Down-sampling Layers Research Articles

Research of improving semantic image segmentation based on a feature fusion model

The context information of images had been lost due to the low resolution of features, and due to repeated combinations of max-pooling layer and down-sampling layer. When the feature extraction process had been performed using a convolutional network, the result of semantic image segmentation loses sensitivity to the location of the object. The semantic image segmentation based on a feature fusion model with context features layer-by-layer had been proposed. Firstly, the original images had been pre-processed by the Gaussian Kernel Function to generate a series of images with different resolutions to form an image pyramid. Secondly, inputting an image pyramid into the network structure in which the plurality of fully convolutional network was been combined in parallel to obtain a set of initial features with different granularities by expanding receptive fields using Atrous Convolutions, and the initialization of feature fusion with different layer-by-layer granularities in a top-down method. Finally, the score map of feature fusion model had been calculated and sent to the conditional random field, modeling the class correlations between image pixels of the original image by the fully connected conditional random field, and the spatial position information and color vector information of image pixels were jointed to optimize and obtain results. The experiments on the PASCAL VOC 2012 and PASCAL Context datasets had achieved better mean Intersection Over Union than the state-of-the-art works. The proposed method has about 6.3% improved to the conventional methods.

ResWnet for Retinal Small Vessel Segmentation

U-Net shows excellent performance in biomedical image segmentation tasks, Variants of U-Net have been proposed, such as U-Net with residual blocks, densely connected convolution, or deformed convolution. However, contracting path of these variants is containing four consecutive downsampling layers and an encoding-decoding structure with skip connections. The feature map transformation paths are limited. Therefore, ResWnet is proposed in this work, the number of downsampling layers of contracting path is modified to two, and the network is changed from the encoding-decoding structure to the encoding-decoding-encoding-decoding structure so that the network can retain more detailed features and extract deeper semantic information. The skip connections are used between the contracting path and the expansion path at the same scale to increase the information transmission path. The convergence speed of the network is accelerated by the residual block replacing convolution layer. The image is scaled to different scales and is inputted into the network from the left to enhance the sensitivity of the network for blood vessels at different scales. The feature maps extracted from different layers of the network are scaled to the same size as the inputted image for fusion and segmentation. The performance of ResWnet is evaluated on DRIVE and STARE databases. The AUC is 0.9799 and 0.9863, and the accuracy is 0.9554 and 0.9723, respectively. According to the results, ResWnet outperforms many other proposed methods in small vessel segmentation. For future research, it can be learned that pre-processing and post-processing are vital to the segmentation accuracy. Furthermore, ResWnet can also be applied to other semantic segmentation tasks such as lung nodule segmentation or cancer cell segmentation. For future research, the pre-processing and post-processing will be tried to further enhance the segmentation accuracy performance of ResWnet.

AutoSegNet: An Automated Neural Network for Image Segmentation

Neural Architecture Search (NAS) has drawn significant attention as a tool for automatically constructing deep neural networks. The generated neural networks are mainly applied for image classification, and natural language processing. However, there are increasing demands for image segmentation in various areas, such as medical image processing, satellite image object location, and autopilot technology. We propose a NAS method called Automated Segmentation Network (AutoSegNet), targeting industrial and medical image segmentation. The search architectures are constructed by stacking the downsampling layer, the bridge layer, and the upsampling layer, which are explored by a recurrent neural network. Compared with other related methods for image segmentation, the proposed method has a small search space but can explore most of the-state-of-the-art supervised image segmentation models. We perform verification on two datasets, and the results show that AutoSegNet achieves superior segmentation results with clear and continuous segmented edges, as well as better image details.

M-GAN: Retinal Blood Vessel Segmentation by Balancing Losses Through Stacked Deep Fully Convolutional Networks

Until now, the human expert segments retinal blood vessels manually in fundus images to inspect human retinal-related diseases, such as diabetic retinopathy and vascular occlusion. Recently, many studies were conducted for automatic retinal vessel segmentation from fundus images through supervised and unsupervised methods to minimize user intervention. However, most of them lack in segmentation robustness and cannot optimize loss functions so that results of the segmentation have made lots of fake or thin branches. This article proposes a new conditional generative adversarial network called M-GAN to conduct accurate and precise retinal vessel segmentation by balancing losses through stacked deep fully convolutional networks. It consists of a newly designed M-generator with deep residual blocks for more robust segmentation and an M-discriminator with a deeper network for more efficient training of the adversarial model. In particular, a multi-kernel pooling block is added between the stacked layers to support the scale-invariance of vessel segmentations of different sizes. The M-generator has down-sampling layers to extract features and up-sampling layers to make segmented retinal blood vessel images from the extracted features. The M-discriminator also has a deeper network similar to the down-sampling of the M-generator, but the final layer is constructed as a fully connected layer for decision making. We conduct pre-processing of the input image using automatic color equalization (ACE) to make the retinal vessels of the input fundus image more clear and perform post-processing that makes the vessel branches smooth and reduces false-negatives using a Lanczos resampling method. To verify the proposed method, we used DRIVE, STARE, HRF, and CHASE-DB1datasets and compared the proposed M-GAN with other studies. We measured accuracy, the intersection of union (IoU), F1 score, and Matthews correlation coefficient (MCC) for comparative analysis. Results of comparison proved that the proposed M-GAN derived superior performance than other studies.

Enhancement of Multimodal Microwave-Ultrasound Breast Imaging Using a Deep-Learning Technique.

We present a deep learning method used in conjunction with dual-modal microwave-ultrasound imaging to produce tomographic reconstructions of the complex-valued permittivity of numerical breast phantoms. We also assess tumor segmentation performance using the reconstructed permittivity as a feature. The contrast source inversion (CSI) technique is used to create the complex-permittivity images of the breast with ultrasound-derived tissue regions utilized as prior information. However, imaging artifacts make the detection of tumors difficult. To overcome this issue we train a convolutional neural network (CNN) that takes in, as input, the dual-modal CSI reconstruction and attempts to produce the true image of the complex tissue permittivity. The neural network consists of successive convolutional and downsampling layers, followed by successive deconvolutional and upsampling layers based on the U-Net architecture. To train the neural network, the input-output pairs consist of CSI’s dual-modal reconstructions, along with the true numerical phantom images from which the microwave scattered field was synthetically generated. The reconstructed permittivity images produced by the CNN show that the network is not only able to remove the artifacts that are typical of CSI reconstructions, but can also improve the detectability of tumors. The performance of the CNN is assessed using a four-fold cross-validation on our dataset that shows improvement over CSI both in terms of reconstruction error and tumor segmentation performance.

Two-step deep neural network for segmentation of deep white matter hyperintensities in migraineurs

Background and ObjectivePatients with migraine show an increased presence of white matter hyperintensities (WMHs), especially deep WMHs. Segmentation of small, deep WMHs is a critical issue in managing migraine care. Here, we aim to develop a novel approach to segmenting deep WMHs using deep neural networks based on the U-Net. Methods148 non-elderly subjects with migraine were recruited for this study. Our model consists of two networks: the first identifies potential deep WMH candidates, and the second reduces the false positives within the candidates. The first network for initial segmentation includes four down-sampling layers and four up-sampling layers to sort the candidates. The second network for false positive reduction uses a smaller field-of-view and depth than the first network to increase utilization of local information. ResultsOur proposed model segments deep WMHs with a high true positive rate of 0.88, a low false discovery rate of 0.13, and F1 score of 0.88 tested with ten-fold cross-validation. Our model was automatic and performed better than existing models based on conventional machine learning. ConclusionWe developed a novel segmentation framework tailored for deep WMHs using U-Net. Our algorithm is open-access to promote future research in quantifying deep WMHs and might contribute to the effective management of WMHs in migraineurs.

Study on Traffic Sign Recognition by Optimized Lenet-5 Algorithm

Traffic sign recognition (TSR) is a key technology of intelligent vehicles, which is based on visual perception for road information. In view of the fact that the traditional computer vision identification technology cannot meet the requirements of real-time accuracy, the TSR algorithm has been proposed on the basis of improved Lenet-5 algorithm. Firstly, we performed picture noise elimination and image enhancement on selected traffic sign images. Secondly, we used Gabor filter kernel in the convolution layer for convolution operation. In the convolution process, we added normalization layer Batch Normality (BN) after each convolution layer and reduced the data dimension. In the down-sampling layer, we replaced Sigmoid with the Relu activator. Finally, we selected the expanded GTSRB traffic sign database for the comparison experiment on the Caff platform. The experimental results showed that the proposed improved Lenet-5 network test set had the recognition accuracy of 96%, which was better than the method that combined Gabor with Support Vector Machine (SVM) in terms of recognition accuracy and real-time performance.

Research on convolutional neural network and its application on medical image

Recent years, convolutional neural network (CNN) is a research hot spot in machine learning and has some application value in computer aided diagnosis. Firstly, this paper briefly introduces the basic principle of CNN. Secondly, it summarizes the improvement on network structure from two dimensions of model and structure optimization. In model structure, it summarizes eleven classical models about CNN in the past 60 years, and introduces its development process according to timeline. In structure optimization, the research progress is summarized from five aspects (input layer, convolution layer, down-sampling layer, full-connected layer and the whole network) of CNN. Thirdly, the learning algorithm is summarized from the optimization algorithm and fusion algorithm. In optimization algorithm, it combs the progress of the algorithm according to optimization purpose. In algorithm fusion, the improvement is summarized from five angles: input layer, convolution layer, down-sampling layer, full-connected layer and output layer. Finally, CNN is mapped into the medical image domain, and it is combined with computer aided diagnosis to explore its application in medical images. It is a good summary for CNN and has positive significance for the development of CNN.

Directly Connected Convolutional Neural Networks

Convolutional neural networks (CNNs) have better performance in feature extraction and classification. Most of the applications are based on a traditional structure of CNNs. However, due to the fixed structure, it may not be effective for large dataset which will spend much time for training. So, we use a new algorithm to optimize CNNs, called directly connected convolutional neural networks (DCCNNs). In DCCNNs, the down-sampling layer can directly connect the output layer with three-dimensional matrix operation, without full connection (i.e., matrix vectorization). Thus, DCCNNs have less weights and neurons than CNNs. We conduct the comparison experiments on five image databases: MNIST, COIL-20, AR, Extended Yale B, and ORL. The experiments show that the model has better recognition accuracy and faster convergence than CNNs. Furthermore, two applications (i.e., water quality evaluation and image classification) following the proposed concepts further confirm the generality and capability of DCCNNs.