In the context of the intelligent era, the processing power of mobile terminals is getting much stronger, and the functions that can be realized are becoming more and more abundant. This project is intended to design a complete application system for power companies, including mobile application, web management system and cloud server, to improve the intelligent level of on-site operations. This topic involves image processing, machine learning, android and python mixed programming, java web development and other technologies, the research results are of great application value.
In recent years, deep learning has made great breakthroughs in image recognition and image classification. We use traditional machine learning methods and deep learning methods to detect, extract, and identify objects in the substation drawing elements. Analyze the relationship between each object element in the analysis, and export the SPCD and other documents that meet the specifications of the intelligent substation fiber loop modeling and coding technology.
We aim to use artificial intelligence technology to achieve accurate speech quality assessment of FM signals in different scenarios, obtain important factors affecting speech quality through research, then collect samples of FM signals in different environments, processing generates raw data for training artificial intelligence algorithms. we try many different algorithms, including support vector machine with low complexity, ensemble learning, or deep learning algorithms with high complexity, such as cyclic neural network and convolution neural network, to select the optimal algorithm, and finally achieve the accurate classification of speech quality.
Because the spectrum of UHF full-band signals from 30Hz to 3000MHz is too dense, it is very difficult to identify a specific modulation signal. The purpose of this study is to detect a specific modulation signal from UHF full-band signals by machine learning and deep learning, and to judge whether it exists or not. Find out where the signal is located.
With the widespread application of compressed sensing algorithms in image processing and signal processing in recent years, we have applied this algorithm in combination with lossless compression algorithms in the processing of astronomical data, improving the transmission of satellites to the ground. Data traffic, effectively reducing data transmission costs
Deep learning activity in multiple directions of research in machine learning and even active in all research directions, is currently mainly used in the field of artificial intelligence. In recent years, deep learning in image recognition and image classification has a greater breakthrough. We try to use deep learning in the signal modulation recognition. The difficulty of signal modulation recognition is that the feature extraction of different modulation signals is difficult. We can use the convolution neural network to extract the feature automatically to extract the different characteristics of the time-frequency spectrum of the modulation signal to identify the modulation scheme. Model before we collected a large number of different modulations modulating signal IQ data, the use of collected IQ data into time-frequency map, the establishment of a database of more than ten million pictures, the use of convolution neural network to train these images, the final has achieved good results.
The technology of 3D garment simulation has been widely used in various areas, such as animation film, online game, virtual fitting and virtual fashion show. In the garment simulation, the sense of reality and requirement for real time are two major factors we consider.Firstly, we construct the 2D garment and generate a 3D garment by means of virtual sewing on the basis of the parameters of the human body. Then, we research on the reusability of garment. We propose a deformation method based on layered adaptive deformation, which is an efficient approach for improving reusability of each set of garment. Finally, we propose a classified strain limiting method which deals with unreal stretch deformation in clothing simulation with physical-based mass-spring model. Results indicate that our method has good performance and efficiency.
In order to balance the motion controllability and the realistic appearance of virtual human, we drive a virtual human animation with a surface model. The surface model consists of an inner virtual human motion skeleton and an external virtual human mesh skin. After the skeleton and skin was bound, the skeleton can drive the surface of the virtual human skin deformation, resulting in a realistic virtual human skinned mesh animation. Virtual human skeleton extraction is based on the principle of Reeb graph topology segmentation, the method can extract the static skeleton points exactly. After the adjustment of the skeleton points and the redefinition of the skeleton, the forward kinematics principle is used to drive the virtual human skeleton. Because of the topology structure of the model was considered in the process of skeleton extraction, the method can drive the virtual human skinned mesh animation with different initial pose successfully, and the animation is vivid and controllable.
头发快速仿真与渲染是计算机动画和虚拟人领域至关重要的两个部分,头发的大批量处理会带来较大的计算开销,同时头发受风力驱动的动态仿真也是一个巨大的挑战。本文在平衡速度与视觉效果的前提下,采用Dynamic follow-the-leader (DFTL) 方法对头发进行快速的仿真,并提出一种新的基于高斯核函数的快速插值算法和一个简单的卷发建模方法。我们还提出一种基于数学和物理的风力模型,此模型可应用于大部分质点弹簧头发模型。最后使用两种光照模型-Kajiya-kay 和Marschner对头发进行实时地渲染和着色,并使用deep opacity maps算法产生头发自阴影。最终实验结果证明了本文提出的风力模型在头发仿真中可得到较好的动态效果,也具有较好的实用性。
The fast simulation and rendering of hair are two significant parts in the field of computer animation and virtual characters. The processing of enormous strands of hair always results in an expensive amount of calculation. Dynamic hair simulation driven by wind forces also brings us a big challenge. In the premise of thinking about the balance between visual effect and speed, we adopt the Dynamic follow-the-leader (DFTL) method to perform hair simulation, present a new Gaussian kernel-based fast interpolation algorithm and a simple frizzy hair modeling method in this paper. We also propose a wind model based on mathematics and physics, and it can be used in various mass-spring hair models. Lastly, we render and shade the hair by two shading models- Kajiya-kay and Marschner in real time, deep opacity maps algorithm is also used to generate hair self-shadowing. The final experiments prove that our wind model performs well in the dynamic hair simulation, and has good practical results as well.
We do a research on leaf carving image which is a traditional art in China. To get the final 3D leaf carving model, the leaf carving image is divided into different parts. Through the distance image combining with 3D reconstruction algorithm, the 3D model is generated, then the height adjustment function is carried on the 3D model to get the suitable 3D model. Later stage, Through the hollow out operation、narrow adjustment and patch repair, we can get the final model.
We implement a deep neural networks (DNN) approach for speech-driven articulators motion synthesis, which can be applied to speech-driven talking avatar. We realize acoustic-articulatory mapping by DNN. The input of the system is acoustic speech and the out is the estimated articulatory movements. Experiments on the croups MNGU0 and two conclusions are obtained. First, the performance of the method based on DNN for speech-driven articulators motion synthesis is superior than ANN. Second, long context acoustic input can effectively reduce the reconstruction error.
Lane markings detection plays an important role in the ITS (intelligent transport system), which is widely used in driver assistance system, lane departure warming system and forward collision warning system, it is significant for improving the driving safety. Edge information is used as the feature information for the extraction of lane markings, then, the line segments detected by Progressive Probabilistic Hough Transform (PPHT) are applied to represent the structural information of lanes and the interferential line segments are eliminated under vanishing point constrains, K-means clustering algorithm is used to classify for fitting the closest left and right lane markings and determining the travel area of the intelligent vehicle, The experiment results demonstrate that this method can detect lane markings in various complex environments.
The main task of this topic is to realize an efficient and real-time traffic sign detection and recognition algorithm. In the detection stage, the first use of the method of color segmentation, Preliminary processing of the acquired images, to get detected traffic signs. According to the special shape of traffic signs, such as circle, rectangle, triangle, etc, to carry out accurate positioning of traffic signs. After the detection of traffic signs, we need to classify the traffic signs, to extract the characteristics of the traffic signs to match with the standard traffic signs, or to use the method of training classifiers to identify the traffic signs.