CN113375676B - Detector landing site positioning method based on impulse neural network - Google Patents
Detector landing site positioning method based on impulse neural network Download PDFInfo
- Publication number
- CN113375676B CN113375676B CN202110576033.0A CN202110576033A CN113375676B CN 113375676 B CN113375676 B CN 113375676B CN 202110576033 A CN202110576033 A CN 202110576033A CN 113375676 B CN113375676 B CN 113375676B
- Authority
- CN
- China
- Prior art keywords
- layer
- neural network
- snn
- ann
- value
- 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
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 41
- 238000000034 method Methods 0.000 title claims abstract description 31
- 238000012549 training Methods 0.000 claims abstract description 11
- 239000000523 sample Substances 0.000 claims abstract description 10
- 238000003062 neural network model Methods 0.000 claims abstract description 5
- 238000007781 pre-processing Methods 0.000 claims abstract description 4
- 238000012360 testing method Methods 0.000 claims abstract description 4
- 230000004913 activation Effects 0.000 claims description 30
- 210000002569 neuron Anatomy 0.000 claims description 22
- 238000003709 image segmentation Methods 0.000 claims description 11
- 230000005284 excitation Effects 0.000 claims description 9
- 238000005070 sampling Methods 0.000 claims description 8
- 238000011176 pooling Methods 0.000 claims description 4
- 238000010586 diagram Methods 0.000 claims description 3
- 210000002364 input neuron Anatomy 0.000 claims description 3
- 238000001208 nuclear magnetic resonance pulse sequence Methods 0.000 claims description 3
- 210000004205 output neuron Anatomy 0.000 claims description 3
- 238000012216 screening Methods 0.000 claims description 3
- 230000001960 triggered effect Effects 0.000 claims description 3
- 238000004088 simulation Methods 0.000 claims description 2
- 238000001514 detection method Methods 0.000 abstract description 7
- 238000005516 engineering process Methods 0.000 description 8
- 238000005265 energy consumption Methods 0.000 description 4
- 238000013507 mapping Methods 0.000 description 3
- 230000006978 adaptation Effects 0.000 description 2
- 238000013527 convolutional neural network Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 239000011435 rock Substances 0.000 description 2
- 230000011218 segmentation Effects 0.000 description 2
- PEDCQBHIVMGVHV-UHFFFAOYSA-N Glycerine Chemical compound OCC(O)CO PEDCQBHIVMGVHV-UHFFFAOYSA-N 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000004888 barrier function Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000000903 blocking effect Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000006835 compression Effects 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000005305 interferometry Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000008054 signal transmission Effects 0.000 description 1
- 239000013589 supplement Substances 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01C—MEASURING DISTANCES, LEVELS OR BEARINGS; SURVEYING; NAVIGATION; GYROSCOPIC INSTRUMENTS; PHOTOGRAMMETRY OR VIDEOGRAMMETRY
- G01C21/00—Navigation; Navigational instruments not provided for in groups G01C1/00 - G01C19/00
- G01C21/24—Navigation; Navigational instruments not provided for in groups G01C1/00 - G01C19/00 specially adapted for cosmonautical navigation
-
- 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
-
- 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
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- 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
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Abstract
The invention discloses a probe landing site positioning method based on a pulse neural network, which comprises the following steps of: preprocessing the meteorite pit DEM data on the surface of the star; training the training set data by adopting a U-Net network architecture based on an artificial neural network model ANN; converting UNET of the artificial neural network model ANN into UNET model of the impulse neural network model SNN; coding the image, and adopting an UNET model based on SNN to carry out test set detection; and (3) carrying out meteorite pit matching and positioning through a template matching algorithm, constructing a terrain road sign library, and realizing positioning. The invention realizes the positioning of the landing point of the detector by first applying the pulse neural network to the field of deep space detection and identifying the meteorite crater.
Description
Technical Field
The invention belongs to the technical field of vision-based descending-section autonomous navigation of a spacecraft in a deep space exploration task, and particularly relates to a probe landing site positioning method based on a pulse neural network.
Background
After entering the 21 st century, the large aerospace state disputes develop a series of deep space exploration tasks, wherein the soft landing of the detector on the surface of the target star is realized, the safety of equipment and personnel is ensured, and the method is the basis for exploring deep space and developing aerospace technology in China. Typically, the probe positioning technology is through ground measurement and control stations or GNSS navigation technology. However, in a deep space environment, the technology has larger delay, and also has the problems of star shielding, signal transmission blocking and the like, so that the autonomous safe landing technology is a necessary technology for the future detector to land extraterrestrial satellites.
The accurate positioning of the landing points is an important basis for establishing a working coordinate system of the detector, provides position information for the subsequent navigation and scientific detection of the detector, and is an important premise for the detector to develop detection task work. However, the unpredictability of the environment of the extraterrestrial celestial body and the presence of obstructions such as merle pits or rock masses on the star floor present significant challenges to the safe landing of the detector. The current common methods are: computer vision-based landing zone location techniques find application in lunar, mars, and other asteroid exploration tasks. In the goddess Chang E task II, landing points are resolved by using landing camera images, navigation camera images and digital orthophotos, but the positioning result is completed in a few hours after the detector falls into the month; chang's III combines with the very long baseline interferometry technique to locate landing sites; and the goddess Chang E No. four landing zone is the back of moon, because the star shielding, traditional ranging, speed measuring and VLBI technology can not be used in the power reduction stage of the detector, the landing zone positioning is realized on the basis of high-precision image and geometric transformation by adopting a computer vision method. However, the above methods have a problem of real-time performance.
The closest approach to the method is a landing zone positioning technology based on a landform landmark database, namely, a landing image acquired in a descending stage is identified through a convolutional neural network, and the landform landmark database is established through identifying barriers such as merle pits and rocks contained in the image, so that landing point positioning is realized. However, in order to improve the accuracy of the model, a deeper neural network and a large number of data sets are required to perform model training, which brings about a large amount of energy consumption loss and also has a real-time problem.
Disclosure of Invention
The invention provides a method for positioning a landing point of a detector based on a pulse neural network, which solves the problems of larger time delay, low accuracy and higher energy consumption of the existing landing point positioning method.
In order to achieve the above purpose, the invention adopts the following technical scheme:
a probe landing site positioning method based on a pulse neural network comprises the following steps:
s1, preprocessing the data of the meteorite crater DEM on the surface of a star;
s2, training set data by adopting a U-Net network architecture based on an artificial neural network model (artificial neural network, ANN);
s3, converting the UNET of the artificial neural network model ANN into a UNET model of a pulse neural network model (SNN);
s4, encoding the image, and detecting a test set by adopting an UNET model based on SNN;
s5, performing merle pit matching and positioning through a template matching algorithm, constructing a terrain road sign library, and realizing positioning.
Further, the step S1 specifically includes:
s11, the original data set is 184340 x 61440 pixels and 16 bits/pixel, firstly downsampling (down-sampled) is carried out on the original data set to convert the original data set into 92160 x 30720 pixels and 8 bits/pixel data;
s12, randomly cutting the DEM image in the S11 into square areas, and converting the square areas into 256-256 pixel data sets through downsampling;
s13, orthographic projection is carried out on the image in the S12, and other parts are filled with zero;
s14, linearly scaling the image in the S13, and improving the contrast of the image to obtain an input data set;
s15, converting the information into 256-by-256 pixel output labels according to the longitude and latitude and the radius information of the meteorite crater.
Further, the step S2 specifically includes:
s21, inputting the 256-by-256-bit pixel data set constructed in the S14 into an image segmentation network UNET, wherein the left side of the network is a coding layer (decoder), and the right side of the network is a decoding layer (decoder); firstly, carrying out downsampling through 4 blocks 1, wherein each block1 is downsampled by 3 effective convolutions and 1 max pooling, so as to obtain a characteristic diagram; then up-sampling is carried out through 4 blocks 2 to improve the resolution ratio, and then up-sampling results are connected with network left coding layer results through jump connection, so that deep features and shallow features are combined, images are thinned, and feature images are obtained;
s22, predicting the feature map obtained in the S21, and judging whether the current position is the meteorite pit boundary or not through a convolution kernel of 1*1, so as to obtain a 256-by-256-bit pixel value prediction map (predicted mask). However, the output result does not contain information such as the position, the radius and the like of the meteorite crater, so that a template matching algorithm in S5 is also needed, and the specific longitude and latitude information of the identified meteorite crater is used for constructing a landmark library.
Further, the step S3 specifically includes:
s31, acquiring a trained UNET model in S22, and converting model weight W and bias B through scale factors (scales) to ensure that pulse signals exist in a deep network; let us assume that there is a layer of l in UNET, where l e {1,2, 3..l-1, l }, W l ann And b l ann For trained ANN-based network layer I weights and biases, W l snn And b l snn For the scaled weight and bias of the first layer of SNN network, acquiring the activation value { a } l 1 ,a l 2 ,a l 3 .., and remembering a maximum activation value of lambda l ＝max{a l 1 ,a l 2 ,a l 3 .. the scaling factor of each layer is the ratio between the maximum activation value of the previous layer and the maximum activation value of the current layer, i.e. the scaling factor scale of the first layer l Maximum activation value lambda for layer l-1 l-1 Divided by the maximum activation value lambda of the first layer l ,scale l ＝λ l-1 /λ l The method comprises the steps of carrying out a first treatment on the surface of the W is then l snn ＝W l ann *scale l ，b l snn ＝b l ann *λ l ；
S32, the activation value of each layer in the traditional neural network is required to be encoded in the impulse neural network and converted into discrete impulses; in the leaky-distributed integral neuron model (Leaky integrate fire, LIF), assuming a total time window of T, the voltage value V at the first layer, time T mem,j l (t) is the voltage value V at the previous time t-1 mem,j l (t-1) And the input current z at the current time t j l (t), i.e. V mem,j l (t)＝V mem,j l (t-1)+z j l (t) wherein:θ i l (t)＝U(V mem,i l (t)-V th )，w i,j l representing the weight of neuron i connected to neuron j of the first layer, where i represents a neuron in the previous layer of neuron l-1, θ i l-1 (t) shows whether or not the layer 1 neuron i is pulsed at time t, V th Is threshold, U is a step function, θ i l (t) indicates whether the first layer neuron i is pulsed at time t, and the value is 0 or 1, and the current voltage value exceeds the threshold value V th When then theta i l (t) 1 and resetting the voltage value to 0, if the current voltage value has not reached the threshold value, θ i l (t) 0, the voltage value continues to accumulate until a pulse can be triggered;
s33, counting the total number of pulses issued by each output neuron in the total time window T in the last layer, and converting the total number of pulses into pixel values of an output image to obtain a prediction result.
Further, the step S4 specifically includes:
s41, because in the pulse neural network, the input and the activation value of each layer are transmitted in the form of pulse signals, the input image is required to be encoded; the common encoding method of the input image is a poisson encoding method, namely, in a time window T with a certain length, the number of pulses issued by the input neurons corresponding to each pixel is distributed in poisson, and the formula is as follows:wherein: p denotes the probability of issuing k pulses within the total time window T; x is the case of issuing k pulses; k is the number of excitation pulses; μ is the excitation frequency proportional to the pixel value; e is the base of the natural logarithm, here taking the value 2.718, then within the total time window T,each cycle is multiplied by the probability of the time step dt of each simulation by the excitation frequency to determine whether a pulse is issued, and the dt is usually 1, so that a frequency code is finally formed;
s42, transmitting the input pulse sequence into a pulsed UNET model to obtain an image segmentation result.
Further, the step S5 specifically includes:
s51, the merle pit obtained through image segmentation in the S4 is subjected to a template matching algorithm (match-template) to obtain the position and the radius of the current merle pit, namely a landmark library based on the merle pit is obtained, and the detector obtains the current position by identifying the information of the current merle pit;
s52, screening and detecting repeated merle pits by comparing the obtained position and radius information, and perfecting a road sign warehouse based on the merle pits.
Compared with the prior art, the invention has the following beneficial effects:
the invention realizes the positioning of the landing point of the detector by first applying the pulse neural network to the field of deep space detection and identifying the meteorite crater. Firstly, processing the DEM image to obtain an image and a label required by training and testing; training a traditional artificial intelligence UNET model by using a training set, and converting the trained UNET model into a pulse neural network version (Spiking-UNET, SUNET); and finally, using a template matching algorithm to find out the specific longitude and latitude information and the radius of each meteorite crater, thereby realizing the positioning based on a road sign library. The invention has the advantages that the impulse neural network is an event-driven mechanism, and only the impulse release is the loss of energy consumption, so that the energy consumption can be greatly reduced; meanwhile, experiments show that the convolution kernel is 3*3, so that the detection effect of the UNET network on the merle pit with larger radius is poor, and SUNET can solve the problem, and the integrity of a terrain library and the positioning accuracy are improved.
Drawings
FIG. 1 is a flow chart of the present invention.
Fig. 2 is a flowchart for converting CNN to SNN.
FIG. 3 is a map of the merle-pit detection network SUNET model based on a pulsed neural network.
Detailed Description
The invention will be further illustrated with reference to examples.
As shown in fig. 1, a deep space probe landing site positioning method based on a pulse neural network comprises the following steps:
s1, preprocessing the data of the meteorite crater DEM on the surface of a star;
s11, the original data set is 184340 x 61440 pixels and 16 bits/pixel, firstly downsampling (down-sampled) is carried out on the original data set to convert the original data set into 92160 x 30720 pixels and 8 bits/pixel data;
s12, randomly cutting the DEM image in the S11 into square areas, and converting the square areas into 256-256 pixel data sets through downsampling;
s13, orthographic projection is carried out on the image in the S12, and other parts are filled with zero;
s14, linearly scaling the image in the S13, and improving the contrast of the image to obtain an input data set;
s15, converting the information into 256-by-256 pixel output labels according to the longitude and latitude and the radius information of the meteorite crater.
S2, training set data by adopting a U-Net network architecture based on an artificial neural network model (artificial neural network, ANN);
s21, inputting the 256-by-256-bit pixel data set constructed in the S14 into an image segmentation network UNET, wherein the left side of the network is an encoding layer (decoder), and the right side of the network is a decoding layer (decoder). Firstly, carrying out downsampling through 4 blocks 1, wherein each block1 is downsampled by 3 effective convolutions and 1 max pooling, so as to obtain a characteristic diagram; and then up-sampling is carried out through 4 blocks 2 to improve the resolution, and then the up-sampling result is connected with the network left coding layer result through jump connection, so that the combination of deep features and shallow features is realized, the image is thinned, and a feature map is obtained.
S22, predicting the feature map obtained in the S21, and judging whether the current position is the meteorite pit boundary or not through a convolution kernel of 1*1, so as to obtain a 256-by-256-bit pixel value prediction map (predicted mask). However, the output result does not contain information such as the position, the radius and the like of the meteorite crater, so that a template matching algorithm in S5 is also needed, and the specific longitude and latitude information of the identified meteorite crater is used for constructing a landmark library.
Specifically, as shown in fig. 2, the image segmentation architecture based on the conventional neural network is trained first, and then the model is converted into the SNN architecture after mapping conversion from ANN to SNN. The main framework of the merle segmentation model of the network model M adopts a U-Net network with good effect on image segmentation, wherein the U-Net network comprises 4 layers of downsampling, each layer comprises 2 convolution kernels of 3*3, feature extraction is carried out, and 2 x 2 average pooling is carried out on the layers for downsampling. After 4-layer downsampling, reduction is performed by 4-layer upsampling. By 2 x 2 up-convolutions between layers, information is supplemented by skip connection because there is information loss in each up-convolution, and each layer contains 2 3*3 convolution kernels, and the last layer outputs the prediction result for each layer. Comparing the predicted result with the label data through a Dice loss functionWherein: x represents a predicted value, y represents a real value, x is equal to y represents a part where the predicted value and the real value coincide, x is equal to y represents a sum of the predicted value and the real value, the smaller the loss function L value is, the more accurate the segmentation result is represented, the weight is continuously adjusted through back propagation, and finally the loss function value is minimum, so that a final training model is obtained.
S3, converting the UNET of the artificial neural network model ANN into a UNET model of a pulse neural network model (SNN);
s31, acquiring a trained UNET model in S22, wherein in order to ensure that pulse signals exist in a deep network, model weights W and offsets B are converted through scale factors (scales); let us assume that there is a layer of l in UNET, where l e {1,2, 3..l-1, l }, W l ann And b l ann For trained ANN-based network layer I weights and biases, W l snn And b l snn For the scaled weight and bias of the first layer of SNN network, acquiring the activation value { a } l 1 ,a l 2 ,a l 3 .., and remembering a maximum activation value of lambda l ＝max{a l 1 ,a l 2 ,a l 3 .. the scaling factor of each layer is the ratio between the maximum activation value of the previous layer and the maximum activation value of the current layer, i.e. the scaling factor scale of the first layer l Maximum activation value lambda for layer l-1 l-1 Divided by the maximum activation value lambda of the first layer l ,scale l ＝λ l-1 /λ l The method comprises the steps of carrying out a first treatment on the surface of the W is then l snn ＝W l ann *scale l ，b l snn ＝b l ann *λ l ；
S32, the activation value of each layer in the traditional neural network is required to be encoded in the impulse neural network and converted into discrete impulses; in the leaky-distributed integral neuron model (Leaky integrate fire, LIF), assuming a total time window of T, the voltage value V at the first layer, time T mem,j l (t) is the voltage value V at the previous time t-1 mem,j l Input current z at (t-1) and current time t j l (t), i.e. V mem,j l (t)＝V mem,j l (t-1)+z j l (t) wherein:θ i l (t)＝U(V mem,i l (t)-V th )，w i,j l representing the weight of neuron i connected to neuron j of the first layer, where i represents a neuron in the previous layer of neuron l-1, θ i l-1 (t) shows whether or not the layer 1 neuron i is pulsed at time t, V th Is threshold, U is a step function, θ i l (t) indicates whether the first layer neuron i is pulsed at time t, and the value is 0 or 1, and the current voltage value exceeds the threshold value V th When then theta i l (t) 1 and resetting the voltage value to 0, if the current voltage value has not reached the threshold value, θ i l (t) 0, the voltage value continues to accumulate until a pulse can be triggered;
s33, counting the total number of pulses issued by each output neuron in the total time window T in the last layer, and converting the total number of pulses into pixel values of an output image to obtain a prediction result.
Specifically, converting the traditional neural network model into a pulsed neural network includes: image pulsing, weight mapping, convolution layer and upsampling layer pulsing, as shown in fig. 3. To avoid too many or how many pulses are in the network, the weights are mapped. The DEM image (input DEM) is then poisson encoded and converted into a pulse input network. Image compression is carried out on the left side of the network, the network comprises 4 layers in total, each layer comprises 2 convolution kernels, the weight value in each convolution kernel is obtained by mapping the weight value in S2, and in a time window T, whether the neuron emits a pulse to enter the next layer at the current moment is judged by calculating a voltage value through a spiking-relu function; the average sampling based on the pulse neural network is adopted between the layers, and the pulse average value at the current moment is counted; the right side of the network performs image decoding, which also includes 4 layers, firstly restores through an up-sampling layer based on SNN, supplements image information through jump connection in order to reduce image information loss, and finally outputs a prediction result (predicted mask).
S4, coding the image, and detecting merle pits by adopting an UNET model based on SNN;
s41, because in the pulse neural network, the input and the activation value of each layer are transmitted in the form of pulse signals, the input image is required to be encoded; the common encoding method of the input image is a poisson encoding method, namely, in a time window T with a certain length, the number of pulses issued by the input neurons corresponding to each pixel is distributed in poisson, and the formula is as follows:wherein: p denotes the probability of issuing k pulses within the total time window T; x is the case of issuing k pulses; k is the number of excitation pulses; μ is the excitation frequency proportional to the pixel value; e is the base of the natural logarithm, here taking the value 2.718, then each cycle is multiplied by each modulus at the excitation frequency within the total time window TThe probability of the quasi-time step dt determines whether a pulse is issued, and the dt is generally 1, and finally a frequency code is formed; .
S42, transmitting the input pulse sequence into a pulsed UNET model to obtain an image segmentation result.
S5, performing merle pit matching and positioning through a template matching algorithm, constructing a terrain road sign library, and realizing positioning.
S51, the merle pit obtained through image segmentation in the S4 is subjected to a template matching algorithm (match-template) to obtain the position and the radius of the current merle pit, so that a landmark library based on the merle pit can be obtained, and the detector can obtain the current position by identifying the information of the current merle pit.
S52, screening and detecting repeated merle pits by comparing the obtained position and radius information, and perfecting a road sign warehouse based on the merle pits.
The foregoing is only a preferred embodiment of the invention, it being noted that: it will be apparent to those skilled in the art that various modifications and adaptations can be made without departing from the principles of the present invention, and such modifications and adaptations are intended to be comprehended within the scope of the invention.
Claims (6)
1. The method for positioning the landing site of the detector based on the impulse neural network is characterized by comprising the following steps of:
s1, preprocessing the data of the meteorite crater DEM on the surface of a star;
s2, training the training set data by adopting a U-Net network architecture based on an artificial neural network model ANN;
s3, converting the UNET of the artificial neural network model ANN into a UNET model of the impulse neural network model SNN;
the step S3 includes a step S31,
s31, acquiring a trained UNET model in the S2, and converting model weight W and bias B through scaling factors to ensure that pulse signals exist in a deep network; let us assume that there is a layer of l in UNET, where l ε {1,2,3 … l-1, l }, W l ann And b l ann For trained basesWeights and biases at layer I of ANN network, W l snn And b l snn For the scaled weight and bias of the first layer of SNN network, acquiring the activation value { a } l 1 ,a l 2 ,a l 3 .., and remembering a maximum activation value of lambda l ＝max{a l 1 ,a l 2 ,a l 3 .. the scaling factor of each layer is the ratio between the maximum activation value of the previous layer and the maximum activation value of the current layer, i.e. the scaling factor scale of the first layer l Maximum activation value lambda for layer l-1 l-1 Divided by the maximum activation value lambda of the first layer l ,scale l ＝λ l-1 /λ l The method comprises the steps of carrying out a first treatment on the surface of the W is then l snn ＝W l ann *scale l ，b l snn ＝b l ann *λ l ；
S4, encoding the image, and detecting a test set by adopting an UNET model based on SNN;
s5, performing merle pit matching and positioning through a template matching algorithm, constructing a terrain road sign library, and realizing positioning.
2. The method for positioning a probe landing site based on a pulsed neural network according to claim 1, wherein the step S1 is specifically:
s11, an original data set is 184340 x 61440 pixels and 16bit/pixel, firstly downsampling is carried out on the original data set, and downsampling is carried out on the original data set to convert the original data set into 92160 x 30720 pixels and 8bit/pixel data;
s12, randomly cutting the DEM image in the S11 into square areas, and converting the square areas into 256-256 pixel data sets through downsampling;
s13, orthographic projection is carried out on the image in the S12, and other parts are filled with zero;
s14, linearly scaling the image in the S13, and improving the contrast of the image to obtain an input data set;
s15, converting the information into 256-by-256 pixel output labels according to the longitude and latitude and the radius information of the meteorite crater.
3. The method for positioning the probe landing site based on the impulse neural network according to claim 2, wherein the step S2 specifically comprises:
s21, inputting the 256-by-256-bit pixel data set constructed in the S14 into an image segmentation network UNET, wherein the left side of the network is a coding layer decoder, and the right side of the network is a decoding layer decoder; firstly, carrying out downsampling through 4 blocks 1, wherein each block1 is downsampled by 3 effective convolutions and 1 max pooling, so as to obtain a characteristic diagram; then up-sampling is carried out through 4 blocks 2 to improve the resolution ratio, and then up-sampling results are connected with network left coding layer results through jump connection, so that deep features and shallow features are combined, images are thinned, and feature images are obtained;
s22, predicting the feature map obtained in the S21, and judging whether the current position is the meteorite pit boundary or not through a convolution kernel of 1*1, so as to obtain a 256-by-256-bit pixel value predicted mask; however, the output result does not contain the position and radius information of the meteorite crater, so that a template matching algorithm in S5 is also needed, and the specific longitude and latitude information of the identified meteorite crater is needed, so that a landmark library is constructed.
4. The method for positioning a probe landing site based on a pulsed neural network according to claim 3, wherein the step S3 is specifically:
s31, acquiring a UNET model trained in the S22, and converting model weight W and bias B through scaling factors to ensure that pulse signals exist in a deep network; let us assume that there is a layer of l in UNET, where l ε {1,2,3 … l-1, l }, W l ann And b l ann For trained ANN-based network layer I weights and biases, W l snn And b l snn For the scaled weight and bias of the first layer of SNN network, acquiring the activation value { a } l 1 ,a l 2 ,a l 3 .., and remembering a maximum activation value of lambda l ＝max{a l 1 ,a l 2 ,a l 3 .. the scaling factor of each layer is the one of the preceding layerThe ratio between the maximum activation value and the maximum activation value of the current layer, i.e. the scale factor of the first layer l Maximum activation value lambda for layer l-1 l-1 Divided by the maximum activation value lambda of the first layer l ,scale l ＝λ l-1 /λ l The method comprises the steps of carrying out a first treatment on the surface of the W is then l snn ＝W l ann *scale l ，b l snn ＝b l ann *λ l ；
S32, the activation value of each layer in the traditional neural network is required to be encoded in the impulse neural network and converted into discrete impulses; in the leakage issue integral neuron model, assuming that the total time window is T, the voltage value V at the first layer and the T moment mem,j l (t) is the voltage value V at the previous time t-1 mem,j l Input current z at (t-1) and current time t j l (t), i.e. V mem,j l (t)＝V mem,j l (t-1)+z j l (t) wherein:θ i l (t)＝U(V mem,i l (t)-V th )，w i,j l representing the weight of neuron i connected to neuron j of the first layer, where i represents a neuron in the previous layer of neuron l-1, θ i l-1 (t) shows whether or not the layer 1 neuron i is pulsed at time t, V th Is threshold, U is a step function, θ i l (t) indicates whether the first layer neuron i is pulsed at time t, and the value is 0 or 1, and the current voltage value exceeds the threshold value V th When then theta i l (t) 1 and resetting the voltage value to 0, if the current voltage value has not reached the threshold value, θ i l (t) 0, the voltage value continues to accumulate until a pulse can be triggered;
s33, counting the total number of pulses issued by each output neuron in the total time window T in the last layer, and converting the total number of pulses into pixel values of an output image to obtain a prediction result.
5. The method for positioning a probe landing site based on a pulsed neural network according to claim 4, wherein the step S4 is specifically:
s41, because in the pulse neural network, the input and the activation value of each layer are transmitted in the form of pulse signals, the input image is required to be encoded; the common encoding method of the input image is a poisson encoding method, namely, in a time window T with a certain length, the number of pulses issued by the input neurons corresponding to each pixel is distributed in poisson, and the formula is as follows:wherein: p denotes the probability of issuing k pulses within the total time window T; x is the case of issuing k pulses; k is the number of excitation pulses; μ is the excitation frequency proportional to the pixel value; e is a base number of natural logarithms, and in a total time window T, determining whether pulses are issued or not according to the probability that the excitation frequency is multiplied by the time step dt of each simulation in each cycle, and taking 1 for dt to finally form a frequency code;
s42, transmitting the input pulse sequence into a pulsed UNET model to obtain an image segmentation result.
6. The method for positioning a probe landing site based on a pulsed neural network according to claim 5, wherein the step S5 is specifically:
s51, the position and the radius of the current meteorite pit are obtained through a template matching algorithm by the meteorite pit obtained through image segmentation in the S4, namely a landmark library based on the meteorite pit is obtained, and the detector obtains the current position by identifying the information of the current meteorite pit;
s52, screening and detecting repeated merle pits by comparing the obtained position and radius information, and perfecting a road sign warehouse based on the merle pits.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110576033.0A CN113375676B (en) | 2021-05-26 | 2021-05-26 | Detector landing site positioning method based on impulse neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110576033.0A CN113375676B (en) | 2021-05-26 | 2021-05-26 | Detector landing site positioning method based on impulse neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113375676A CN113375676A (en) | 2021-09-10 |
| CN113375676B true CN113375676B (en) | 2024-02-20 |
Family
ID=77572231
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110576033.0A Active CN113375676B (en) | 2021-05-26 | 2021-05-26 | Detector landing site positioning method based on impulse neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113375676B (en) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114519723A (en) * | 2021-12-24 | 2022-05-20 | 上海海洋大学 | Meteorite crater automatic extraction method based on pyramid image segmentation |
| WO2023179516A1 (en) * | 2022-03-23 | 2023-09-28 | 北京华视诺维医疗科技有限公司 | Detection system and method based on intelligent mobile device, and detection model construction method |
| CN115054209B (en) * | 2022-04-14 | 2023-09-26 | 杭州华视诺维医疗科技有限公司 | Multi-parameter physiological information detection system and method based on intelligent mobile equipment |
| CN116432703B (en) * | 2023-06-12 | 2023-08-29 | 成都大学 | Pulse height estimation method, system and terminal based on composite neural network model |
| CN116989800B (en) * | 2023-09-27 | 2023-12-15 | 安徽大学 | Mobile robot visual navigation decision-making method based on pulse reinforcement learning |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019104767A1 (en) * | 2017-11-28 | 2019-06-06 | 河海大学常州校区 | Fabric defect detection method based on deep convolutional neural network and visual saliency |
| CN110210563A (en) * | 2019-06-04 | 2019-09-06 | 北京大学 | The study of pattern pulse data space time information and recognition methods based on Spike cube SNN |
| CN111652896A (en) * | 2020-05-29 | 2020-09-11 | 北京理工大学 | Inertial navigation auxiliary meteorite crater coarse-to-fine detection method |
| CN112580502A (en) * | 2020-12-17 | 2021-03-30 | 南京航空航天大学 | SICNN-based low-quality video face recognition method |
| CN112819794A (en) * | 2021-02-04 | 2021-05-18 | 青岛科技大学 | Small celestial body meteorite crater detection method based on deep learning |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| DE102019212907A1 (en) * | 2019-08-28 | 2021-03-04 | Robert Bosch Gmbh | Method, device and computer program for creating a pulsed neural network |
-
2021
- 2021-05-26 CN CN202110576033.0A patent/CN113375676B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019104767A1 (en) * | 2017-11-28 | 2019-06-06 | 河海大学常州校区 | Fabric defect detection method based on deep convolutional neural network and visual saliency |
| CN110210563A (en) * | 2019-06-04 | 2019-09-06 | 北京大学 | The study of pattern pulse data space time information and recognition methods based on Spike cube SNN |
| CN111652896A (en) * | 2020-05-29 | 2020-09-11 | 北京理工大学 | Inertial navigation auxiliary meteorite crater coarse-to-fine detection method |
| CN112580502A (en) * | 2020-12-17 | 2021-03-30 | 南京航空航天大学 | SICNN-based low-quality video face recognition method |
| CN112819794A (en) * | 2021-02-04 | 2021-05-18 | 青岛科技大学 | Small celestial body meteorite crater detection method based on deep learning |
Non-Patent Citations (1)
| Title |
|---|
| 基于深度学习的大型陨石坑识别方法研究;郑磊等;《北京航空航天大学学报》;第46卷(第5期);第994-1003页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113375676A (en) | 2021-09-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113375676B (en) | Detector landing site positioning method based on impulse neural network | |
| CN111461291B (en) | Long-distance pipeline inspection method based on YOLOv3 pruning network and deep learning defogging model | |
| CN110599537A (en) | Mask R-CNN-based unmanned aerial vehicle image building area calculation method and system | |
| Ma et al. | Automatic detection and counting system for pavement cracks based on PCGAN and YOLO-MF | |
| CN112651437B (en) | Spatial non-cooperative target pose estimation method based on deep learning | |
| CN113052106B (en) | Airplane take-off and landing runway identification method based on PSPNet network | |
| Yu et al. | Capsule feature pyramid network for building footprint extraction from high-resolution aerial imagery | |
| CN104778699B (en) | A kind of tracking of self adaptation characteristics of objects | |
| CN114049356B (en) | Method, device and system for detecting structure apparent crack | |
| CN114241296A (en) | Method for detecting meteorite crater obstacle during lunar landing, storage medium and electronic device | |
| CN114842681A (en) | Airport scene flight path prediction method based on multi-head attention mechanism | |
| Yuan et al. | Encoding time-series ground motions as images for convolutional neural networks-based seismic damage evaluation | |
| CN116309348A (en) | Lunar south pole impact pit detection method based on improved TransUnet network | |
| CN116206098A (en) | Moon surface safety landing zone selection system and method thereof | |
| Zuo et al. | A remote sensing image semantic segmentation method by combining deformable convolution with conditional random fields | |
| CN116309536A (en) | Pavement crack detection method and storage medium | |
| Meng et al. | A modified fully convolutional network for crack damage identification compared with conventional methods | |
| Potnis et al. | Multi-class segmentation of urban floods from multispectral imagery using deep learning | |
| CN116758363A (en) | Weight self-adaption and task decoupling rotary target detector | |
| CN115147720A (en) | SAR ship detection method based on coordinate attention and long-short distance context | |
| CN114842078A (en) | Dual-channel satellite attitude estimation network based on deep learning | |
| CN113609999A (en) | Human body model establishing method based on gesture recognition | |
| Xu et al. | Segmentation of building footprints with xception and iouloss | |
| Xi et al. | Oriented target detection algorithm based on transformer | |
| CN116628251B (en) | Method, device, equipment and medium for searching moon surface safety area |
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 |