Fine-grained vehicle classification is a challenging topic in computer vision due to the high intraclass variance and low interclass variance. Recently, considerable progress has been made fine-grained huge success of deep neural networks. Most studies based on networks, focus network structure improve performance. In contrast existing works classification, we loss function network. We add regularization term cross-entropy propose new function, Dual Cross-Entropy Loss. The places constraint...

With the rapid development of deep learning, convolutional neural networks have achieved milestones in synthetic aperture radar (SAR) image object detection. However, detection SAR images is still a great challenge due to difficulty distinguishing targets from complex backgrounds. At same time, most are small and unevenly distributed, which makes it challenging extract sufficient feature information. To solve these issues mentioned above, an efficient network for based on Swin transformer...

Oriented object detection is a challenging task in scene text and remote sensing image analysis, which has attracted extensive attention recent years with the development of deep learning. Currently, mainstream oriented detectors are based on preset anchor boxes. This method increases computational load network cause large amount box redundancy. To solve this problem, we propose anchor-free Gaussian centerness(AOGC), single-stage method. Our uses contextual FPN(CAFPN) to obtain information...

This paper presents a new project about UAV landing. Firstly, the applies zigbee to communicate and estimate position of UAV, which is wireless network technology in world. Secondly, vision guidance used realize identifying confirming aircraft accurate position. To identify multi-targets CCD vision, an improved morphological gradient image edge detection operator equipped with orientation was provided for better quality gargets' contour moment algorithm. It useful that blurring processing...

Traditional methods of access control such as password token or identification card are being replaced by biometrics recognition technology in many fields because their limitation reliability and usability. Vein pattern is outstanding compared to other fingerprint iris due its dependability ease use. At the same time, current systems applications have some disadvantages usability, cost, supported user age range. In this paper, we propose MyPalmVein, which a low-cost system based on vein...

Convolutional Neural Networks (CNNs) have been successfully used in various image classification tasks and gradually become one of the most powerful machine learning approaches. To improve capability model generalization performance on small-sample classification, a new trend is to learn discriminative features via CNNs. The idea this paper decrease confusion between categories extract enlarge inter-class variance, especially for classes which indistinguishable features. In paper, we propose...

This paper designs a novel algorithm for automatic guided vehicle. The uses the CCD in front of vehicle to identify routing indicator and calculate yaw angle from track. It applies odometer get displacement. DR can provide direction position Machine vision guidance track indicator. adaptive fuzzy feedback faster better, further simple arithmetic reduce picture jitter. system used this avoid error accumulated with time as general cost lower. Plenty examinations prove that efficiently improve precision

为了提高系统特征提取算法的计算效率、减少占用的存储空间和简化程序设计，该文基于Riemann 流形上优化算法的几何框架，提出了改进的Stiefel 流形上的梯度下降算法。根据不同要求采用不同的测地线计算公式，并使用多项式逼近测地线方程，同时采用了秦九韶-Horner 多项式算法及线搜索、变步长的方法。以主分量分析问题为例，详细讨论了Stiefel 流形上的梯度算法在其中的应用。理论分析和实验结果均表明，此方法可以在确保迭代矩阵列向量单位正交性的同时获得更好的计算效率和收敛速度，并且更容易实现。

In order to solve the problem of deep learning models that are difficult balance accuracy and speed, this paper selects YOLOv3 model with higher detection speed but lower improve it, intends construct a high-speed high-precision target model. First, we used ResNeXt residual block backbone network few parameters high accuracy. Then K-means++ clustering algorithm obtain anchor boxes is closer real box help perform regression detection. The experimental results on VOC data set show improved...

Abstract The method of object detection has been applied to all aspects in our lives. Although methods based on deep learning have widely used various fields, there are still some overlooked problems the candidate box selection stage. results traditional can only select a relatively optimal maximum box. If is not accurate enough, this type will be able do adjust it. To solve problem, an multiple fusion proposed. retain and delete non‐maximum box, but also position again. Thereby more...

This article analyzed the data of survey 2004-2009 graduates' professional development status Shanghai University Traditional Chinese Medicine working in community health centers Pudong New District , which had been conducted a deep discussion with career as guideline and existing merits problems breakthrough. Besides, it also discussed feasibility suggestions TCM employment promotion primary care units combined current cultivating mechanism. It concluded that adjustment colleges' students'...

Traffic signs contain important traffic information. The traditional sign detection method cannot solve the problem of low accuracy caused by small, occupied area images. Based on this, a algorithm based improved YOLOv4 is proposed. Firstly, 13 × large receptive field layer removed structure, and 104 added. It obtains more global feature information improves accuracy. attention mechanism introduced into algorithm, that is, backbone network extracts three layers then adds scSE module. Make...

In this paper, we propose the nonlinearity generation method to speed up and stabilize training of deep convolutional neural networks. The proposed modifies a family activation functions as generators (NGs). NGs make linear symmetric for their inputs lower model capacity, automatically introduce enhance capacity during training. can be considered an unusual form regularization: parameters are obtained by relatively low-capacity model, that is easy optimize at beginning, with only few...

In order to estimate the potential hazards and adopt strategies for preventing accidents, this paper is about research of auto-recognition pedestrians. Using captured images vehicle recorder, based on HOG feature processing technology SVM classifier, automatic detection recognition pedestrians are realized. past, cost pedestrian identification system too high its universality poor. research, optimization methods reasonable sample allocation difficult training used make model more general effective.

Deep learning has made substantial progress in crowd density estimation, but there are still some problems existing methods, such as large population density, background interference, and scale change, which makes it difficult to count people. To solve the above problems, we proposed a counting method based on cross column fusion attention mechanism. First, first ten layers of VGG16 with good migration ability feature extraction used front-end network preliminarily extract human head...