CN110929811B - Deep learning method for acquiring full waveform laser radar high-resolution data - Google Patents
Deep learning method for acquiring full waveform laser radar high-resolution data Download PDFInfo
- Publication number
- CN110929811B CN110929811B CN201911001708.8A CN201911001708A CN110929811B CN 110929811 B CN110929811 B CN 110929811B CN 201911001708 A CN201911001708 A CN 201911001708A CN 110929811 B CN110929811 B CN 110929811B
- Authority
- CN
- China
- Prior art keywords
- data
- waveform
- laser radar
- full
- deep learning
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
- G06V30/19—Recognition using electronic means
- G06V30/192—Recognition using electronic means using simultaneous comparisons or correlations of the image signals with a plurality of references
- G06V30/194—References adjustable by an adaptive method, e.g. learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Abstract
The invention relates to a deep learning method for acquiring full-waveform laser radar high-resolution data. Belongs to the field of deep learning and radar data processing. The method comprises the steps of constructing a deep learning network framework model by constructing a data processing platform, training the deep learning network framework model, adjusting parameters in the deep learning network framework model according to Loss functions Loss and PSNR, and storing the deep learning network framework model with the best training effect, wherein the model is used for processing full-waveform laser radar data, so that the space-time resolution of the full-waveform laser radar data is improved multiple times, the problem that the resolution is limited under the constraint of the limitation of full-waveform laser radar hardware scanning equipment is solved, and when new full-waveform laser radar data is processed, only the stored deep learning network framework model is required to be called, so that the method is suitable for occasions with high requirements on the volume and portability of acquisition equipment.
Description
Technical Field
The invention relates to a deep learning method for acquiring full-waveform laser radar high-resolution data. In particular to a method for acquiring higher resolution data by adopting a residual neural network to process and train full waveform laser radar data in batches based on a deep learning strategy. Belongs to the field of deep learning and radar data processing.
Background
In recent years, the academia and industry have paid a great deal of attention to the field of deep learning. Deep learning is a method of machine learning, and in the past decades, the deep learning absorbs and references knowledge about human brain nerves, statistics and applied mathematics, and uses the rapidly improved computer computing power, the rapidly increased training data set and the more complete and exquisite deep neural network training technology in recent years, so that the deep learning machine learning can be widely developed in various application fields.
Deep learning has been successfully applied to an increasingly wide range of practical problems since the 80 s of the 20 th century. The earliest deep learning model is used for identifying single objects in the image which is cut to be extremely small in size, the later developed deep convolutional neural network can identify high-resolution pictures with rich processing details, cutting is not needed, and at least 1000 different types of objects are identified from early identification of second objects to annual imaging net large-scale visual identification challenge. Furthermore, deep learning has been remarkably successful in pedestrian detection and image segmentation. In 2012, ciresan et al, in the paper, mention that deep learning was performed on traffic sign classification, has achieved performance beyond humans.
With the increasing scale and precision of deep learning networks, the task that they can perform is increasingly complex, contributing to the development of other scientific research. The brain neuroscientist can rely on the deep convolutional neural network of object recognition to build a visual processing model for research; statistics and finance researchers can process mass scientific data and make effective predictions in the professional field by using the tool of deep learning; in addition, deep learning can predict interactions between molecules in the field of biopharmaceuticals, helping pharmaceutical companies to design new drugs; deep learning can also be used for automatically analyzing and constructing microscopic images of three-dimensional images of human brain. In recent years, deep learning has been greatly developed in popularity and practicality in various research fields. In the future, with the development of the artificial intelligence revolution in the new era, the powerful tool of deep learning is taken as a powerful tool, and the novel opportunity and challenge are filled in the deeper and wider field.
The airborne laser radar is an active detection device integrating a laser scanning system, a global positioning system and an inertial navigation system and used for rapidly acquiring three-dimensional information of the earth surface. The earliest airborne laser radar can only record a limited number of discrete echo signals in a period of time, and calculates the spatial distance between the radar and the detected ground object according to the time interval between the received echo signals and the emitted laser pulses. Whereas a full waveform lidar system at the end of the twentieth century, in comparison to a discrete lidar, can provide more abundant target information because the full waveform lidar records the backscattered energy of each height Cheng Dian of the detected target in the form of a waveform, in comparison to conventional point cloud data-only lidar. Each echo waveform can theoretically be regarded as the result of the superposition of several gaussian functions. The width, peak position, peak-to-peak spacing, and amplitude of the waveform are all important parameters for analysis of the detection target. In recent years, the abundant information content of the full waveform laser radar complete back scattering echo waveform arouses the interest of researchers in various fields, and the new era of full waveform laser radar data processing and application research is brought forward.
Because the space-time resolution of the full-waveform laser radar is limited by factors such as the pulse width, the intensity, the detection bandwidth of the detector and the like of the emitted laser, the improvement of the data resolution of the full-waveform laser radar can be considered to update hardware equipment, a laser light source with higher intensity and narrower pulse width is adopted, a detector with wider bandwidth, a digital-to-analog converter and the like are used, however, the hardware equipment needs to be replaced firstly at a huge cost, and secondly, the technical indexes of corresponding hardware also have bottlenecks, so that the improvement of the data resolution of the full-waveform laser radar is restricted. In terms of software methods, methods for full-waveform lidar data, such as a waveform decomposition method, a deconvolution method, etc., are mostly dedicated to extracting a detection target in an original waveform, but the problem of improving the resolution of the original data is not considered. In 2017, researchers propose to apply the compressed sensing technology to the full-waveform laser radar data acquisition process so as to recover and reconstruct with a small amount of measurement observation signals to obtain high-resolution data, and also, from the aspect of hardware system design, researchers propose a method for acquiring time-domain super-resolution full-waveform laser radar data by arranging a plurality of detectors in a staggered and linear array manner.
However, at the software level, the method of acquiring the higher resolution data for the full-waveform laser radar data is to be improved accordingly, and innovative methods are to be proposed.
Disclosure of Invention
The invention aims to provide a deep learning method for acquiring full-waveform laser radar high-resolution data, which combines a neural network and a large amount of full-waveform laser radar data based on a deep learning strategy, so that the resolution of original full-waveform laser radar data is improved multiple times, and the problem that the resolution of the full-waveform laser radar data is limited when the full-waveform laser radar hardware scanning equipment is limited is solved.
The invention aims at realizing the following technical scheme:
firstly, a data processing platform is built, and a hardware environment requires computer configuration: a central processor not lower than four cores, a running memory not lower than 4GB and an image processing unit not lower than 2 GB; the software environment comprises a compiling environment capable of running a deep learning TensorFlow framework; then, constructing a network architecture model based on deep learning; next, full waveform lidar data for training is prepared; then, the prepared data are sent into a network architecture model for training, and the number of training iterations is preset; subsequently, a TensorBoard is adopted to observe the change condition of a Loss function Loss and a PSNR after each round of training, and parameters in a network architecture model are dynamically adjusted according to the change condition; and finally, saving a network architecture model with the best training effect, and processing full-waveform laser radar data by using the model to improve the space-time resolution of the original data.
The deep learning method for acquiring the full-waveform laser radar high-resolution data comprises the following steps of:
step one: constructing a network architecture model based on a residual error method in deep learning;
the network architecture model comprises data input, data preprocessing, residual convolution feature extraction, data output and loop optimization;
step two: the network architecture model reads full-waveform laser radar data in txt format to realize data input;
step three: carrying out original full-waveform laser radar data preprocessing on the original full-waveform laser radar data read in the second step, summing every n adjacent points of the original full-waveform laser radar data sequence, and taking an average value to generate a new data point by the data preprocessing, so as to obtain a new full-waveform laser radar data sequence with the length of 1/n of the original full-waveform laser radar data;
step four: the residual stacking module of the network architecture model adopts a residual convolution method to extract waveform characteristics of the novel full-waveform laser radar data sequence generated in the step three;
the method for adopting the residual neural network can effectively solve the gradient dispersion and disappearance problem caused by the too deep network layer, the residual block structure of the constructed network architecture model is different from the residual block structure of the pattern recognition, a single residual block is composed of a convolution layer, an activation function and a convolution layer, and a plurality of residual blocks form a residual stacking module of the network architecture model;
step five: connecting a residual stacking input node with an output node based on a residual neural network jump connection method;
step six: performing up-sampling processing on the data after the convolution output of the last layer of the residual stacking module;
based on a pixel rearrangement mapping method, finally stacking multi-layer residual blocks, convoluting the output data, up-sampling by adopting a pixel rearrangement mapping method in similar image processing, and if the data preprocessing sums every n adjacent points of an original full-waveform laser radar data sequence to obtain an average value to generate a new data point, the space-time resolution is improved by n, the number of convolution kernels of the last layer is also n, so that the number of data points obtained through convolution is matched with the number of data points expected by up-sampling output, and finally mapping to corresponding positions one by one to obtain the output data sequence;
step seven: defining and calculating a Loss function Loss;
defining a Loss function of the network as an absolute value difference value of data at a position corresponding to the new data sequence output in the step six and the input data sequence, accumulating and averaging, namely L1 norm, and calculating a Loss value after initial data input;
step eight: presetting the number of iteration rounds of a network architecture model, and performing training optimization;
the training process aims at optimizing the Loss function Loss, and under the condition that the parameters are reasonably set, the Loss is gradually reduced and converged to smaller values;
step nine: the network architecture model with the minimum loss function and the best training effect in training is stored, the model is used for processing the original full-waveform laser radar data, and the full-waveform laser radar data sequence with the best space-time resolution improving effect can be obtained.
Advantageous effects
1. According to the deep learning-based residual neural network method, a network architecture model for improving the space-time resolution of full-waveform laser radar data is constructed, and the problem that the space-time resolution of the full-waveform laser radar data is limited by a radar hardware system is solved.
2. The full-waveform laser radar data is processed in a deep learning training mode, after the network architecture model is trained, the model with excellent training effect is saved, the saved network architecture model is only required to be called for processing the new full-waveform laser radar data, the calculation resources consumed by program operation are very little, and the full-waveform laser radar data processing method can be applied to various occasions with high requirements on the volume and portability of the acquisition equipment.
Drawings
FIG. 1 is a diagram of a residual neural network architecture model for processing full-wave lidar data according to the present invention;
FIG. 2 is a diagram of a residual block structure employed by a computer vision neural network for pattern recognition;
FIG. 3 is a diagram of a residual block structure constructed in accordance with the present invention that is different from that used for pattern recognition;
fig. 4 is a waveform diagram of various full waveform lidar data: fig. 4 (a) is a full-waveform lidar waveform diagram added with gaussian white noise, the waveform is the most original input waveform, fig. 4 (b) is a full-waveform lidar waveform diagram without gaussian white noise, the waveform is used for loop comparison and Loss, fig. 4 (c) is a full-waveform lidar waveform diagram obtained by downsampling in a preprocessing step, and fig. 4 (d) is a full-waveform lidar waveform diagram output in a training process after 1000 iterations of a network loop;
fig. 5 is a diagram showing a result of processing full-waveform lidar data using a network architecture model with good effect after parameter adjustment in the case of resolution enhancement scale n=4: FIG. 5 (a) is a waveform diagram of input raw data, the sequence of segments containing 10000 data points, and FIG. 5 (b) is a comparative waveform diagram of input raw full-waveform laser radar data without noise;
fig. 6 is a detailed enlarged effect diagram of the output waveform: fig. 6 (a) is a detailed enlarged view of an input full-waveform lidar waveform, and fig. 6 (b) is a detailed enlarged view of an output full-waveform lidar waveform;
fig. 7 is a detailed diagram of a full waveform laser radar waveform training process with four different shapes, wherein the horizontal axis represents the detected ground feature height, the vertical axis represents the echo intensity, and three detailed waveform diagrams from top to bottom in fig. 7 (a), 7 (b), 7 (c) and 7 (d) are sequentially: inputting a trained waveform with noise, obtaining a smooth waveform with distortion although noise after downsampling, and outputting a corresponding waveform after training iteration is finished;
fig. 8 is an output waveform diagram of a full waveform laser radar waveform of two different shapes processed by using a trained and well-effective network architecture model under the condition of resolution upscaling by 4 times, wherein the horizontal axis represents the detected ground object height, and the vertical axis represents the echo intensity: the top waveforms in fig. 8 (a) and 8 (b) are the original waveforms with serious noise pollution in the input process, the bottom waveforms in fig. 8 (a) and 8 (b) are full waveform laser radar waveforms with substantially disappeared noise after the processing, the data sequence is expanded, and the resolution is improved.
Detailed Description
The process according to the invention is described in further detail below with reference to the accompanying drawings and examples.
Example 1
The embodiment processes the output data of a large-broadband full-waveform laser radar system of a certain model SK-300, the peak power of a laser used by the full-waveform laser radar system is 10kw, the pulse width is 2ns, the effective pulse frequency is 10kHz, and 10000 data points can be obtained by scanning a ground object target for 1 second in theory;
step one: constructing a data processing platform for running a deep learning method;
the hardware devices used in the embodiments are: the embodiment of the computer platform is provided with an Intel Core i9-9900k central processor and a Nvidia GeForce RTX 2080-Ti graphic processor, and the software used by the embodiment comprises an open source operating system Ubuntu 16.04,CUDA 10.0.105 driver based on a Linux kernel and a Tensorflow_GPU-1.4.0 neural network framework;
step two: based on the constructed data platform, constructing a residual neural network architecture model for processing full-waveform laser radar data shown in the figure 1, wherein the network architecture model comprises data input, data preprocessing, residual convolution characteristic extraction, data output and loop-back optimization;
step three: constructing a residual block structure shown in figure 3, wherein the constructed residual block structure is different from the residual structure for pattern recognition shown in figure 2, and extracting waveform characteristics of a full waveform laser radar data sequence by adopting a residual convolution method;
step four: reading a full waveform laser radar data file in a txt format;
for reading and calling of full-waveform laser radar original data, data files in different formats can be firstly converted into a. Txt format by using a lastol toolbox and then uniformly processed, in the embodiment, a full-waveform laser radar data file with 10000 data points is input into a network architecture model for data preprocessing, and input data is recorded as x in As shown in fig. 4 (a), the original full-waveform lidar waveform diagram contains gaussian white noise, each gaussian waveform can be resolved into a tree/shrub layering model, and a total of 200 pieces of full-waveform lidar data of this type are used for training, and the waveform detail characteristics of the data are shown in the top graphs in fig. 7 (a), fig. 7 (b), fig. 7 (c) and fig. 7 (d);
step five: performing downsampling operation on input data;
the method comprises the steps of collecting, quantifying, analog-to-digital converting and receiving original echo signals of the analog full-waveform laser radar by a system, summing every four data points in original sequence data to obtain an average value, obtaining 2500 new data points in total, and recording the one-dimensional sequence data as x down_sample X obtained by downsampling original sequence down_sample The waveform is shown in fig. 4 (c), the downsampling process can be regarded as the discrete sampling process of the full waveform laser radar system on the echo signals, and the waveform detail characteristics of the obtained new data after the mean value sampling are shown in fig. 7 (a) and attached drawingsFig. 7 (b), fig. 7 (c), and fig. 7 (d) are central diagrams;
step three: a first layer convolution operation of the residual neural network;
extracting waveform characteristics in original full-waveform laser radar data by adopting convolution operation, dynamically adjusting a training network, extracting characteristics of the full-waveform laser radar data by adopting one-dimensional convolution, and obtaining x down_sample Obtaining x through one-layer convolution conv_1 The convolution kernel size selected by convolving the one-dimensional data is 3*1, the moving step length is set to be 1, and then the one-dimensional data is input into a residual block stacking part of the network architecture model for processing;
step four: processing the data after the first layer convolution by using a residual error network module;
the residual network module is composed of a plurality of residual blocks in a stacked manner, each individual residual block can be divided into three parts of convolution, activation function and convolution, and data after the output of the residual block is convolved with x after one layer of convolution conv_1 Adding to obtain x conv_n In an embodiment, the convolutional layer parameter setting in the residual block is consistent with the initial round convolution, the activation function uses a ReLU activation function, the residual network module is composed of 9 residual blocks stacked, and the data processed and output by the residual network module is recorded as x Res ，x Res Then a convolution layer composed of two convolution kernels is passed to make the dimension of the data to be output match with the expected output, and the data to be output after the convolution of the layer is recorded as x conv-n ；
Step five: up-sampling the data based on the pixel rearrangement idea;
x for upper layer conv-n The sequence data can be regarded as a plurality of low-resolution full-waveform laser radar signals containing different information, a one-to-one mapping method based on a pixel rearrangement method is adopted to map data points in the low-resolution full-waveform laser radar signals to corresponding positions one by one to obtain a brand new data sequence, which is recorded as x est Where x is est 10000 data points are contained in the sequence, the dimension of the data points is consistent with that of input data, and a Loss function Loss is defined as x in And x est Absolute value difference and accumulated average valueFIG. 4 (b) is a waveform diagram of a full waveform lidar for loop comparison to calculate Loss to eliminate Gaussian white noise, x est Comprises ten thousand data points, is dynamically updated along with iterative training of a network architecture model, and is x after the training stage is completed est The waveform of (a) is shown in FIG. 4 (d), x est The waveform details of (a) are shown in the lower graphs of fig. 7 (a), fig. 7 (b), fig. 7 (c) and fig. 7 (d);
step six: the training round number is preset for iterative training, and a neural network parameter model with the best effect is stored;
presetting the number of training iteration rounds as 1000 times, processing 10 pieces of data in batches each time, starting a training process, observing the change condition of a Loss function Loss in real time through a TensorBoard, optimizing the Loss function by adopting an Adam optimizer in a deep learning framework TensorFlow, and storing a network architecture model with the lowest Loss function in training as a ckpt format file after the training process is finished so as to obtain full-waveform laser radar data with improved resolution based on the parameter model;
step seven: the parameter model stored in the step is used for improving the resolution of the newly input full-waveform laser radar data;
based on the residual neural network parameter model in the file with the format of the ckpt stored in the step six, the residual neural network parameter model is used for processing a brand-new full waveform laser radar data file, after the new original full waveform laser radar data comprises 10000 data points to be input, as shown in fig. 5 (a), waveform detail characteristics of the data are shown in fig. 6 (a), four times of resolution improvement can be obtained through the residual neural network parameter model processing, and sequence output containing 40000 data points can be obtained, as shown in fig. 5 (b), the waveform detail characteristics of the data are shown in fig. 6 (b);
in this stage, the original data waveform containing 10000 data points is tested, the detail characteristic diagram of the input waveform is shown in fig. 5 (a) and the top diagram in fig. 8 (b), the new data waveform containing 40000 data points is shown in fig. 5 (b), the detail characteristic diagram of the output waveform is shown in the bottom diagram in fig. 8 (a) and fig. 8 (b), the noise is generally lost after processing, the data sequence is expanded, and the resolution is improved.
While the foregoing is directed to embodiments of the present invention, other and further embodiments of the invention may be devised without departing from the basic scope thereof, and the scope thereof is determined by the claims that follow.
Claims (2)
1. A deep learning method for acquiring full waveform laser radar high-resolution data is characterized in that: the method comprises the following steps:
step one: constructing a network architecture model based on a residual error method in deep learning;
step two: reading full-waveform laser radar data in txt format by a network architecture model;
step three: carrying out original full-waveform laser radar data preprocessing on the original full-waveform laser radar data read in the second step, summing every n adjacent points of the original full-waveform laser radar data sequence, and taking an average value to generate a new data point by the data preprocessing, so as to obtain a new full-waveform laser radar data sequence with the length of 1/n of the original full-waveform laser radar data;
step four: a residual stacking module of a network architecture model is constructed, the waveform characteristics of the new full waveform laser radar data sequence generated in the step three are extracted by adopting a residual convolution method, and the convolution operation mode involved in the step is conv1d, namely one-dimensional unidirectional convolution operation;
step five: connecting a residual stacking input node with an output node based on a residual neural network jump connection method;
step six: performing up-sampling processing on the data after the convolution output of the last layer of the residual stacking module;
based on a pixel rearrangement mapping method, finally stacking multi-layer residual blocks, convoluting the output data, up-sampling by adopting a pixel rearrangement mapping method in image processing, if the data preprocessing sums every n adjacent points of an original full-waveform laser radar data sequence and then takes an average value to generate a new data point, the space-time resolution is improved by n, the number of convolution kernels of the last layer is also n, so that the number of the data points obtained through convolution is matched with the number of the data points expected by up-sampling output, and finally mapping to corresponding positions one by one to obtain an output one-dimensional waveform sequence;
step seven: defining and calculating a Loss function Loss;
defining a Loss function of the network as an absolute value difference value of data at a position corresponding to the new one-dimensional waveform sequence output in the step six and the input one-dimensional waveform sequence, accumulating and averaging, and calculating a Loss value after initial data input;
step eight: presetting the number of iteration rounds of a network architecture model, and performing training optimization;
the training process aims at optimizing the Loss function Loss, and under the condition that the parameters are reasonably set, the Loss is gradually reduced and converged to smaller values;
step nine: the network architecture model with the minimum loss function and the best training effect in training is stored, the model is used for processing the original full-waveform laser radar data, and the time-space resolution of the full-waveform laser radar data sequence can be improved.
2. The deep learning method for obtaining full waveform lidar high resolution data of claim 1, wherein the deep learning method comprises the steps of: a single residual block is formed by conv1d convolution layer-activation function-conv 1d convolution layer, and multiple residual blocks form a residual stacking module of the network architecture model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911001708.8A CN110929811B (en) | 2019-10-21 | 2019-10-21 | Deep learning method for acquiring full waveform laser radar high-resolution data |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911001708.8A CN110929811B (en) | 2019-10-21 | 2019-10-21 | Deep learning method for acquiring full waveform laser radar high-resolution data |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110929811A CN110929811A (en) | 2020-03-27 |
| CN110929811B true CN110929811B (en) | 2023-07-28 |
Family
ID=69849374
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911001708.8A Active CN110929811B (en) | 2019-10-21 | 2019-10-21 | Deep learning method for acquiring full waveform laser radar high-resolution data |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110929811B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112881812B (en) * | 2021-01-11 | 2022-02-22 | 中国气象科学研究院 | Full-flash real-time positioning method and device based on machine learning coding |
| CN116338628B (en) * | 2023-05-16 | 2023-09-15 | 中国地质大学（武汉） | Laser radar sounding method and device based on learning architecture and electronic equipment |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108734660A (en) * | 2018-05-25 | 2018-11-02 | 上海通途半导体科技有限公司 | A kind of image super-resolution rebuilding method and device based on deep learning |
| CN109064405A (en) * | 2018-08-23 | 2018-12-21 | 武汉嫦娥医学抗衰机器人股份有限公司 | A kind of multi-scale image super-resolution method based on dual path network |
| CN109636716A (en) * | 2018-10-29 | 2019-04-16 | 昆明理工大学 | A kind of image super-resolution rebuilding method based on wavelet coefficient study |
| CN109991602A (en) * | 2019-04-10 | 2019-07-09 | 中国人民解放军国防科技大学 | ISAR image resolution enhancement method based on depth residual error network |
| CN110020989A (en) * | 2019-05-23 | 2019-07-16 | 西华大学 | A kind of depth image super resolution ratio reconstruction method based on deep learning |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108780569B (en) * | 2016-01-08 | 2023-05-05 | 特利丹菲力尔有限责任公司 | System and method for image resolution enhancement |
-
2019
- 2019-10-21 CN CN201911001708.8A patent/CN110929811B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108734660A (en) * | 2018-05-25 | 2018-11-02 | 上海通途半导体科技有限公司 | A kind of image super-resolution rebuilding method and device based on deep learning |
| CN109064405A (en) * | 2018-08-23 | 2018-12-21 | 武汉嫦娥医学抗衰机器人股份有限公司 | A kind of multi-scale image super-resolution method based on dual path network |
| CN109636716A (en) * | 2018-10-29 | 2019-04-16 | 昆明理工大学 | A kind of image super-resolution rebuilding method based on wavelet coefficient study |
| CN109991602A (en) * | 2019-04-10 | 2019-07-09 | 中国人民解放军国防科技大学 | ISAR image resolution enhancement method based on depth residual error network |
| CN110020989A (en) * | 2019-05-23 | 2019-07-16 | 西华大学 | A kind of depth image super resolution ratio reconstruction method based on deep learning |
Non-Patent Citations (1)
| Title |
|---|
| Enhanced Deep Residual Networks for Single Image Super-Resolution;Bee Lim 等;《ArXiv》;20170710;第1-9页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110929811A (en) | 2020-03-27 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109829399B (en) | Vehicle-mounted road scene point cloud automatic classification method based on deep learning | |
| Jin et al. | Stem–leaf segmentation and phenotypic trait extraction of individual maize using terrestrial LiDAR data | |
| CN110363215B (en) | Method for converting SAR image into optical image based on generating type countermeasure network | |
| Zhang et al. | A progressive morphological filter for removing nonground measurements from airborne LIDAR data | |
| CN106408011B (en) | Laser scanning three-dimensional point cloud tree automatic classification method based on deep learning | |
| CN111402227B (en) | Bridge crack detection method | |
| Gobeawan et al. | Modeling trees for virtual Singapore: From data acquisition to CityGML models | |
| CN110929811B (en) | Deep learning method for acquiring full waveform laser radar high-resolution data | |
| CN112859011B (en) | Method for extracting waveform signals of single-wavelength airborne sounding radar | |
| Morsdorf et al. | Close-range laser scanning in forests: towards physically based semantics across scales | |
| CN111898688A (en) | Airborne LiDAR data tree species classification method based on three-dimensional deep learning | |
| CN111652059B (en) | Target identification model construction and identification method and device based on computational ghost imaging | |
| Hui et al. | A mean shift segmentation morphological filter for airborne LiDAR DTM extraction under forest canopy | |
| CN113204998A (en) | Airborne point cloud forest ecological estimation method and system based on single wood scale | |
| Liu et al. | A novel entropy-based method to quantify forest canopy structural complexity from multiplatform lidar point clouds | |
| Yu et al. | LFPNet: Lightweight network on real point sets for fruit classification and segmentation | |
| Feng et al. | PyFLEXTRKR: A flexible feature tracking Python software for convective cloud analysis | |
| CN113989343A (en) | Attention mechanism-based sensor fusion depth reconstruction data driving method | |
| CN115797184B (en) | Super-resolution extraction method for surface water body | |
| CN105931184A (en) | SAR image super-resolution method based on combined optimization | |
| Steinbacher et al. | Integrated full-waveform analysis and classification approaches for topo-bathymetric data processing and visualization in hydroVISH | |
| CN111291815A (en) | Landslide sensitivity map drawing method based on wavelet packet transformation and support vector machine | |
| Tang et al. | Accuracy test of point-based and object-based urban building feature classification and extraction applying airborne LiDAR data | |
| Peters et al. | As good as human experts in detecting plant roots in minirhizotron images but efficient and reproducible: The Convolutional Neural Network “RootDetector” | |
| Cui et al. | Remote sensing image target recognition system based on heapsort |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |