the new approaches have been proposed that are based on artificial neural networks and image processing that can detect and recognize the butterfly types. In this article, specifically, we'll scrutiny the butterfly species detecting by using image processing and smart classification methods, also we are looking for the performance improvement by employing texture of butterfly wings features. In this context, the quantization feature extraction method of the local phase is used that resist against the blur in the butterfly pictures. As well as, in order to classification, the MLP and wavelet neural network is used that the result demonstrates, the wavelet neural network achieve 100% classification accuracy in 14 butterfly species.

Refining image annotation is an effective approach to improve tag base image retrieval. Many images in social networks and search engines have vague tags, incomplete and irrelevant content. However the unreliable tags, reducing the precision of image retrieval, recently some of the tag refinement (TR) algorithms have been suggested as labels noise removal and enrichment of images. In order to achieve optimal result in TR, extracting features that have a good description of visual content of images will have direct impact on accuracy of TR process. Achieving the appropriate description and relevant to the content of images, is the major challenges in the refining image annotation. Due to effectiveness of deep learning in research fields, in this paper we will use deep convolutional neural network (DCNN) in order to extract efficient features for computing images visual and semantic similarity. Employing transfer learning based ImageNet image database in DCNN, for large scale NUSWIDE dataset, indicating the effectiveness of this approach in refining image annotation.

Proper spatial resolution has great importance in many image types since it conveys significant details. Performance of feature extraction methods, in some image types such as textual, facial, and fingerprint ones, highly depend on the image quality. Spatial resolution is one of important factors affecting the image quality; but, high spatial resolution increases the storage memory of the corresponding images, showing the importance of image compression methods. In the proposed image compression approach of this paper, dimension of the input image is first decreased based on the wavelet transform and then is compressed using any desired image compression method. In the decompression stage, first, the dimensionally reduced image is reconstructed and then, the initial dimension is restored by our proposed technique of estimating the detail sub-bands in the wavelet domain. In the evaluation stage, we chose two image types of textual and facial as two case studies having band-pass and low-pass spectrums respectively. We evaluated the compression and recognition performance of proposed approach by combining it with any of conventional compression methods of JPEG, JPEG2000 and SPIHT. Simulation results showed the noticeable effect of the proposed approach on reducing the storage memory and simultaneously, preserving o the compression/recognition performance.

Digital image watermarking is a common topic in the field of information security and image abuse prevention on the internet and communication areas. One of the applications of digital watermarking is authentication and recovery tampered region. These methods can discover correctness and integrity of the received image by using information which is embedded. In this paper, a method is proposed to identify and recover tampered regions in both color and grayscale images. The proposed method provides a second chance for recovery of tampered region, based on the ability of dual watermark embedding in the image. In addition, due to use of discrete cosine transform, it improves robustness against compression. In order to guarantee the security of embedded watermark, a chaotic map is used with a secret key which is transferred with the image. In this proposed method, to prevent copy-move attack, the information which is embedded for detection of tampered block depends on its own key. The experimental results show that the proposed method can correctly identify the tampered region under a compression with a quality factor of more than 30. This means the reduction of the false-positive error, and also the recovery of the tampered regions where half of the image is destroyed with structuralsimilarity index about 0. 9.

Nuclear medicine planar imaging is the most important medical imaging methods in detecting of lesions and abnormality of tissues and their functions. Analysis and interpretation of the nuclear medicine images plays an important role in diagnostic medicines. The images usually have low contrast, high noise and small sizes in the injury region. Abnormal region identifications are depended to images quality and resolution. In this research, the dual domain method is used and tried to improve the quality of nuclear medicine planar images. In nuclear medicine, noise usually has a high-frequency component and it seems that removing the frequency components with other contrast enhancement algorithms can be useful in noise removal. For 46 chosen images from kidney and other part of body, the dual domain method was applied. The images were very noisy that the contrast was improved by the method. Comparisons between the images show that the dual domain method by eliminating high frequency component of image can be considered as one of the most important methods for noise removal of nuclear medicine planar images. Also, the contrast enhancement method is effective for some images. For evaluation, the opinions of nuclear medicine physicians and medical physics were used. The experts’ opinions show that the quality and contrast of images have been improved significantly.

The aim of the present paper is to give an efficient scheme for denoising and edge detection of images. The main idea is based on rationalized Haar functions. Up to now, these functions were being used in solving integral equations and differential equations. But in this paper, we have used rationalized Haar functions for denoising and edge detection of the standard images. Experimental results show the accuracy of using rationalized Haar functions in edge detection and denoising.

This paper proposes a 3-D non-local means (NLMs) method for the application of video inpainting. To do this, first it assigns to the target pixels a priority and then restores them. The priority assignment of the target pixels is done based on structure and texture information of their neighbors’ pixels. To determine the type of each patch, which can be either texture or structure, the entropy criterion is used. The proposed method, to estimate damaged pixels, uses several non-local patches instead of the best matched patch. The subjective and objective experiments confirm the superiority of proposed method in the application of video inpainting and removing moving objects from the video in comparison with state-of-the-art methods.

Ever increasing number of radiology images makes their storage and classification a challenging issue. In this paper a new method for compression and retrieval of radiology images in HEVC compressed domain is introduced. In this method the radiology images are coded as I-frames in HEVC standard. The prediction mode histogram along with PU size histogram of the coded images are used for classification and retrieval of radiology images. Experimental results indicate that the proposed method achieves in average 95% accuracy in classification and P@35 of 89% for radiology images which are superior to the other methods.