This PDF file contains the front matter associated with SPIE Proceedings Volume 11431, including the Title Page, Copyright information, Table of Contents, Author and Conference Committee lists.

This paper proposes a real-time FPGA-based architecture of improved ORB. It proposes a strategy of redistribution of ORB feature points, which solves the problem of sorting FAST points of the whole image by response score. Besides, a strategy for offline generation of rBrief point pair patterns is proposed, which avoids online rotation of neighborhood pixels of feature points. These two strategies greatly reduce the resource consumption and processing clock cycles of the whole architecture. What’s more, the data throughput of the feature extraction step and feature description step is maximized, and finally a completely pipeline architecture is obtained. Due to the tips for parallel processing and resource reuse, the hardware implementation of the proposed architecture costs very few resources and processing cycles. The experimental results show that this architecture can detect feature and extract descriptor from video streams of 1280x720 resolution at 161 frames per second (161 fps), and the extracted ORB features perform well.

Blind image deblurring is a challenging problem which has drawn a lot of attention in recent years. Previous work states shows that image details caused by blur could adversely affect the kernel estimation, especially when the blur kernel is large. In this paper, we focus on how to extract the suitable salient structure for kernel estimation from a single blurred image. A fast method for estimating the salient structure of an image is proposed in the paper. The image is divided into two layers with different smoothness, and the local relative smoothness layer eliminates the image structure that adversely affects the kernel estimation. Further kernel estimation using the layer can obtain more accurate results. Substantial experiment shows that our method is effective on some challenging examples.

At present, there are more and more urgent demands to realize image/video adaptive enhancement in many different fields. In the case of large data, it is of great practical significance to study how to remove redundant noise from image/video, design a denoising image restoration technology based on large data antagonism generation network, and realize image/video enhancement technology in many fields. This paper mainly studies the further improvement and optimization of GAN, including image denoising oriented GAN model construction, GAN model improvement and training optimization, mobile phone image enhancement based on large data, etc. The experimental results show that GAN network denoising technology has been successfully applied in many image processing applications.

A synthetic aperture de-occlusion algorithm based on the microlens array (MLA) is proposed. This kind of light field camera could obtain the original image data. This paper utilized the fusion synthetic aperture technique to identify occlusion foreground information. Moreover, OTSU was used to distinguish the pixel value range of the occlusion and the target object. The obscured object data can be identified. The image obtained by removing the occlusion information was refocused. The light field camera was utilized to extract the target image. Experiments show that the target image of this proposed algorithm is much better compared with other algorithms. The proposed algorithm in this paper also presents 3D scene images with high contrast and SNR. The evaluation value of unreferenced images is also increased by 20.03% on average.

This paper presents a new algorithm to compute correlation in digital domain by transforming correlation formula into a calculation form based on a first-order moment. As a result, the arbitrary-length digital correlation can be implemented efficiently through rapidly computing the first-order moment in this new correlation formula. It is acknowledged that correlation’s computation performance depends on its multiplication complexity, so we introduce and improve an algorithm of first-order moment to implement correlation without multiplication and through an iterative procedure. Also, a systolic array without multiplier is designed for correlation’s hardware implementation according to the proposed algorithm. The comparisons with some algorithms have proven this algorithm’s efficiency

Deep convolutional neural networks (CNNs) have contributed to the significant progress of the single image super resolution (SISR) field. However, most of existing CNN-based SR models require high computing power, which is not conducive to daily use. In addition, these algorithms need to use a large number of CNN to obtain global features. Therefore, this paper proposes an image super-resolution framework based on adaptive residual neural network, using the adaptive framework to switch between global and local reasoning for internal features in a flexible way, it can extract a large number of global features without neglecting key information, which is conducive to the comprehensiveness of residual images. After the adaptive block, SENet is added to conduct channel modeling for the extracted features, and the importance of each feature channel is automatically acquired by learning method. Then, according to this importance, useful features are promoted and those that are not useful for the current task are suppressed. In this way, with more nonlinearity, the complex correlation between channels can be better fitted, and the number of parameters and computation can be reduced, which can improve the performance of super resolution to a certain extent.

Deep convolutional neural networks are increasingly used in various parallel embedded platforms such as mobile GPUs, AMD APUs, and FPGAs. At the same time, many new models have been developed for embedded platforms, such as MobileNet. In order to balance accuracy, speed and resource requirements and achieve cross-platform versatility, we have developed a software framework for in-depth research. Generated an OpenCL code that takes full advantage of parallel resources and improves the parallel efficiency of OpenCL code. Another advantage is that it optimizes and consolidates the network and compiles offline, making the entire application most efficient. MobileNets uses nonlongitudinal separable convolution (deep separable convolution) instead of standard convolution. Experiments with MobileNet have shown that the OpenCL code generation framework can significantly improve the efficiency of use.

The presence of enlarged lymph nodes is a signal of malignant disease or infection. Lymph nodes detection plays an important role in clinical diagnostic tasks. Previous lymph nodes detection methods achieve high sensitivity at the cost of a high false positive rate. In this paper, we propose a method that helps reject false positives. Features are extracted separately from 2D CT slices by using a deep convolutional neural network with multi-view input. Separated feature layers can extract the most suitable features from each input slice individually. We validate the approach on a public dataset and improve the sensitivity by reducing the false positive rate.

Recent progress in deep learning, especially deep convolutional neural networks (DCNNs), has led to significant improvement in natural image classification. However, research is still ongoing in the domain of medical image analysis in part due to the shortage of annotated data sets for training DCNNs, the imbalanced number of positive and negative samples, and the difference between medical images and natural images. In this paper, two strategies are proposed to train a DCNN for pathological lymph node image classification. Firstly, the transfer learning strategy is used to deal with the shortage of training samples. Second, a novel exponential loss function is presented for the imbalance in training samples. Four state-of-the-art DCNNs (GoogleNet, ResNet101, Xception, and MobileNetv2) are tested. The experiments demonstrate that the two strategies are effective to improve the performance of pathological lymph node image classification in terms of accuracy and sensitivity with a mean of 0.13% and 1.50%, respectively, for the four DCNNs. In particular, the proposed exponential loss function improved the sensitivity by 3.9% and 4.0% for Xception and ResNet101, respectively.

Automatic and accurate segmentation of bladder walls and tumors in magnetic resonance imaging (MRI) is a challenging task, due to significant bladder shape variations, strong intensity inhomogeneity in urine and very high variability across tumors appearance. To tackle such issues, we propose to leverage the representation capacity of an improved U-Net networks using stacked dilated convolutions. The proposed structure includes stacked dilated convolutions to increase the receptive field without incurring gridding artifacts. In addition, we embed stacked dilated convolution network into the U-Net architecture, thus enabling extracting multi-scale features for segmentation of multi structures with different shapes and scales. Finally, we apply a focal loss function to make all classes contribute equally to the loss function in our model. Evaluations on T2-weighted MRI show the proposed model achieves a higher level of accuracy than state-of-the-art methods, with a mean Dice similarity coefficient of 0.95, 0.81 and 0.66 for inner wall, outer wall and tumor region segmentation, respectively. These results demonstrate a strong agreement with reference standards and a high performance gain compared with existing methods.

X-ray angiograms, which suffer from low-contrast and noise, need to be improved by both the image enhancement and denoising techniques. However, the goals of these two tasks usually conflict, which makes it difficult to efficiently combine the enhancement and denoising in one scheme. To solve this problem, we propose a novel spatial-frequency filtering (SFF) scheme to simultaneously enhance and denoise low-quality X-ray cardiovascular angiogram images. The proposed scheme includes three key components: Firstly, a relative total variation method is employed as a guide filter to separate an input image into two parts, including the base layer with strong structures and the detail layer with weak structures and noise. Then the base layer is enhanced by a proposed improved histogram equalization (IHE) method while the detail layer is extracted by a short-time Fourier transform and is further enhanced by using a proposed adaptive correction parameter. Finally, the improved image is the combination of results obtained by the two components. Both quantitative and qualitative results of experiments on real-world low-quality X-ray angiogram images demonstrate that the proposed method outperforms the state-of-the-arts in terms of contrast enhancement, structure preservation, and noise reduction.

Medical image have the characteristics of the complex overlapping of organ and tissue, and accompanied by noise, local volume effect, artifact. So the traditional segmentation method is not ideal. To solve this problem, a medical image segmentation algorithm based on tree-structured MRF in wavelet domain (WTS-MRF) was proposed. For expressing medical image information. WTS-MRF model defines the same tree structure at every scale of wavelet decomposition. At the same time, wavelet transform has good directional selectivity, non-redundancy and multi-scale characteristics. Multiscale and multi direction expression by wavelet decomposition improved the ability of TS-MRF to describe the non-stationary characteristics of images. Then, it can more accurately describe the statistical characteristics of images, and effectively extract the feature information of medical image. In the WTS-MRF model, there are two structures in the layer TS-MRF structure and the interlayer four fork tree structure of wavelet coefficient. The TS-MRF model is built in the layer, and the node potential function is modeled by Potts model. The Gaussian model is used to build the model for the observed characteristics with the same label. The interlayer wavelet coefficients have the property of first-order Markov. The maximum posterior probability is obtained by recursive operation, and the classification hierarchy tree label is implemented to realize medical image segmentation. the experiment results indicate that the algorithm not only can effectively extract the details but also can relatively completely extract target area of medical image, and has higher segmentation accuracy and robustness．

Myocardial ischemia or coronary artery disease can be identified and located by analyzing the movement and deformation of the heart. Therefore, to accurately and non-invasively diagnose the location and extent of ischemic or infarcted myocardium, it is of great practical significance to quantitatively determine the motion/deformation parameters of myocardial tissue. In this paper, the myocardial material parameters are used as a priori information and combined with a continuum mechanics model to restore the cardiac cycle motion under the spatial constraints of the graph total variation (GTV). In the motion reconstruction, the biomechanical model establishes the relationship between stress and deformation through system dynamics. The total variation of the graph proposed in this paper ignores the spatial distance, establishes the connection between similar regions in the image, overcomes the limitation of considering only the similarity with adjacent regions, and preserves the texture details and fine structure. Because GTV uses the K-nearest neighbor algorithm (KNN) to classify regional similarity, the connection between similar regions is stronger, therein achieving computational scalability and lower computational complexity. The accuracy of the strategy with and promising application results from synthetic data, magnetic resonance (MR) phase contrast, and gradient echo cine MR image sequence are demonstrated.

Recently, diagnosis, therapy and monitoring of human diseases involve a variety of imaging modalities , such as magnetic resonance imaging(MRI),computed tomography(CT),Ultrasound(US) and Positron-emission tomography(PET)as well as a variety of modern optical techniques .The degeneration of lumbar intervertebral disc has become a common disease in modern society. Currently, the most commonly used method is the diagnostic grade based on MRI technology, among which Pfirrmann grading system is most widely used in clinic. The Pfirrmann grading system is mainly based on the measurement of the average height of the lumbar intervertebral disc and the intensity of the signal of the nucleus pulposus and the inner and outer edge of the fiber ring in MR images. With the degeneration of the intervertebral disc, the signal of the inner and outer edge of the annulus also decreases, so the error caused by the method of measuring the average height of the lumbar intervertebral disc is larger. Therefore, we proposed an algorithm based on morphology to detect lumbar intervertebral disc in MRI spinal images. First, the median filter is used to remove noise in MRI and then the lumbar intervertebral disc is extracted through morphological processing. Then, the image is smoothed by combining with gaussian filtering. Finally, the result map of lumbar intervertebral disc is obtained and its area is calculated. In the analysis and comparison of the detection results of the lumbar intervertebral disc, the skeleton extraction diagram of the detection results of the lumbar intervertebral disc was obtained after processing the image of the detection results of the lumbar intervertebral disc with the thinning algorithm. According to the analysis, the degree of laminar disc skeleton and upper and lower vertebral body is as high as 90%. This paper also briefly introduces the application direction of this measurement algorithm in medicine: 1. Improve doctors' ability to detect early lumbar disc degeneration.2. Assist doctors to observe postoperative recovery of patients.

Due to its unique advantages, infrared temperature measurement technology has been used more and more widely in medical auxiliary diagnosis, and the change of ambient temperature is an important factor affecting medical infrared temperature measurement components. This paper builds a simulation platform for ambient temperature. At a fixed distance, by changing the ambient temperature and the target temperature, a large amount of infrared temperature measurement data is collected, and a two-dimensional compensation model of the scene temperature variation is obtained by combining the infrared focal plane non-uniformity correction and the least squares fitting. The actual scene experiment proves that the temperature measurement accuracy and stability have achieved satisfactory results.