基于修正CycleGAN的声呐图像库构建方法研究

声呐图像数据集获取困难,导致很多水下工作无法正常开展,如水下目标检测与跟踪、声呐图像的超分辨等,因此构建充足的声呐图像数据库成为很多水下研究工作的重要前提条件。受光学图像与合成孔径雷达（Synthetic Aperture Radar,SAR）图像转换研究工作的启发,提出了基于CycleGAN实现声呐图像库的构建,即利用光学图像合成声呐图像,实现光学到声呐的图像风格迁移。通过对CycleGAN网络损失函数的改进,提高了声呐图像的合成效果。通过与Pix2Pix等图像风格迁移网络进行比较的实验结果证明,修正后的CycleGAN网络具有更好的图像风格迁移效果。最后用合成的声呐图像训练Mask RCNN目标检测网络,并用真实的声呐图像进行测试,训练得到的模型能够成功地检测出真实声呐图像中对应的目标,进一步验证了利用光学图像构建声呐图像库的有效性。     

Optical implementation of neural networks for face recognition by the use of nonlinear joint transform correlators

We describe a nonlinear joint transform correlator-based two-layer neural network that uses a supervised learning algorithm for real-time face recognition. The system is trained with a sequence of facial images and is able to classify an input face image in real time. Computer simulations and optical experimental results are presented. The processor can be manufactured into a compact low-cost optoelectronic system. The use of the nonlinear joint transform correlator provides good noise robustness and good image discrimination.     

Optical dual-scale architecture for neural image recognition

A novel neural-network architecture that combines image data reduction with focus of attention to achieve reduced training cost, improved noise tolerance, and better generalization performance than comparable conventional networks for image-recognition tasks is presented. The dual-scale architecture is amenable to optical implementation, and an example optical system is demonstrated. For one example problem, the best-case improvements of the dual-scale network over its conventional counterpart were found through simulation to be a factor of 6.7 in training cost, 67.3% in noise tolerance, and 61.6% in generalization to distortions. The dual-scale network is also applied to one instance of a human face recognition problem.     

NEURAL NETWORK PATTERN RECOGNITION OF BLAST FRAGMENT SIZE DISTRIBUTIONS

Neural networks have considerable potential for applications in particulate image analysis. An area of great current interest is to use image analysis techniques to characterize particle size distributions in video images of blasted rock. A simulated neural network was trained to recognize fragmented rock size classes taken from images of blasted ore in a large open pit mining operation. Size distributions were assigned to categories such as 40% and 60% minus six inches. Pattern recognition features were extracted from digitized images using two-dimensional Fourier transforms. These features were then used as a training set to enable the neural network to recognize the size category of subsequent images of blasted rock taken from the mining operation. Training sets were developed for a back propagation algorithm by hand sorting and sizing the blast fragments from photographed piles. Within the limits of this experiment, the trained network consistently recognized the size distribution categories. A trained neural network can be readily calibrated to adjust for changes in light and shadow, a problem which plagues algorithm-based blast fragmentation analysis routines. Neural network techniques may provide a solution to the problem of rapid and reliable on-line and on-site size distribution recognition and assessment.     

Convolutional Neural Network Based Intelligent Handwritten Document Recognition

This paper presents a handwritten document recognition system based on the convolutional neural network technique. In today’s world, handwritten document recognition is rapidly attaining the attention of researchers due to its promising behavior as assisting technology for visually impaired users. This technology is also helpful for the automatic data entry system. In the proposed system prepared a dataset of English language handwritten character images. The proposed system has been trained for the large set of sample data and tested on the sample images of user-defined handwritten documents. In this research, multiple experiments get very worthy recognition results. The proposed system will first perform image pre-processing stages to prepare data for training using a convolutional neural network. After this processing, the input document is segmented using line, word and character segmentation. The proposed system get the accuracy during the character segmentation up to 86%. Then these segmented characters are sent to a convolutional neural network for their recognition. The recognition and segmentation technique proposed in this paper is providing the most acceptable accurate results on a given dataset. The proposed work approaches to the accuracy of the result during convolutional neural network training up to 93%, and for validation that accuracy slightly decreases with 90.42%.     

基于卷积神经网络的模糊车牌自动识别

目前,清晰的车牌识别算法已经成熟,但是对于人眼不能识别的模糊车牌,传统车牌识别算法的识别率较低或者根本无法识别。鉴于此,提出了一种基于卷积神经网络的车牌字符识别算法。制作了含9 720幅模糊字符样本集,用8 748幅样本对卷积神经网络进行训练,测试样本时,先对模糊车牌字符进行盲分割等预处理,再调用训练好的卷积神经网络对盲分割后的字符进行识别。实验结果表明:该算法对训练集的准确识别率约为99.17%,对测试集的准确识别率约为93.32%,这说明该算法对模糊车牌的识别具有鲁棒性,能应用于各种场景。     

Face recognition by using feature-specific imaging

We present a face-recognition system based on the optical measurement of linear features. We describe a polarization-based optical system that computes linear projections of an incident irradiance distribution. We quantify the fundamental limitations of optical feature measurement. We find that higher feature fidelity can be obtained by feature-specific imaging than by postprocessing a conventional image. We present feature-fidelity results for wavelet, principal component, and Fisher features. We study face recognition by using a k-nearest neighbors classifier and two different feed-forward neural networks. Each image block is reduced to either a one- or a two-dimensional feature space for input to these recognition algorithms. As high as 99% recognition has been achieved with one-dimensional wavelet feature projections and 100% has been achieved with two-dimensional projections. A 95-fold increase in noise tolerance by use of feature-specific imaging has been demonstrated for an example of the face-recognition problem. An optical experiment is performed to validate these results.     

Improved classification robustness for noisy cell images represented as principal-component projections in a hybrid recognition system

Different types of cells are recognized from their noisy images by use of a hybrid recognition system that consists of a learning principal-component analyzer and an image-classifier network. The inputs to the feed-forward backpropagation classifier are the first 15 principal components of the 10 x 10 pixel image to be classified. The classifier was trained with clear images of cells in metaphase, unburst cells, and other erroneous patterns. Experimental results show that the recognition system is robust to image scaling and rotation, as well as to image noise. Cell recognition is demonstrated for images that are corrupted with additive Gaussian noise, impulse noise, and quantization errors. We compare the performance of the hybrid recognition system with that of a conventional three-layer feed-forward backpropagation network that uses the raw image directly as input.     

Optical implementation of a translation-invariant second-order neural network for multiple-pattern classification

A novel approach to the optical implementation of second-order neural networks that can recognize multiple patterns is reported. The systems issues, especially the accuracy required for the weighted interconnections, are discussed for numeric character (0-9) recognition. It is shown that the accuracy of the weighted interconnections has a far greater influence on the network performance during training than on classification. To lessen the problem, we introduce an adaptive learning rule, whereby the optical power is adjusted during training. Finally, numeric character recognition using an experimental system with a liquid-crystal display is demonstrated.     

基于卷积神经网络的印刷品颜色缺陷检测

目的 提取样本图像颜色直方图特征对卷积神经网络进行训练,达到快速、高准确率检测图像颜色缺陷的目的.方法 将标准图像从RGB颜色空间转换至HSV颜色空间,通过改变图像H,S,V三分量值获取训练样本和测试样本;在HSV颜色空间中非均匀量化图像的颜色直方图,得到所有训练样本和测试样本的颜色直方图特征;利用样本图像颜色直方图特征训练卷积神经网络,然后对测试样本进行检测,研究检测的速度、准确率,并将该检测方法与逐像素、超像素、BP神经网络和支持向量机方法进行对比.结果 对于图片尺寸为512×512的彩色图像,卷积神经网络检测单幅图片的平均检测时间约为57.66 ms,训练样本图像为50000张时,卷积神经网络方法对10000张测试样本进行检测的准确率为99.77％.结论 卷积神经网络方法在保证高准确率的前提下大幅提高检测精度,对于印刷品色差缺陷在线检测具有良好的应用价值.     

Spatially resolved absolute diffuse reflectance measurements for noninvasive determination of the optical scattering and absorption coefficients of biological tissue

The absorption and transport scattering coefficients of biological tissues determine the radial dependence of the diffuse reflectance that is due to a point source. A system is described for making remote measurements of spatially resolved absolute diffuse reflectance and hence noninvasive, noncontact estimates of the tissue optical properties. The system incorporated a laser source and a CCD camera. Deflection of the incident beam into the camera allowed characterization of the source for absolute reflectance measurements. It is shown that an often used solution of the diffusion equation cannot be applied for these measurements. Instead, a neural network, trained on the results of Monte Carlo simulations, was used to estimate the absorption and scattering coefficients from the reflectance data. Tests on tissue-simulating phantoms with transport scattering coefficients between 0.5 and 2.0 mm(-1) and absorption coefficients between 0.002 and 0.1 mm(-1) showed the rms errors of this technique to be 2.6% for the transport scattering coefficient and 14% for the absorption coefficients. The optical properties of bovine muscle, adipose, and liver tissue, as well as chicken muscle (breast), were also measured ex vivo at 633 and 751 nm. For muscle tissue it was found that the Monte Carlo simulation did not agree with experimental measurements of reflectance at distances less than 2 mm from the incident beam.     

Neural network implementation using self-lensing media

An all-optical implementation of a feed-forward artificial neural network is presented that uses self-lensing materials in which the index of refraction is irradiance dependent. Many of these types of material have ultrafast response times and permit both weighted connections and nonlinear neuron processing to be implemented with only thin material layers separated by free space. Both neuron processing and weighted interconnections emerge directly from the physical optics of the device. One creates virtual neurons and their connections simply by applying patterns of irradiance to thin layers of the nonlinear media. This is a result of a variation of the refractive-index profile of the self-lensing nonlinear media in response to the applied irradiance. An optical-backpropagation training method for this network is presented. The optical backpropagation is a training method that can be implemented potentially within the same optical device as the forward calculations, although several issues crucial to this po sibility remain to be addressed. Such a network was numerically simulated and trained to solve many benchmark classification problems, and some of these results are presented. To demonstrate the feasibility of building such a network, we also describe experimental work in the construction of an optical network trained to perform a logic XNOR function. This network, as a proof of concept, uses a relatively slow thermal nonlinear material with ~1-s response time.     

神经网络在二维图像识别中的应用

本文提出了一种基于神经网络的二维图像识别技术。选取一组机械零件的二维图像,对每张图像进行放缩和旋转变换,并分析、提取对应图像的nmi特征和7个不变矩特征作为BP网络的输入样本,图像的二进制编号为输出样本构建BP神经网络。并对网络进行抗干扰训练,使网络对理想输入及带噪声的输入均有较好的识别率。实验证明该网络具有一定的工程实用性。     

Gaussian synapse ANNs in multi- and hyperspectral image data analysis

A new type of artificial neural network is used to identify different crops and ground elements from hyperspectral remote sensing data sets. These networks incorporate Gaussian synapses and are trained using a specific algorithm called Gaussian synapse back propagation described here. Gaussian synapses present an intrinsic filtering ability that permit concentrating on what is relevant in the spectra and automatically discard what is not. The networks are structurally adapted to the problem complexity as superfluous synapses and/or nodes are implicitly eliminated by the training procedure, thus pruning the network to the required size straight from the training set. The fundamental difference between the present proposal and other ANN topologies using Gaussian functions is that the latter use these functions as activation functions in the nodes, while in our case, they are used as synaptic elements, allowing them to be easily shaped during the training process to produce any type of n-dimensional discriminator. This paper proposes a multi- and hyperspectral image segmenter that results from the parallel and concurrent application of several of these networks providing a probability vector that is processed by a decision module. Depending on the criteria used for the decision module, different perspectives of the same image may be obtained. The resulting structure offers the possibility of resolving mixtures, that is, carrying out a spectral unmixing process in a very straightforward manner.     

An Improved Deep Fusion CNN for Image Recognition

With the development of Deep Convolutional Neural Networks (DCNNs), the extracted features for image recognition tasks have shifted from low-level features to the high-level semantic features of DCNNs. Previous studies have shown that the deeper the network is, the more abstract the features are. However, the recognition ability of deep features would be limited by insufficient training samples. To address this problem, this paper derives an improved Deep Fusion Convolutional Neural Network (DF-Net) which can make full use of the differences and complementarities during network learning and enhance feature expression under the condition of limited datasets. Specifically, DF-Net organizes two identical subnets to extract features from the input image in parallel, and then a well-designed fusion module is introduced to the deep layer of DF-Net to fuse the subnet’s features in multi-scale. Thus, the more complex mappings are created and the more abundant and accurate fusion features can be extracted to improve recognition accuracy. Furthermore, a corresponding training strategy is also proposed to speed up the convergence and reduce the computation overhead of network training. Finally, DF-Nets based on the well-known ResNet, DenseNet and MobileNetV2 are evaluated on CIFAR100, Stanford Dogs, and UECFOOD-100. Theoretical analysis and experimental results strongly demonstrate that DF-Net enhances the performance of DCNNs and increases the accuracy of image recognition.     

基于卷积神经网络的印刷套准识别方法

目的 针对目前印刷套准识别方法依赖于经验人工设计特征提取的问题,提出一种不需要人工提取图像特征的卷积神经网络模型,实现印刷套准状态的识别.方法 采用图像增强技术实现不均衡训练集的均衡化,增加训练集图像的数量,提高模型的识别准确率.设计基于AlexNet网络结构的印刷套准识别模型的结构参数,分析批处理样本数量和基础学习率对模型性能的影响规律.结果 文中方法获得的总印刷套准识别准确率为0.9860,召回率为1.0000,分类准确率几何平均数为0.9869.结论 文中方法能自动提取图像特征,不依赖于人工设计的特征提取方法.在构造的数据集上,文中方法的分类性能优于实验中的支持向量机方法.     

Spectral peak verification and recognition using a multilayered neural network

The verification and recognition of peak-shaped signals in analytical data are ubiquitous scientific problems. Experimental data contain overlapping signals and noise, which make sensitive and reliable peak recognition difficult. A peak detection system based on a class of neural networks known as "multilayered perceptrons" has been created. The network was trained and evaluated with use of vapor-phase infrared spectral data. The results of varying the network architecture on system training and prediction performance along with refinement of the form of the input pattern are presented.     

融合人脸表情的手语到汉藏双语情感语音转换

针对聋哑人与正常人之间存在的交流障碍问题,提出了一种融合人脸表情的手语到汉藏双语情感语音转换的方法。首先使用深度置信网络模型得到手势图像的特征信息,并通过深度神经网络模型得到人脸信息的表情特征。其次采用支持向量机对手势特征和人脸表情特征分别进行相应模型的训练及分类,根据识别出的手势信息和人脸表情信息分别获得手势文本及相应的情感标签。同时,利用普通话情感训练语料,采用说话人自适应训练方法,实现了一个基于隐Markov模型的情感语音合成系统。最后,利用识别获得的手势文本和情感标签,将手势及人脸表情转换为普通话或藏语的情感语音。客观评测表明,静态手势的识别率为92.8%,在扩充的Cohn-Kanade数据库和日本女性面部表情(Japanese Female Facial Expression,JAFFE)数据库上的人脸表情识别率为94.6%及80.3%。主观评测表明,转换获得的情感语音平均情感主观评定得分4.0分,利用三维情绪模型(Pleasure-Arousal-Dominance,PAD)分别评测人脸表情和合成的情感语音的PAD值,两者具有很高的相似度,表明合成的情感语音能够表达人脸表情的情感。     

Handling dataset dependence with model ensembles for skin lesion classification from dermoscopic and clinical images

Dataset dependence affects many real-life applications of machine learning: the performance of a model trained on a dataset is significantly worse on samples from another dataset than on new, unseen samples from the original one. This issue is particularly acute for small and somewhat specific databases in medical applications; the automated recognition of melanoma from skin lesion images is a prime example. We document dataset dependence in dermoscopic skin lesion image classification using three publicly available medium size datasets. Standard machine learning techniques aimed at improving the predictive power of a model might enhance performance slightly, but the gain is small, the dataset dependence is not reduced, and the best combination depends on model details. We demonstrate that simple differences in image statistics account for only 5% of the dataset dependence. We suggest a solution with two essential ingredients: using an ensemble of heterogeneous models, and training on a heterogeneous dataset. Our ensemble consists of 29 convolutional networks, some of which are trained on features considered important by dermatologists; the networks' output is fused by a trained committee machine. The combined International Skin Imaging Collaboration dataset is suitable for training, as it is multi-source, produced by a collaboration of a number of clinics over the world. Building on the strengths of the ensemble, it is applied to a related problem as well: recognizing melanoma based on clinical (non-dermoscopic) images. This is a harder problem as both the image quality is lower than those of the dermoscopic ones and the available public datasets are smaller and scarcer. We explored various training strategies and showed that 79% balanced accuracy can be achieved for binary classification averaged over three clinical datasets.     

基于YOLOv5s网络的垃圾分类和检测

目的 为了实现垃圾自动按类处理,通过研究基于视觉的垃圾检测与分类模型,实现对垃圾的自动识别和检测.方法 采用YOLOv5s网络作为垃圾检测与分类的模型,在自制垃圾分类数据集上对网络进行训练,利用训练好的YOLOv5s网络提取不同种类垃圾图像的特征和位置信息,实现垃圾的分类与检测.结果 在真实场景中进行了测试,基于YOLOv5s的垃圾分类检测模型可以有效识别6种不同形态的垃圾,检测mAP值为99.38％,测试精度为95.34％,目标检测速度达到6.67FPS.结论 实验结果表明,基于YOLOv5s网络的垃圾分类检测模型在不同光照、视角等条件下,检测准确率高,鲁棒性好、计算速度快.同时,有助于促进垃圾处理公司实现智能分拣,提高工作效率.