CN110378880A - The Cremation Machine burning time calculation method of view-based access control model - Google Patents
The Cremation Machine burning time calculation method of view-based access control model Download PDFInfo
- Publication number
- CN110378880A CN110378880A CN201910585276.3A CN201910585276A CN110378880A CN 110378880 A CN110378880 A CN 110378880A CN 201910585276 A CN201910585276 A CN 201910585276A CN 110378880 A CN110378880 A CN 110378880A
- Authority
- CN
- China
- Prior art keywords
- burning time
- cremation machine
- view
- access control
- control model
- 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.)
- Pending
Links
- 238000004364 calculation method Methods 0.000 title claims abstract description 17
- 238000000034 method Methods 0.000 claims abstract description 19
- 230000011218 segmentation Effects 0.000 claims abstract description 15
- 238000000605 extraction Methods 0.000 claims abstract description 7
- 238000002485 combustion reaction Methods 0.000 claims abstract description 5
- 239000000284 extract Substances 0.000 claims description 8
- 208000020816 lung neoplasm Diseases 0.000 claims description 7
- 238000013139 quantization Methods 0.000 claims description 7
- 238000006243 chemical reaction Methods 0.000 claims description 3
- 238000013528 artificial neural network Methods 0.000 abstract description 10
- 238000013135 deep learning Methods 0.000 abstract description 4
- 230000002708 enhancing effect Effects 0.000 abstract description 3
- 238000013507 mapping Methods 0.000 description 11
- 238000005070 sampling Methods 0.000 description 11
- 238000013527 convolutional neural network Methods 0.000 description 9
- 238000010586 diagram Methods 0.000 description 8
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 description 7
- 230000006870 function Effects 0.000 description 6
- 239000013598 vector Substances 0.000 description 6
- 230000000694 effects Effects 0.000 description 5
- 239000011159 matrix material Substances 0.000 description 5
- 230000009466 transformation Effects 0.000 description 5
- 238000002474 experimental method Methods 0.000 description 4
- 230000006872 improvement Effects 0.000 description 4
- 238000012549 training Methods 0.000 description 4
- 230000004913 activation Effects 0.000 description 3
- 230000008859 change Effects 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 239000000203 mixture Substances 0.000 description 3
- 238000011176 pooling Methods 0.000 description 3
- 241000282326 Felis catus Species 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- PCHJSUWPFVWCPO-UHFFFAOYSA-N gold Chemical compound [Au] PCHJSUWPFVWCPO-UHFFFAOYSA-N 0.000 description 2
- 239000010931 gold Substances 0.000 description 2
- 229910052737 gold Inorganic materials 0.000 description 2
- 238000003475 lamination Methods 0.000 description 2
- 210000002569 neuron Anatomy 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 230000007704 transition Effects 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 241001465754 Metazoa Species 0.000 description 1
- 230000009471 action Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000001186 cumulative effect Effects 0.000 description 1
- 230000004069 differentiation Effects 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 230000002685 pulmonary effect Effects 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000005549 size reduction Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
- 230000017105 transposition Effects 0.000 description 1
- 238000005303 weighing Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10004—Still image; Photographic image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Landscapes
- Engineering & Computer Science (AREA)
- Quality & Reliability (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses the Cremation Machine burning time calculation methods of view-based access control model, the step of including are as follows: the position of corpse is fixed first, the photo of corpse is obtained by camera, it is then based on enhancing Mask-RCNN deep learning network and carries out corpse segmentation, human body pixel number and funeral cremator burning time relationship are judged again by four layers of full link neural network, obtain optimal funeral cremator burning time.The Mask-RCNN method includes three phases: feature extraction, region suggest and prediction.The invention belongs to contactless burning time Predicting Technique, Combustion System program differentiates that result is adjusted Cremation Machine burning time by four layers of full link neural network, so that Cremation Machine burning time parameter setting is more reasonable.
Description
Technical field
The invention belongs to incinerator technical fields, are related to a kind of method for being specially adapted for incinerating human or animal's corpse, tool
Body is related to a kind of Cremation Machine burning time calculation method of view-based access control model.
Background technique
Funeral cremator burning time has particularity, currently, time for being burnt of the corpse of different weight substantially according to
Manually micro-judgment, therefore occur often because judgement inaccuracy, so that burning time is too long or too short, causes bad shadow
It rings.
The key factor for influencing Cremation Machine burning time is the weight of corpse.Because of the particularity of Cremation Machine burning time, no
Contact weighing-appliance convenient to use carries out the weight weighing of corpse, this allows for judging based on contactless carcass weight non-
It is often necessary.
Summary of the invention
To solve the problems, such as that above-mentioned existing funeral cremator burning time judgement inaccuracy, the present invention propose that view-based access control model passes
The estimation of contactless corpse burning time of sense belongs to the estimation of contactless funeral cremator burning time based on deep learning
Method.The photo for obtaining corpse by camera first is then based on enhancing Mask-RCNN deep learning network and carries out corpse point
It cuts, judges human body pixel number and funeral cremator burning time relationship again by four layers of full link neural network, obtain best
Funeral cremator burning time.
To achieve the above object, the technical solution adopted by the present invention is the Cremation Machine burning time calculating side of view-based access control model
Method, comprising the following steps:
Step 1: the position of corpse is fixed;
Step 2: photo is extracted;
Step 3: the region where corpse is determined by Mask-RCNN method;
Step 4: burning time is determined according to the time relationship of corpse pixel and corpse combustion state.
Above-mentioned Mask-RCNN method includes three phases: feature extraction, region suggest and prediction.
In feature extraction phases, using the FPN based on ResNet as the core network of Mask-RCNN, according to Lung neoplasm figure
As designing fine essential characteristic extract layer and simplified subsequent characteristics extract layer.
The FPN includes three parts: path from bottom to top, from top to bottom path and lateral connection.
In the region proposal stage, region suggests network RPN by carrying out sliding window on more size characteristic figures that FPN is provided
Mouth operation is to extract rectangle candidate region.
The Feature Conversion in the frame of candidate region is tieed up with fixed space using the method for RoIAlign in forecast period
The small characteristic pattern of H*W is spent, to solve region mismatch problem caused by quantifying twice in RoIPool operation.
Preferably, tubercle is identified as 7*7, nodule segmentation 14*14 in above-mentioned small characteristic pattern.
It is above-mentioned to be quantified as the quantization of the boundary RoI and elementary boundary quantization twice.
Compared with prior art, the present invention has following advantageous effects:
1, the present invention is based on enhancing Mask-RCNN deep learning networks to carry out human body segmentation, then passes through four layers of full link
Neural network judges human body pixel number and funeral cremator burning time relationship, when available optimal funeral cremator burns
Between.
2, the invention belongs to contactless burning time Predicting Technique, Combustion System program is by above-mentioned differentiation result to fire
Change machine burning time is adjusted, so that Cremation Machine burning time parameter setting is more reasonable.
Detailed description of the invention
Fig. 1 is that the present invention is based on the flow charts of the Cremation Machine burning time calculation method of vision;
Fig. 2 is the Lung neoplasm automatic identification flow chart based on Mask-RCNN;
Fig. 3 is the segmentation schematic diagram based on Mask-RCNN frame;
Fig. 4 is that Mask-RCNN is that three phases schematic diagram is established in segmentation;
Fig. 5 is the FPN flow chart based on ResNet;
Fig. 6 is the system assumption diagram of RPN;
Fig. 7 is the system assumption diagram of ResNet-50;
Fig. 8 is three kinds of ResNet structure optimization schematic diagrames;
Fig. 9 is four layers of fully-connected network schematic diagram.
Specific embodiment
Now in conjunction with attached drawing, the present invention will be further described in detail.
The process of the Cremation Machine burning time calculation method of view-based access control model as shown in Figure 1, the present invention in order to make full use of sky
Between information and priori knowledge, using improved Mask-RCNN algorithm, the spy based on ResNet that is modified according to pulmonary nodule image
It levies pyramid network (FPN), the core network as Mask-RCNN.Then first by the Pixel-level of tubercle using Mask-RCNN
It tests knowledge and extracts the pyramid characteristic pattern for being suitable for image, the tubercle side provided followed by extension by Mask-RCNN identification
Boundary's frame obtains the identification and segmentation that area-of-interest (ROI) carries out tubercle, and the flow chart of this method is as shown in Figure 2.For letter
For the sake of list, the only characteristic pattern after display core network in RPN framework, but after all characteristic patterns all have each layer.
The segmentation schematic diagram of Mask-RCNN frame is as shown in figure 3, the architectural framework of Mask-RCNN is based on convolutional Neural net
The architectural framework of network (CNN), CNN are a kind of deep neural networks being mainly made of various types of layers.The input of CNN is usual
It is no complicated pretreated original image, and the output of each layer of CNN has the characteristic pattern of sizes and a variety of ratios.One
As for, extended with region is experienced due to the intensification of layer, extracted feature from low-level become high level.CNN is used
Back-propagation algorithm optimizes the weight of all weighting layers by minimizing loss, and the design of loss function can be defined as not
Same image task.In general, there are three types of major type of layers: convolutional layer (Convolutional Layer), pond layer for CNN tool
(Pooling Layer), full articulamentum (Fully Connected Layer).
Traditional neural network by establishing connection, weight multiplied by weight matrix for input matrix between input and output
Matrix is the operation being fully connected, therefore very time-consuming, so only allow to train some features, in order to more effectively extract feature,
For convolutional layer using part connection, it both horizontally and vertically moves the convolutional filtering of a small fixed size along input matrix
Device generates linear activation response in the form of characteristic pattern.
Each neuron in characteristic pattern is connected to the local receptor field of input matrix, and is obtained by two steps.?
In the first step, receptive field is multiplied by it with the value of corresponding position in filter, is then added the value of multiplication to obtain end value, institute
There is neuron to share the weight of filter.In order to carry out feature extraction in a manner of more, multiple filters in convolutional layer can
With parallel work-flow to generate multiple characteristic patterns, local feature is then extracted.Gradient disappears or explodes in order to prevent and to improve CNN anti-
To the efficiency of propagation, amendment linear unit (ReLU) is fed to by the characteristic pattern that filter generates.Then after ReLU being handled
Characteristic pattern be input to the layer (pond layer) of next type, and by select the whole statistical nature of small neighbourhood replace field come into
Row double sampling, whole statistical nature in this case use maximum value (referred to as maximum pond).Pond layer has CNN
There are translation, rotation and scaling invariance, meanwhile, the size reduction of feature vector can prevent network over-fitting and further drop
Low computation complexity, also results in loss in detail.Another important layer is full articulamentum, and filter size is that it is to tool
There are the advanced features of the strong semantic information suitable for different images task to be encoded, learns and adjust convolutional layer and full articulamentum
Weight can preferably indicate input data in the training process.
As shown in figure 4, Mask-RCNN is that three phases are established in segmentation: feature extraction, region suggest and prediction.Wherein, 4
(a) learn the flow chart of block, the convolution block that 4 (b) present invention use, 4 (c) identical pieces used for the present invention for residual error.
In the first stage, it using the FPN based on ResNet as the core network of Mask-RCNN, is set according to Lung neoplasm image
The subsequent characteristics extract layer counting out fine essential characteristic extract layer and simplifying.Recent studies indicate that network depth is to net
Network study is of great significance, but has certain side effect to e-learning, such as gradient disappearance/explosion, training accuracy.
Previous problem is solved via specification initialization and intermediate cannonical layer, and ResNet can be by being applied to CNN for residual block
Solve Second Problem, as shown in Figure 4.
Residual block is defined as
Y=F (x)+x (1.1)
Wherein input feature vector mapping is expressed as x, and the output for merging layer is expressed as y, and output y can pass through x and F (x) corresponding positions
The weight set is added to obtain.Therefore, the output F (x) of three convolutional layers with ReLU becomes the mapping of the residual error from x to y.It is identical
Mapping is that input feature vector maps and merges the quick connection between layer, need to only execute the identical mapping from x to y.So due to defeated
The weight ratio of the reference entered, convolutional layer is easier to learn without fast connection.In ResNet, other than first layer convolutional layer,
Nearly all weighting layer is all made of residual block.Due to the requirement of area size, the necessary extended area range of block, therefore there are two types of
The block of type, i.e. convolution block (convolutional block) and identical piece (identity blocks).Convolution block can lead to
Crossing increases to 2 from 1 by the step-length of first layer convolutional layer to increase regional scope, and for matching dimensionality, quick connection should also be incited somebody to action
Step-length increases to 2..Identical piece is calibrated bolck used in ResNet, and it is identical to output and input dimension.Scheme in (b) and (c) the
One, two, the size of three-layer coil product core is 1 × 1,3 × 3,1 × 1, and the effect of first layer and third layer convolutional layer is to reduce feature dimensions
Degree, dimension here refers to port number, so that second layer convolutional layer has less weight to go to learn.This design is reliable
, and be proved that precision can be effectively improved.The following table 1 provides the details for the ResNet structure that the present invention uses.
Table 1
Although ResNet has powerful expressive force, verified to solve multiple dimensioned figure by using pyramid representation
As task can obtain better performance.Therefore, using the multiple dimensioned spy extracted based on the FPN of ResNet in method
Sign.
FPN includes three parts: path from bottom to top, from top to bottom path and lateral connection, as shown in Figure 5.FPN includes
Three parts: path from bottom to top, top-down path and lateral connection.Path from bottom to top is the feedforward of ResNet
It calculates, there is identical scale in each stage, and layer is deeper, and character representation is better, so utilizing the last one residual error of each stage
Block output activates to form the individual features mapping set of pyramid level.
Top-down path first up-samples characteristic pattern with rougher spatial resolution, but from higher gold
The semantic information that word tower rank is extracted wants being higher by twice for relatively low pyramid rank extraction.Then by the feature of up-sampling with
Feature from bottom to top merges accordingly, these features from bottom to top are by one 1 × 1 convolutional layer (to reduce pyramid
The port number of all ranks, and make its port number 256) be added by element, which is known as lateral connection, by top-down
Each stage in path executes, as shown in the figure.Then, one 3 × 3 convolution is added in each combined Feature Mapping
Layer, to reduce the aliasing effect generated by up-sampling.The output of all 3 × 3 convolutional layers constitutes us based on ResNet's
The final Feature Mapping set of FPN core network, the set expression be { P2, P3, P4, P5 }, pay attention to P6 on the basis of P5 into
The maximum pondization that step-length of having gone is 2 operates, and for detecting lesser target, high-rise characteristic pattern is used for the characteristic pattern of such bottom
Detect biggish target.
In second stage, region suggests network (RPN) by carrying out sliding window on more size characteristic figures that FPN is provided
Operation is to extract rectangle candidate region.
RPN is a mininet, as shown in fig. 6, rectangular window therein can be in the feature gold word of above-mentioned core network
It is slided on tower, the feature pyramid rank for executing slide is determined by ROI size.Then, sliding window is passed through into filter
Convolutional layer having a size of 3 × 3 is mapped to 512 dimensional vectors, and by two 1 × 1 parallel convolutional layers of the vector feed-in, and one point
For classifying, another branch returns branch for bounding box.K candidate region frame is extracted in each sliding window position, therefore
Bounding box returns layer and exports 4K coordinate, and classification layer exports 2K score (probability for predicting each candidate region prospect, background),
K candidate region frame is parameterized relative to reference block, these reference blocks are referred to as anchor and centered on sliding window.
The selection of anchor is based on anchor and true value (Ground Truth) friendship and than (IOU), and the highest anchor of IOU is be overlapped with single GT phase
Folded, anchor of the IOU higher than 0.7 is assigned to positive sample in all GT, and anchor of the IOU lower than 0.3 is assigned to negative sample.Red block
It is the window slided on the feature pyramid of above-mentioned core network, the loss function of RPN is defined as:
Wherein, piIndicate that i-th of anchor of prediction is the probability of target.If anchor is positive sample, GT labelIt is 1, it is on the contrary
?.tiIndicate the parametrization coordinate that the bounding box of RPN prediction returns,Indicate the GT coordinate for corresponding to positive sample anchor point.Lcls
Indicate the logarithm loss of binary classification (target/non-targeted), LregIndicate the smooth L1 loss of predicted boundary frame and GT frame.RPN's
Loss function is by NclsAnd NregStandardize, and is weighted by a balance parameters λ.Finally, RPN passes through training adjustment
Network weight obtains a series of candidate region frames.
In the phase III, by a kind of technology for being known as RoIAlign, by the Feature Conversion in the frame of candidate region be with
The small characteristic pattern (tubercle is identified as 7*7, due to required segmentation precision, nodule segmentation 14*14) of fixed space dimension H*W,
The technology has well solved region caused by quantization twice in RoIPool operation (quantization of the boundary RoI and elementary boundary quantization)
Mismatch problem.The bounding box identification and mask segmentation of tubercle are completed followed by Liang Ge branch, as shown in the figure.
In first branch, each Feature Mapping is fed to the full articulamentum of two connections, followed by two parallel
Full articulamentum.One of them is returned for bounding box, and exports four values, is the side of each candidate region of n object class
Boundary frame position carries out coding adjusting, another classifies for frame, exports the probability Estimation of N+1 (1 is background) a classification.Classification damage
Lose LclsIt is the logarithm loss of N+1 class, returns loss LregBe by RPN generate bounding box and
The smooth L1 of GT frame loses.In this experiment, target category only has Lung neoplasm, so N is 1.
Second branch is parallel with the first branch, is made of RoIAlign and small-sized FCN, the FCN is only by several heaps
Folded convolutional layer and a transfer convolution composition.Due to the corresponding relationship of pixel to pixel, FCN can carry out semantic segmentation instruction
Practice.It should be noted that the input due to small-sized FCN is made of the Feature Mapping of the RoIAlign of low resolution, and small-sized
The step-length of the transposition convolution of FCN is set as 2 to control the quantity of weight, therefore the corresponding relationship of pixel to pixel is actually spy
The GT mask of a pixel-map to the RoI fixed size of sign (is sized four pixels for 28*28).
The Lung neoplasm binary segmentation mask of each RoI in mask branch prediction RPN, for each mask, small-sized FCN's
Sigmoid is executed in each pixel of the last layer Feature Mapping:
Wherein ωi,jIndicate the value at Feature Mapping position (i, j), LmaskIndicate all S (ωi,j) average binary pair
Number loss.
Therefore, network is trained by using the multitask loss definition of each sampling RoI:
L=Lcls+Lbox+Lmask (1.4)
Mask-RCNN method is a kind of semantic segmentation technology based on region, can be used for predicting object boundary frame and corresponding
Mask, the region that mask prediction is provided based on RPN is suggested, and mask loss is added in multitask loss, raising area
The accuracy that domain is suggested.Mask prediction and region suggest mutually promoting, and improve the positioning accuracy on Lung neoplasm boundary.
Model adjustment is the minor adjustment to network architecture, such as changes the stride of specific convolutional layer, this adjustment
Usually hardly change computation complexity, but the image that can not ignore may be generated to model accuracy.In the present invention, will make
The effect of network model is promoted with three kinds of adjustment of ResNet and DenseNet.
ResNet framework is presently described, module especially relevant to model adjustment.ResNet network is led by an input
Dry, four follow-up phases and a final output layer composition, as shown in fig. 7, in show convolution kernel, output channel number and stride
Size (default value 1), pond layer are similar.Input trunk has 7 × 7 convolutional layers, and output channel number is 64, stride 2,
The maximum pond layer for being 2 followed by stride.Width will be inputted and highly reduce 2 times by inputting trunk, and its port number is increased to
64。
Since Stage 2, each stage is started with one down-sampled piece (conv block), followed by several residual
Poor block (identity block).In down-sampled piece, there are path A and path B, there are three convolutional layer, core sizes by path A
Respectively 1 × 1,3 × 3 and 1 × 1, the stride of first convolutional layer are 2, and input width and height halve, concluding volume lamination
Output channel number is 4 times bigger than the first two, referred to as bottleneck structure.Path B will input dimension transformation using 1 × 1 convolution that stride is 2
For the output dimension of path A, therefore can sum the output in two paths to obtain the output of down-sampling block.Residual block
The convolution for being 1 using only stride, it is other to be similar to down-sampling block.
It can change the quantity of residual block in each stage, to obtain different ResNet network models, such as ResNet-
50 and ResNet-152, wherein the convolution layer number in digital representation network.
Two popular ResNet adjustment are first recalled, ResNet-B and ResNet-C is referred to as, introduces one later
The new model of kind adjusts ResNet-D.
(1) ResNet-B, improvement part are exactly that the residual block of downsample will be in stage
Downsample operation changes second 3 × 3 convolutional layer into from first 1 × 1 convolutional layer, if downsample operation is placed on
1 × 1 convolutional layer that stride is 2, then will lose compared with multicharacteristic information (default is to be reduced to 1/4), it can be understood as have
3/4 characteristic point is all not engaged in calculating, and downsample operation, which is placed on 3 × 3 convolutional layers then, can reduce this loss,
Even if because stride is set as 2, but convolution kernel size is enough big, therefore can be with almost all of position on Cover Characteristics figure.
(2) ResNet-C, improvement part are exactly by 7 × 7 convolutional layers of the part input stem in Fig. 7 with 3 volume 3 × 3
Lamination replacement.The thought of Inception v2 has been used for reference in this part, and main consideration is calculation amount, after all large scale convolution kernel band
The calculation amount come is many more than small size convolution kernel, if can find 33 × 3 convolutional layers in ResNet-C under carefully calculating
Calculation amount lack unlike original.
(3) ResNet-D, improvement part be by stage part be downsample residual block branch from
1 × 1 convolutional layer that stride is 2 changes the convolutional layer that stride is 1 into, and adds a pond layer in front and be used to do
downsample.Although pond layer can also lose information, at least lost again after selecting (for example being averaging operation here)
Redundancy is lost, it is much better compared to 1 × 1 convolutional layer that stride is set as 2.
Fig. 8 is three kinds of ResNet structure optimizations.ResNet-B has modified the down-sampling block of ResNet, and ResNet-C is further
Have modified input stem, it is most important that, ResNet-D has modified down-sampling block again.
Assuming that input is a picture X0, by one L layers of neural network, wherein i-th layer of nonlinear transformation is denoted as
Hi(*), Hi(*) can be cumulative such as BN, ReLU, Pooling or Conv of many kinds of function operation.I-th layer of feature exports note
Make Xi。
In order to advanced optimize the propagation of information flow, DenseNet proposes the network structure of diagram.
As shown, i-th layer of input is not only related to i-1 layers of output, there are also the related notes of the output of all front layers
Make:
Xl=Hl([X0,X1,...,Xl-1]) (1.5)
Wherein [] represents concatenation (splicing), both by X0To Xl-1All output feature map of layer are pressed
Channel is combined.Here used nonlinear transformation H is the combination of BN+ReLU+Conv (3 × 3).
Due to needing the feature map to different layers to carry out cat operation in DenseNet, so needing different layers
Feature map keeps identical feature size, and which limits the realizations of Down sampling in network.In order to make
With Down sampling, author divides DenseNet for multiple Denseblock.
Feature size is required to keep same size in the same Denseblock, between different Denseblock
Transition layers is set and realizes Down sampling, transition layer is by BN+ in the experiment of author
Conv (1 × 1)+2 × 2average-pooling composition.
In Denseblock, it is assumed that the output of each nonlinear transformation H is K feature map, then i-th layer
The input of network is just K0+ (i-1) × K, here we can see that a main difference of DenseNet and existing network
Point: DenseNet can receive output of the less characteristic pattern quantity as network layer.
Reason is exactly that each layer in the same Denseblock is all associated with all layers before, if handle
Feature regards the global state of a Denseblock as, then each layer of training objective is by the existing overall situation
State, judgement need to be added to global state updated value thus the characteristic pattern quantity K of each network layer output be also known as
Growth rate, equally decide each layer need how much of the information updated to global state after us it will be seen that
Lesser K is only needed just to be enough to realize the performance of state-of-art in the experiment of author.
Although DenseNet receives less k, that is, the quantity of feature map is used as output, due to different layers
Between feature map together by cat operative combination, still can be finally feature map channel it is larger and become
The burden of network.Author is used herein 1 × 1Conv (Bottleneck) as the method for Feature Dimension Reduction to reduce channel
Quantity, to improve computational efficiency.Nonlinear transformation after improving becomes BN-ReLU-Conv (1 × 1)-BN-ReLU-Conv
(3 × 3), the DenseNet using Bottleneck layers are known as DenseNet-B by author.In an experiment, author uses 1
× 1 convolution generates the feature map that channel quantity is 4k.
After being extracted by corpse area pixel, the corpse burning time anticipation based on full link neural network is needed to difference
Size corpse carries out forecast analysis to Cremation Machine burning time.Combustion System program differentiates that result burns to Cremation Machine by this
Time is adjusted, so that Cremation Machine burning time parameter setting is more reasonable.No matter subjective assessment or objectively evaluate
An intuitive score is required to embody Cremation Machine burning time.In the present embodiment, by four layers of neural network to cremate
Machine burning time is estimated.The input of neural network is the pixel quantity after human region identification.Network is connected entirely using four layers
Network is connect, network structure is as shown in Figure 9.
Input is 1 node, is that two layers of the hidden layer of element of pixel shared by human body uses 32 units, exports as burning time
Classification, Web vector graphic relu activation primitive, output layer use sigmod activation primitive, learning rate 0.001, and loss function uses
Cross entropy loss function.
It should be noted that the description of the above specific embodiment is not intended to limit the invention, it is all in essence of the invention
Any modification, equivalent replacement, improvement and so within mind and principle, should all be included in the protection scope of the present invention.
Claims (8)
1. the Cremation Machine burning time calculation method of view-based access control model, it is characterised in that comprise the steps of:
Step 1: the position of corpse is fixed;
Step 2: photo is extracted;
Step 3: the region where corpse is determined by Mask-RCNN method;
Step 4: burning time is determined according to the time relationship of corpse pixel and corpse combustion state.
2. the Cremation Machine burning time calculation method of view-based access control model according to claim 1, it is characterised in that the Mask-
RCNN method includes three phases: feature extraction, region suggest and prediction.
3. the Cremation Machine burning time calculation method of view-based access control model according to claim 2, which is characterized in that mentioned in feature
The stage is taken, using the FPN based on ResNet as the core network of Mask-RCNN, fine base is designed according to Lung neoplasm image
Eigen extract layer and the subsequent characteristics extract layer simplified.
4. the Cremation Machine burning time calculation method of view-based access control model according to claim 3, which is characterized in that the FPN
Include three parts: path from bottom to top, from top to bottom path and lateral connection.
5. the Cremation Machine burning time calculation method of view-based access control model according to claim 3, which is characterized in that built in region
Network RPN is suggested by carrying out sliding window operation on more size characteristic figures that FPN is provided to extract rectangle in view stage, region
Candidate region.
6. the Cremation Machine burning time calculation method of view-based access control model according to claim 2, which is characterized in that in prediction rank
Feature Conversion in the frame of candidate region is the small characteristic pattern with fixed space dimension H*W using the method for RoIAlign by section,
To solve region mismatch problem caused by quantifying twice in RoIPool operation.
7. the Cremation Machine burning time calculation method of view-based access control model according to claim 6, it is characterised in that the small spy
Tubercle is identified as 7*7, nodule segmentation 14*14 in sign figure.
8. the Cremation Machine burning time calculation method of view-based access control model according to claim 6, it is characterised in that it is described twice
It is quantified as the quantization of the boundary RoI and elementary boundary quantization.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910585276.3A CN110378880A (en) | 2019-07-01 | 2019-07-01 | The Cremation Machine burning time calculation method of view-based access control model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910585276.3A CN110378880A (en) | 2019-07-01 | 2019-07-01 | The Cremation Machine burning time calculation method of view-based access control model |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110378880A true CN110378880A (en) | 2019-10-25 |
Family
ID=68251476
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910585276.3A Pending CN110378880A (en) | 2019-07-01 | 2019-07-01 | The Cremation Machine burning time calculation method of view-based access control model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110378880A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111666988A (en) * | 2020-05-22 | 2020-09-15 | 哈尔滨理工大学 | Target detection algorithm based on multi-layer information fusion |
CN111914947A (en) * | 2020-08-20 | 2020-11-10 | 华侨大学 | Image instance segmentation method, device and equipment based on feature fusion and storage medium |
-
2019
- 2019-07-01 CN CN201910585276.3A patent/CN110378880A/en active Pending
Non-Patent Citations (2)
Title |
---|
STONE: "《令人拍案称奇的Mask RCNN》", 《HTTP://ZHUANLAN.ZHIHU.COM/P/37998710》 * |
程旭东 等: "《基于红外热像的热塑性材料二维火蔓延速率》", 《燃烧科学与技术》 * |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111666988A (en) * | 2020-05-22 | 2020-09-15 | 哈尔滨理工大学 | Target detection algorithm based on multi-layer information fusion |
CN111914947A (en) * | 2020-08-20 | 2020-11-10 | 华侨大学 | Image instance segmentation method, device and equipment based on feature fusion and storage medium |
CN111914947B (en) * | 2020-08-20 | 2024-04-16 | 华侨大学 | Image instance segmentation method, device, equipment and storage medium based on feature fusion |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Thuan | Evolution of Yolo algorithm and Yolov5: The State-of-the-Art object detention algorithm | |
KR102102161B1 (en) | Method, apparatus and computer program for extracting representative feature of object in image | |
CN110428428B (en) | Image semantic segmentation method, electronic equipment and readable storage medium | |
CN110210551A (en) | A kind of visual target tracking method based on adaptive main body sensitivity | |
WO2022252272A1 (en) | Transfer learning-based method for improved vgg16 network pig identity recognition | |
CN109584248A (en) | Infrared surface object instance dividing method based on Fusion Features and dense connection network | |
CN108549893A (en) | A kind of end-to-end recognition methods of the scene text of arbitrary shape | |
Islalm et al. | Recognition bangla sign language using convolutional neural network | |
CN107977932A (en) | It is a kind of based on can differentiate attribute constraint generation confrontation network face image super-resolution reconstruction method | |
CN108764281A (en) | A kind of image classification method learning across task depth network based on semi-supervised step certainly | |
CN108985217A (en) | A kind of traffic sign recognition method and system based on deep space network | |
CN106682569A (en) | Fast traffic signboard recognition method based on convolution neural network | |
CN108416266A (en) | A kind of video behavior method for quickly identifying extracting moving target using light stream | |
CN110070107A (en) | Object identification method and device | |
CN109977955A (en) | A kind of precancerous lesions of uterine cervix knowledge method for distinguishing based on deep learning | |
CN114758288A (en) | Power distribution network engineering safety control detection method and device | |
CN108537121A (en) | The adaptive remote sensing scene classification method of environment parament and image information fusion | |
CN112950780B (en) | Intelligent network map generation method and system based on remote sensing image | |
CN108230330A (en) | A kind of quick express highway pavement segmentation and the method for Camera Positioning | |
CN113140019A (en) | Method for generating text-generated image of confrontation network based on fusion compensation | |
CN113159067A (en) | Fine-grained image identification method and device based on multi-grained local feature soft association aggregation | |
CN110378880A (en) | The Cremation Machine burning time calculation method of view-based access control model | |
CN115222998A (en) | Image classification method | |
CN110188621B (en) | Three-dimensional facial expression recognition method based on SSF-IL-CNN | |
Manzari et al. | A robust network for embedded traffic sign recognition |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191025 |
|
RJ01 | Rejection of invention patent application after publication |