Maintaining filter structure: A Gabor-based convolutional neural network for image analysis

In image segmentation and classification tasks, utilizing filters based on the target object improves performance and requires less training data. We use the Gabor filter as initialization to gain more discriminative power. Considering the mechanism of the error backpropagation procedure to learn the data, after a few updates, filters will lose their initial structure. In this paper, we modify the updating rule in Gradient Descent to maintain the properties of Gabor filters. We use the Left Ventricle (LV) segmentation task and handwritten digit classification task to evaluate our proposed method. We compare Gabor initialization with random initialization and transfer learning initialization using convolutional autoencoders and convolutional networks. We experimented with noisy data and we reduced the amount of training data to compare how different methods of initialization can deal with these matters. The results show that the pixel predictions for the segmentation task are highly correlated with the ground truth. In the classification task, in addition to Gabor and random initialization, we initialized the network using pre-trained weights obtained from a convolutional Autoencoder using two different data sets and pre-trained weights obtained from a convolutional neural network. The experiments confirm the out-performance of Gabor filters comparing to the other initialization method even when using noisy inputs and a lesser amount of training data.     

基于深度学习的单幅图像三维人脸重建研究综述

在计算机视觉领域中,三维人脸重建是一个具有研究价值的方向,高质量地重建出三维人脸在人脸识别、防伪、游戏娱乐、影视动画和美容医疗等领域具有重要的意义.近二十年来,虽然基于单幅图像的三维人脸重建领域已经取得很大的进展,但使用传统算法进行重建的结果仍会受到人脸表情、遮挡、环境光的影响,并且会出现重建效果精度不佳和鲁棒性不足等...     

卷积神经网络在目标检测中的应用综述

深度学习作为机器学习的一个分支,在各个领域的应用越来越广,已经成为语音识别、自然语言处理、信息检索等方面的一个主要发展方向;其在图像分类、目标检测等方面更是不断取得新的突破。文中首先梳理了卷积神经网络在目标检测中的典型应用;其次,对几种典型卷积神经网络的结构进行了对比,并总结了各自的优缺点;最后,讨论了深度学习现阶段存在的问题以及未来的发展方向。     

Water body segmentation of Synthetic Aperture Radar image using Deep Convolutional Neural Networks

Deep Convolutional Neural Networks are finding their way into modern machine learning tasks and proved themselves to become one of the best contenders for future development in the field. Several proposed methods in image segmentation and classification problems are giving us satisfactory results and could even perform better than humans in image recognition tasks. But also at the cost of their performance, they also require a huge amount of images for training and huge amount of computing power and time that makes them unrealistic in some situations where obtaining a large dataset is not feasible. In this work, an attempt is made for segmentation of Synthetic Aperture Radar (SAR) images which are not usually abundant enough for training, and are heavily affected by a kind of multiplicative noise called speckle noise. For the segmentation task, pre-defined filters are first applied to the images and are fed to hybrid CNN that is resulted from the concept of Inception and U-Net. The outcome of our proposed method has been examined for their effectiveness of application in a complete set of SAR images that are not used for training. The accuracy has also been compared with the manually annotated SAR images.     

Deep topology network: A framework based on feedback adjustment learning rate for image classification

Convolutional Neural Network (CNN) has demonstrated its superior ability to achieve amazing accuracy in computer vision field. However, due to the limitation of network depth and computational complexity, it is still difficult to obtain the best classification results for the specific image classification tasks. In order to improve classification performance without increasing network depth, a new Deep Topology Network (DTN) framework is proposed. The key idea of DTN is based on the iteration of multiple learning rate feedback. The framework consists of multiple sub-networks and each sub-network has its own learning rate. After the determined iteration period, these learning rates can be adjusted according to the feedback of training accuracy, in the feature learning process, the optimal learning rate is updated iteratively to optimize the loss function. In practice, the proposed DTN framework is applied to several state-of-the-art deep networks, and its performance is tested by extensive experiments and comprehensive evaluations of CIFAR-10 and MNIST benchmarks. Experimental results show that most deep networks can benefit from the DTN framework with an accuracy of 99.5% on MINIST dataset, which is 5.9% higher than that on the CIFAR-10 benchmark.     

An imitation from observation approach for dozing distance learning in autonomous bulldozer operation

Bulldozers are indispensable heavy equipment for earthwork construction, and improving the intelligence level of bulldozers is of great significance to the construction industry. An efficient autonomous construction of earthmoving machinery requires imitating and learning the expert knowledge of operators under complex environments, and imitation from observation is an effective way. In this work, the expert knowledge of operators was imitated using the proposed hybrid method for rational decision-making of dozing distance, which is one of the key factors affecting the construction efficiency of bulldozers. The proposed method is established based on the modified deep convolutional neural networks (DCNNs) and observation dataset, combined with transfer learning to apply the pre-trained deep learning model to the target task through fine tuning. Comparing the results of different methods reveals that our proposed method obtains the smallest root mean squared error (RMSE) and average error when the expert knowledge of different operators is integrated. The proposed method has universal applicability in solving the observation-based expert knowledge imitation problem. This method also breaks through the imitations of big datasets and computing resource requirements and provides an effective technical route for the practical engineering application of expert knowledge.     

RXDNFuse: A aggregated residual dense network for infrared and visible image fusion

This study proposes a novel unsupervised network for IR/VIS fusion task, termed as RXDNFuse, which is based on the aggregated residual dense network. In contrast to conventional fusion networks, RXDNFuse is designed as an end-to-end model that combines the structural advantages of ResNeXt and DenseNet. Hence, it overcomes the limitations of the manual and complicated design of activity-level measurement and fusion rules. Our method establishes the image fusion problem into the structure and intensity proportional maintenance problem of the IR/VIS images. Using comprehensive feature extraction and combination, RXDNFuse automatically estimates the information preservation degrees of corresponding source images, and extracts hierarchical features to achieve effective fusion. Moreover, we design two loss function strategies to optimize the similarity constraint and the network parameter training, thus further improving the quality of detailed information. We also generalize RXDNFuse to fuse images with different resolutions and RGB scale images. Extensive qualitative and quantitative evaluations reveal that our results can effectively preserve the abundant textural details and the highlighted thermal radiation information. In particular, our results form a comprehensive representation of scene information, which is more in line with the human visual perception system.     

A new convolutional neural network model for peripapillary atrophy area segmentation from retinal fundus images

Peripapillary atrophy (PPA) is a clinical finding, which reflects the atrophy of retina layer and retinal pigment epithelium. The size of PPA area is a useful medical indicator, as it is highly associated with many diseases such as glaucoma and myopia. Therefore, separating the PPA area from retinal images, which is called PPA area segmentation, is very important. It is a challenging task, because PPA areas are irregular and non-uniform, and their contours are blurry and change gradually. To solve these issues, we transform the PPA area segmentation task into a task of segmenting another two areas with relatively regular and uniform shapes, and then propose a novel multi-task fully convolutional network (MFCN) model to jointly extract them from retinal images. Meanwhile, we take edge continuity of the target area into consideration. To evaluate the performance of the proposed model, we conduct experiments on images with PPA areas labelled by experts and achieve an average precision of 0.8928, outperforming the state-of-the-art models. To demonstrate the application of PPA segmentation in medical research, we apply PPA related features based on the segmented PPA area on differentiating glaucomatous and physiologic large cup cases. Experiment conducted on real datasets confirms the effectiveness of using these features for glaucoma diagnosis.     

Automatic Channel Detection Using DNN on 2D Seismic Data

Geologists interpret seismic data to understand subsurface properties and subsequently to locate underground hydrocarbon resources. Channels are among the most important geological features interpreters analyze to locate petroleum reservoirs. However, manual channel picking is both time consuming and tedious. Moreover, similar to any other process dependent on human intervention, manual channel picking is error prone and inconsistent. To address these issues, automatic channel detection is both necessary and important for efficient and accurate seismic interpretation. Modern systems make use of real-time image processing techniques for different tasks. Automatic channel detection is a combination of different mathematical methods in digital image processing that can identify streaks within the images called channels that are important to the oil companies. In this paper, we propose an innovative automatic channel detection algorithm based on machine learning techniques. The new algorithm can identify channels in seismic data/images fully automatically and tremendously increases the efficiency and accuracy of the interpretation process. The algorithm uses deep neural network to train the classifier with both the channel and non-channel patches. We provide a field data example to demonstrate the performance of the new algorithm. The training phase gave a maximum accuracy of 84.6% for the classifier and it performed even better in the testing phase, giving a maximum accuracy of 90%.     

Texture evolution: 3D texture synthesis from single 2D growable texture pattern

An efficient model-independent 3D texture synthesis algorithm based on texture growing and texture turbulence is presented to create vivid 3D solid texture from a single 2D growable texture pattern. Given a 2D texture pattern of some growable material, our technique is able to create an anisotropic 3D volumetric texture cube to simulate the evolution of the material in 3D. An effective tiling scheme is designed to save computation and storage costs. Target objects are directly dipped into the synthesized 3D texture volume to generate creative, sculpture-like models that can be visualized with interactive speed. Our method is conceptually intuitive, computationally fast, and storage efficient compared with other solid texturing methods. As opposed to conventional 2D texture mapping work on polygonal surfaces, our approach is capable of decorating 3D point-rendering systems seamlessly. Furthermore, our combination of texture turbulence and texture growing techniques provides an attractive way to synthesize and tile natural 2D texture patterns, or generate simple but interesting motion textures.     

Excavator 3D pose estimation using deep learning and hybrid datasets

Earthwork operations are crucial parts of most construction projects. Heavy construction equipment and workers are often required to work in limited workspaces simultaneously. Struck-by accidents resulting from poor worker and equipment interactions account for a large proportion of accidents and fatalities on construction sites. The emerging technologies based on computer vision and artificial intelligence offer an opportunity to enhance construction safety through advanced monitoring utilizing site cameras. A crucial pre-requisite to the development of safety monitoring applications is the ability to identify accurately and localize the position of the equipment and its critical components in 3D space. This study proposes a workflow for excavator 3D pose estimation based on deep learning using RGB images. In the proposed workflow, an articulated 3D digital twin of an excavator is used to generate the necessary data for training a 3D pose estimation model. In addition, a method for generating hybrid datasets (simulation and laboratory) for adapting the 3D pose estimation model for various scenarios with different camera parameters is proposed. Evaluations prove the capability of the workflow in estimating the 3D pose of excavators. The study concludes by discussing the limitations and future research opportunities.     

Learning a generic 3D face model from 2D image databases using incremental Structure-from-Motion

Over the last decade 3D face models have been extensively used in many applications such as face recognition, facial animation and facial expression analysis. 3D Morphable Models (MMs) have become a popular tool to build and fit 3D face models to images. Critical to the success of MMs is the ability to build a generic 3D face model. Major limitations in the MMs building process are: (1) collecting 3D data usually involves the use of expensive laser scans and complex capture setups, (2) the number of available 3D databases is limited, and typically there is a lack of expression variability and (3) finding correspondences and registering the 3D model is a labor intensive and error prone process.     

3D computer vision based on machine learning with deep neural networks: A review

Recent advances in the field of computer vision can be attributed to the emergence of deep learning techniques, in particular convolutional neural networks. Neural networks, partially inspired by the brain's visual cortex, enable a computer to “learn” the most important features of the images it is shown in relation to a specific, specified task. Given sufficient data and time, (deep) convolutional neural networks offer more easily designed, more generalizable, and significantly more accurate end‐to‐end systems than is possible with previously employed computer vision techniques. This review paper seeks to provide an overview of deep learning in the field of computer vision with an emphasis on recent progress in tasks involving 3D visual data. Through a backdrop of the mammalian visual processing system, we hope to also provide inspiration for future advances in automated visual processing.     

CycleSense: Detecting near miss incidents in bicycle traffic from mobile motion sensors

In cities worldwide, cars cause health and traffic problems which could be partly mitigated through an increased modal share of bicycles. Many people, however, avoid cycling due to a lack of perceived safety. For city planners, addressing this is hard as they lack insights into where cyclists feel safe and where they do not. To gain such insights, we have in previous work proposed the crowdsourcing platform SimRa, which allows cyclists to record their rides and report near miss incidents via a smartphone app.In this paper, we present CycleSense, a combination of signal processing and Machine Learning techniques, which partially automates the detection of near miss incidents, thus making the reporting of near miss incidents easier. Using the SimRa data set, we evaluate CycleSense by comparing it to a baseline method used by SimRa and show that it significantly improves incident detection.     

Deterministic learning from neural control for uncertain nonlinear pure‐feedback systems by output feedback

The essence of intelligence lies in the acquisition/learning and utilization of knowledge. However, how to implement learning in dynamical environments for nonlinear systems is a challenging issue. This article investigates the deterministic learning (DL) control problem for uncertain pure‐feedback systems by output feedback, which achieves the human‐like learning and control in a simple way. To reduce the complexity of control design and analysis, first, by combining an appropriate system transformation, the original pure‐feedback system is transformed into a simple normal nonaffine system. An observer is then introduced to estimate the transformed system states. Based on the backstepping and dynamic surface control techniques, a simple adaptive neural control scheme is first developed to guarantee the finite time convergence of the tracking error using only one neural network (NN) approximator. Second, through DL, the exponential convergence of the NN weights is obtained with the satisfaction of partial persistent excitation condition. Thus, locally accurate approximation/learning of the transformed unknown system dynamics is achieved and stored as constant NNs. Finally, by utilizing the stored knowledge, an experience‐based controller is constructed and a novel learning control scheme is further proposed to improve the control performance without any further adaptation online for the estimate neural weights. Simulation results have been given to illustrate that the proposed scheme not only can learn and memorize knowledge like humans but also can utilize experience to achieve superior control performance.     

Online Personalised Non‐photorealistic Rendering Technique for 3D Geometry from Incremental Sketching

This paper presents an online personalised non‐photorealistic rendering (NPR) technique for 3D models generated from interactively sketched input. This technique has been integrated into a sketch‐based modelling system. It lets users interact with computers by drawing naturally, without specifying the number, order, or direction of strokes. After sketches are interpreted as 3D objects, they can be rendered with personalised drawing styles so that the reconstructed 3D model can be presented in a sketchy style similar in appearance to what have been drawn for the 3D model. This technique captures the user's drawing style without using template or prior knowledge of the sketching style. The personalised rendering style can be applied to both visible and initially invisible geometry. The rendering strokes are intelligently selected from the input sketches and mapped to edges of the 3D object. In addition, non‐geometric information such as surface textures can be added to the recognised object in different sketching modes. This will integrate sketch‐based incremental 3D modelling and NPR into conceptual design.     

CustomCut: On‐demand Extraction of Customized 3D Parts with 2D Sketches

Several applications in shape modeling and exploration require identification and extraction of a 3D shape part matching a 2D sketch. We present CustomCut, an on‐demand part extraction algorithm. Given a sketched query, CustomCut automatically retrieves partially matching shapes from a database, identifies the region optimally matching the query in each shape, and extracts this region to produce a customized part that can be used in various modeling applications. In contrast to earlier work on sketch‐based retrieval of predefined parts, our approach can extract arbitrary parts from input shapes and does not rely on a prior segmentation into semantic components. The method is based on a novel data structure for fast retrieval of partial matches: the randomized compound k‐NN graph built on multi‐view shape projections. We also employ a coarse‐to‐fine strategy to progressively refine part boundaries down to the level of individual faces. Experimental results indicate that our approach provides an intuitive and easy means to extract customized parts from a shape database, and significantly expands the design space for the user. We demonstrate several applications of our method to shape design and exploration.     

Comparative analysis on the application of neuro‐fuzzy models for complex engineered systems: Case study from a landfill and a boiler

This work aims at developing an explicit neuro‐fuzzy (NF) model to characterize complex engineered systems associated with high nonlinearity, uncertainties, and multivariable couplings. The NF model synergistically exploits the advantages of fuzzy belongingness of each input variable to all output variables and learning ability of neural networks. Owing to the inherent complexities associated with 2 complex engineered systems, a landfill and a boiler were selected to develop models that provide intelligent decisions for optimizing the operational parameters. Data compiled from field‐scale investigation/real plant operation involving various operating scenarios were used to develop the models. Predicting capability of the developed models was evaluated through the correlation coefficient and mean absolute percentage error values. Superiority of the proposed NF model to other similar models has been justified and demonstrated.