CN107423745A - Fish activity classification method based on neural network - Google Patents
Fish activity classification method based on neural network Download PDFInfo
- Publication number
- CN107423745A CN107423745A CN201710187860.4A CN201710187860A CN107423745A CN 107423745 A CN107423745 A CN 107423745A CN 201710187860 A CN201710187860 A CN 201710187860A CN 107423745 A CN107423745 A CN 107423745A
- Authority
- CN
- China
- Prior art keywords
- mrow
- fish
- layer
- msubsup
- output
- 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
- 238000000034 method Methods 0.000 title claims abstract description 34
- 230000000694 effects Effects 0.000 title claims abstract description 20
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 15
- 241000251468 Actinopterygii Species 0.000 claims abstract description 188
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 claims abstract description 66
- 238000012544 monitoring process Methods 0.000 claims abstract description 40
- 238000012549 training Methods 0.000 claims abstract description 40
- 238000001514 detection method Methods 0.000 claims abstract description 20
- 230000008569 process Effects 0.000 claims abstract description 17
- 230000002159 abnormal effect Effects 0.000 claims abstract description 10
- 231100000419 toxicity Toxicity 0.000 claims abstract description 9
- 230000001988 toxicity Effects 0.000 claims abstract description 9
- 210000002569 neuron Anatomy 0.000 claims description 37
- 238000012360 testing method Methods 0.000 claims description 14
- 238000003062 neural network model Methods 0.000 claims description 13
- 230000007935 neutral effect Effects 0.000 claims description 13
- 238000012545 processing Methods 0.000 claims description 11
- 239000013598 vector Substances 0.000 claims description 11
- 230000006870 function Effects 0.000 claims description 10
- 210000004027 cell Anatomy 0.000 claims description 8
- 238000004364 calculation method Methods 0.000 claims description 6
- 239000011159 matrix material Substances 0.000 claims description 6
- 230000004083 survival effect Effects 0.000 claims description 6
- 230000009467 reduction Effects 0.000 claims description 4
- 239000000463 material Substances 0.000 claims description 3
- 210000005036 nerve Anatomy 0.000 claims description 3
- 230000001473 noxious effect Effects 0.000 claims description 3
- 231100001234 toxic pollutant Toxicity 0.000 claims description 3
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 claims description 2
- 230000009471 action Effects 0.000 claims description 2
- 230000004913 activation Effects 0.000 claims description 2
- 210000002364 input neuron Anatomy 0.000 claims description 2
- 230000004807 localization Effects 0.000 claims description 2
- 238000005259 measurement Methods 0.000 claims description 2
- 230000001105 regulatory effect Effects 0.000 claims description 2
- 230000003252 repetitive effect Effects 0.000 claims description 2
- 238000010998 test method Methods 0.000 claims description 2
- 241000208340 Araliaceae Species 0.000 claims 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 claims 1
- 235000003140 Panax quinquefolius Nutrition 0.000 claims 1
- 235000008434 ginseng Nutrition 0.000 claims 1
- 239000010410 layer Substances 0.000 description 57
- 239000003905 agrochemical Substances 0.000 description 12
- 230000006399 behavior Effects 0.000 description 12
- 238000004458 analytical method Methods 0.000 description 11
- 239000003814 drug Substances 0.000 description 10
- 230000008859 change Effects 0.000 description 6
- 238000010586 diagram Methods 0.000 description 6
- 238000000605 extraction Methods 0.000 description 6
- 229940079593 drug Drugs 0.000 description 5
- 238000013461 design Methods 0.000 description 4
- 238000007705 chemical test Methods 0.000 description 3
- 238000002474 experimental method Methods 0.000 description 3
- 230000003993 interaction Effects 0.000 description 3
- 230000033001 locomotion Effects 0.000 description 3
- 231100000614 poison Toxicity 0.000 description 3
- 239000002574 poison Substances 0.000 description 3
- IJGRMHOSHXDMSA-UHFFFAOYSA-N Atomic nitrogen Chemical compound N#N IJGRMHOSHXDMSA-UHFFFAOYSA-N 0.000 description 2
- 230000005856 abnormality Effects 0.000 description 2
- 230000007812 deficiency Effects 0.000 description 2
- 238000009313 farming Methods 0.000 description 2
- 230000004044 response Effects 0.000 description 2
- 210000000697 sensory organ Anatomy 0.000 description 2
- 238000003911 water pollution Methods 0.000 description 2
- 206010000117 Abnormal behaviour Diseases 0.000 description 1
- 241001269238 Data Species 0.000 description 1
- OAICVXFJPJFONN-UHFFFAOYSA-N Phosphorus Chemical compound [P] OAICVXFJPJFONN-UHFFFAOYSA-N 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 231100000481 chemical toxicant Toxicity 0.000 description 1
- 238000010835 comparative analysis Methods 0.000 description 1
- MCWXGJITAZMZEV-UHFFFAOYSA-N dimethoate Chemical compound CNC(=O)CSP(=S)(OC)OC MCWXGJITAZMZEV-UHFFFAOYSA-N 0.000 description 1
- 230000019637 foraging behavior Effects 0.000 description 1
- 239000002917 insecticide Substances 0.000 description 1
- 239000011229 interlayer Substances 0.000 description 1
- 230000002045 lasting effect Effects 0.000 description 1
- 230000035800 maturation Effects 0.000 description 1
- 229910052757 nitrogen Inorganic materials 0.000 description 1
- 238000011017 operating method Methods 0.000 description 1
- 229910052698 phosphorus Inorganic materials 0.000 description 1
- 239000011574 phosphorus Substances 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 238000013179 statistical model Methods 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
- 231100000167 toxic agent Toxicity 0.000 description 1
- 239000003440 toxic substance Substances 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/05—Underwater scenes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Multimedia (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Biophysics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Biomedical Technology (AREA)
- Evolutionary Biology (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
A fish activity detection method based on a neural network comprises the following steps: 1) monitoring crucian in real time as a biological monitoring object; 2) extracting the outline of a target fish by a background difference method, and monitoring a crucian group in real time to obtain a tracking video sequence of the crucian; 3) automatically obtaining the dead and alive texture characteristic information of the monitored fish through a neural network, and the method comprises the following processes: 3.1) collecting and extracting contour information data of the fish target, and using the contour information data as an index for detecting whether the water quality is abnormal; 3.2) training the data based on a neural network algorithm to generate a mature classifier structure model; 3.3) detecting and judging whether the detection model is mature or not by using the new characteristic data; 3.4) detecting real-time water quality data on line through a mature detection model, and finally realizing the on-line detection of water quality toxicity.
Description
Technical field
The present invention relates to the detection of biological formula water quality toxicity, neural network classification technical field, it is proposed that one kind utilizes nerve
Network differentiates the sorting technique of fish activity, and this method can improve the accuracy of biological formula water quality toxicity detection.
Technical background
Normal water quality refers to that the sense organ (transparency, peculiar smell etc.) of water quality and biochemical indicator (phosphorus, nitrogen content, acid-base value etc.) exist
In national standard allowed band.Water pollution often changes the sense organ and biochemical indicator of water quality, by checking the current of water quality
The degree of deviation of data target and normal data index is assured that water quality with the presence or absence of abnormal.Therefore, water quality detection belongs to point
Class problem, i.e., determine that water quality is in normal or abnormality according to organoleptic parameters or biochemical indicator parameter.
Traditional water quality detection method uses the processing method of Physico-chemical tests combination statistical model mostly.Although Physico-chemical tests
Method can accurately provide the achievement data for causing water quality abnormal, but there is also following deficiency:1) water quality data is most
Collection is timed using manual type, causes data that there is hysteresis quality, it is impossible to tackle sudden water pollution;2) due to using
Physico-chemical tests cause testing cost higher, it is necessary to a series of instrument and equipments for analyzing, detecting.
In order to solve that the deficiency of water quality method is detected using physics and chemistry mode, scholars propose the water quality based on Fish behavior
Detection method.Fish behavior is larger to be influenceed by change of water quality, and water quality once pollutes, and lives in Fish behavior therein
Often occur abnormal.By the abnormal behaviour for obtaining and analyzing fish, it is possible to realize to the abnormal real-time early warning of water quality.Generally
Monitored water sample is drained into a monitoring cylinder, monitors in cylinder and places several monitored fishes.Pass through the shooting being placed on monitoring cylinder
Head, automatically obtain the Fish behavior in water sample.Fish behavior in algorithm for design analysis video, so as to realize whether water quality is located
In real-time, the quick and accurate early warning of abnormality.Algorithm therefore, it is possible to accurate judgement Fish behavior is biological formula water quality prison
One of key technique of survey.
If being detected in water sample and toxic compounds or agricultural chemicals composition be present, it will result in and live in Fish behavior therein
It is abnormal, for example trip speed is speeded, hard hit monitoring cylinder, the serious death that can also cause fish in monitoring cylinder.However, in monitoring cylinder
Fish ceaselessly staggeredly travelling and light change both increase judge in monitoring cylinder fish whether the difficulty of death.
The content of the invention
In order to overcome the disadvantages mentioned above of prior art, the present invention proposes a kind of fish Activity determination side based on neutral net
Method, this method can accurately judge to the shoal of fish in monitoring cylinder with the presence or absence of dead fish, so as to be water sample poison that may be present
Property provides real-time early warning signal.
1. a kind of fish activity test method based on neutral net, comprises the following steps:
1) it is monitored in real time using crucian as biological monitoring object, according to realtime graphic of the crucian in cylinder is detected
Feature, judge that it is in dead or existing state with this, so as to realize in monitoring water sample whether by noxious material dirt
Dye carries out real-time early warning.The device that the present invention uses is as shown in figure 1, the current in detection cylinder are realized by water inlet and delivery port
Gentle flowing, not only reduces influence of the current to Fish behavior, also simulates the natural flows environment that fish are lived;
2) characteristics of image of target fish is gathered and extracted, fish objective contour (Fig. 2) is extracted using the method for background difference.
By identifying, splitting the target image taken by monitor video, target fish is tracked and demarcates its real time kinematics position,
Fixed Time Interval is set as measurement period, obtains the characteristics of image of the cycle shoal of fish;
3) state that monitored fish is dead or survives is judged by neutral net.Define dead fish characteristic data set and live fish
Characteristic data set, by importing the dead fish collected and live fish data set, neural network model is trained and judged so as to obtain
The characteristic information of monitored fish death or survival.Detailed process is as follows:
3.1) training data of fish death or existing state is built, as the monitored fish survival state of judgement
According to so as to detect whether water quality occurs abnormal index, such as scheme shown in (3).
By carrying out real-time tracking fish target to the target shoal of fish in video image, mesh in each time interval cycle is extracted
The contour area information of fish individual is marked, then information in target fish region is pre-processed, the area information after processing is preserved
And the training input data (Fig. 4) using this regional image information as training neutral net.The extracting method of training sample data
It is as follows:
A) the fish target arrived according to track and localization, target fish edge contour information is found;
B) according to marginal information, the background information of fish target information and surrounding is intercepted in a manner of minimum enclosed rectangle;
C) thresholding processing is carried out to the information in interception rectangular area, at the background outside target fish profile in region
Manage as invalid information;Retain fish target information not deal with;
D) rectangle picture is subjected to gray processing processing, carries out the training of neutral net as input sample data afterwards.
During the real-time tracking positioning and the extraction of characteristic information of fish target preserve, in order to avoid being deposited in extraction information
In the interference of non-fish target information, especially make c) operating procedure.The purpose is to be invalid information by the processing of non-fish target area.
When so may be such that input data of the input picture as neutral net, effective information only has the information of fish target, so as to reduce
Noise is on influence caused by classifying quality.
Network model is built first, and the weight that network is secondly obtained using the training data of dead fish and live fish as input is joined
Number, so as to obtain the disaggregated model (Fig. 5) of maturation.The model by judging whether there is fish to be in dead state in flow cistern,
So as to make early warning with the presence or absence of toxicity to water sample.
3.2) generation classification structure model is trained to data based on neural network algorithm.
The neutral net reversely passes using a kind of Multi-layered Feedforward Networks trained according to error backward propagation algorithm
Broadcast network (Backpropagation, BP).The topological structure of BP neural network model includes input layer, hidden layer and output layer,
The connection established inside it by certain weight matrix between current layer neuron and next layer of neuron, each layer neuron is only
Mutually it is connected entirely between adjacent layer neuron, with connectionless between neuron in layer, feedback-less connects between each layer neuron,
The feed-forward type nerve network system with hierarchical structure is formed, is exported after sample learning, when reality output and desired output
When not being inconsistent, into the back-propagation phase of error.Error is by output layer, to each connection weight in the way of error gradient declines
Dynamic is made to adjust, and to the successively anti-pass of hidden layer, input layer.The information forward-propagating to go round and begin again and error back propagation mistake
Journey, is the process that each layer weights constantly adjust, and the process of neural network learning training, in continuous repetitive exercise process
In, when the error sum of squares of network is minimum, training terminates and completes the training process of network model.
The basic structure model of BP neural network as shown in fig. 6, the network model number of plies used herein is four layers, wherein
Including an input layer, two hidden layers and an output layer.BP neural network algorithmic procedure is as follows:
A) input layer information is defined by vector x=[x1, x2,…,xm] represent, by the way that samples pictures are adjusted into unified chi
Very little size, and make local binarization processing (LBP) to every width picture and obtain characteristic vector, as network inputs sample data, note
For x.Each xiCharacteristic information represents the number of input layer as an input neuron, m.Wherein to LBP features to
The step of amount extraction, is as follows:
1) Target Photo is divided into 16 × 16 zonule (cell);
2) for a pixel in each cell, by the gray value of 8 adjacent pixels compared with it, if all
Enclose pixel value and be more than center pixel value, then the position of the pixel is marked as 1, is otherwise 0.So, 8 in 3*3 neighborhoods
Point can produce 8 bits through comparing, that is, obtain the LBP values of the window center pixel;
3) each cell histogram, i.e., the frequency that each digital (it is assumed that decimal number LBP values) occurs are calculated;Then
The histogram is normalized.
4) finally obtained each cell statistic histogram is attached as a characteristic vector, that is, view picture
The LBP texture feature vectors of figure;
B) defining network connection weights isIt by weights size is w to have one between layer and layer neuronijConnection
Parameter, and between the neuron of current layer independently of each other.WhereinRepresent i-th neuron of preceding layer and current layer j-th
The connection weight of neuron, p value take 1,2,3, represent respectively input layer and hidden layer, hidden layer and hidden layer, hidden layer and
Weight matrix between output layer.As shown in formula (1).
C) target activation primitive f is defined, as shown in formula (2), for normalizing the output result of hidden layer and output layer,
So that output result is the value between 0 to 1;The output of given current layer is Y, and the output of preceding layer is v, and it is defeated to produce current layer
Go out the output relation expression formula with last layer, see formula (3).
Y=f (W × v) (3)
D) it is n to define hidden layer neuron nodes, and the output of j-th of neuron of l layers is usedRepresent, l layers
The error amount of j-th of neuron is usedRepresent.Using two hidden layers in present network architecture, therefore l values are 1 and 2.Wherein
Shown in the output of first hidden layer such as formula (5), shown in the output such as formula (6) of the second hidden layer.
E) output of the second hidden layer is again by exporting weight matrixThe output layer of a neuron node is merged into, is passed through
Finally give the output of this neuron after successively calculating, what this output valve represented is fish target current time it is dead or
Existing state.
F) for n given training sample, its desired output ynIt is known, then to pass through training net for input sample
The output of network and the value of desired output will produce error e, define shown in error function expression formula such as formula (9)
G) the partial derivative δ of the error function e calculation error function pair output layer neurons obtained by output layero, then profit
Use δoConnection weight is corrected in output with each neuron of preceding layer hidden layerAs shown in formula (10).
Partial derivatives of the following calculation error function e to the second hidden layer neuronRecycleImplied with preceding layer
Connection weight is corrected in the output of each neuron of layerAs shown in formula (11).
Partial derivatives of the last calculation error function e to the first hidden layer neuronUtilize each neuron of the first hidden layer
'sInput with each neuron of input layer is to connection weightAmendment, as shown in formula (12).
Coupled later layer is multiplied by the output that the gradient of each weight is equal to coupled previous node layer
The output of backpropagation.
When error is more than zero to the partial derivatives of weights, weighed value adjusting is negative, and reality output is more than desired output, weights to
Reduce direction adjustment so that reality output and the subtractive of desired output are few.When error it is small to the partial derivative of weights zero when, weights adjust
Whole is that just, reality output is less than desired output, and weights adjust to augment direction so that reality output and the subtractive of desired output
Judge whether network error meets to require.It is more than the maximum secondary of setting when error reaches default precision or learns number
Number, then terminate algorithm.Otherwise, next learning sample and corresponding desired output are chosen, returns c), continues to learn.
3.3) generalization ability of the detection model is judged using new data to detect.It is, the figure of input target fish
As feature, by the neural network model trained, output result between zero and one, i.e., to death or is deposited by its value size
Classification is made in work.It is x, output result y to calculate current input image characteristic vector according to formula (13)0Value.
y0=f (x) (13)
Work as y0Value be less than or equal to 0.1 when, then judge that target fish is in existing state, otherwise in dead state.Pass through
New shoal of fish video is tested the network after training, examine its during whole T time live fish judge into dead fish and
Dead fish is judged into the time t of live fish, if t exceed must threshold value, the grader is not mature enough, it is necessary in regulating networks
Some parameters re-start study, until t is less than the threshold value, what the grader can be ripe is applied to real-time early warning water quality
In.
3.4) early warning and monitoring is carried out to water quality according to the neural network model after test;
Obtained ripe network model is applied into camera to monitor in picture in real time, so as to monitor picture to each moment
The fish targeted activity state of middle extraction carries out accurately classification and judged, pre- in real time so as to be made to water quality toxicity that may be present
It is alert., when time length is to judge the ratio of dead state more than 50% in 30s time windows, one is denoted as when detecting target fish in cylinder
Secondary death.But during the monitoring to fish activity judges, in order to ensure the Stability and veracity of monitoring, it is spaced every two minutes
Take monitoring result to carry out final early warning to judge, determine whether water quality by toxic pollutant and send early warning letter
Number.
The technical advantage of the present invention is special to water sample Mesichthys logo image by the neural network model by repeatedly training
Sign classification in real time, and then quick early warning is made with the presence or absence of toxicity and pollution to water quality.In addition, the neural network model is to fish
The classification of class activity has the characteristics of response time is fast, accuracy is high.Especially, in shoal of fish target can also accurate judgement be
It is no to there are fish and be in dead state, so as to improve the accuracy of water quality toxicity early warning.
Brief description of the drawings
Fig. 1 flows water source fish jar schematic diagram;
The extraction schematic diagram of Fig. 2 objective contours;
Fig. 3 a are fish object delineations under normal water quality, and Fig. 3 b are fish object delineations under exception water quality;
The pretreated fish target signature schematic diagrames of Fig. 4;
Fig. 5 detecting system overall flow schematic diagrams;
Fig. 6 neural network structure schematic diagrames;
Fig. 7 algorithm for training network schematic diagrames.
Fig. 8-1 is learning curve
Fig. 8-2 is 1,3,5 fish target lower network test results
Fig. 8-3 is different agricultural chemicals lower network test results
Fig. 8-4 is 1 fish test result
Test result when Fig. 8-5 is 3 fishes
Test result when Fig. 8-6 is 5 fishes
Fig. 8-7 is the detection response time under different pharmaceutical concentration
Fig. 8-8 is dead time of fire alarming
Embodiment
With reference to experiment accompanying drawing, the invention will be further described.
The present invention carries out classification judgement based on multilayer neural network to the fish target image characteristics in monitoring video frame, detects
Whether whether it is dead, abnormal so as to analyze water quality.For the effect of testing model, different experimental rings is selected in experimentation
Border and the multivariable situation of different monitoring fish destination numbers, are carried out to the situation of normal water matter and the reflection of exception water quality drag
Repeatedly relatively.Simultaneously in order to verify influence degree of the common farming medicine to fish activity, several frequently seen agricultural chemicals medicine is devised
Aquatic environment under thing and different pharmaceutical concentration, examine monitoring situation of the network model in water quality monitoring.
Experimentation obtains the fish body motion video image under normal and exception water quality using digital camera head, per two field picture
Size is 640 × 480, and frame speed is 30f/s.Experimental program is as follows:
1) fish target signature picture sample set is gathered, network structure is trained;
The travelling video of fish target under normal water quality is shot, a fish target is changed within every 30 minutes and re-shoots, to 1,3,5
Fish target individual under several is acquired, each 10 groups of the video of different fish numbers.Fish target is dead under same shooting exception water quality
Video, video duration 30 minutes is each under the fish target conditions of different bar numbers to clap 10 groups.
60 groups of video datas are pre-processed, obtain sample data, it is (normal that set of data samples includes 30 groups of training sample
With each 15 groups under exception water quality), wherein 15 under exception water quality group include adding the video sample in the case of different pharmaceutical.And survey
Different pharmaceutical and concentration under sample sheet 30 groups of sample data sets and exception water quality including different fish individual amounts under normal water quality
Sets of video data.
During training, network model is made according to input and corresponding desired output data to the weight of each interlayer
Adjustment, so that the output of network approaches actual data gradually.Represent to select from training video as shown in Fig. 8-1, in figure
Taking ten groups of videos, analysis network model is during training to the convergence of fish Activity determination as training data.
Fig. 8-1 shows the increase with the training time, and the connection weight in network structure model is by constantly adjusting, most
Gradual convergent process is progressivelyed reach eventually.Additionally, learning rate determines caused weights change in circuit training each time
Amount, therefore certain comparative analysis has been made in the selection to e-learning rate., can when learning rate selection is excessive from Fig. 8-1
Network can be caused finally to be difficult to restrain, the too small locally optimal solution or longer training time of may being absorbed in causes convergence slow
Slowly.By the interpretation of result trained to the network model under different learning rates, carried out in the fish target existing state to detecting
Classification when, learning rate selection network at 0.05 or so is relatively more stable.
2) the bar number of fish and the correctness for changing test model of medicine are passed through;
30 groups of videos outside followed by test sample are tested network model.Target fish is equally first traced into,
Then the target traced into each moment detects, and each individual images feature is as input information.Network now
Each layer weight no longer changes, and allows network model to produce output according to input information.
As shown in Fig. 8-2, measure of merit (each is monitored to the video in the case of 1,3 and 5 live fish respectively
The situation of different bar numbers is respectively repeated 10 times, 30 minutes every time), the target fish then extracted using under current background as input, 1
Bar, 3,5 fishes population sample root-mean-square error without significant difference (F (1,3,5)=1.57, p=0.2476>
0.05)。
The error of every group of sample all fluctuates in a small range in test set, and does not occur bright with the difference of fish bar number
Aobvious change, illustrate that the network model can make effective identification classification in real time to fish target individual.
In order to ensure detection agricultural chemicals to biology impact until death process in, can the network model to fish
Active Accurate classification judges that there is provided the sliding window that time size is 30 for this detection to represent the activated state of current fish, if
The ratio of dead state is judged in window time more than 50%, then current time be judged as death.Video is shot according to camera
Duration has 30 frames in 1 second, and the travelling situation of fish target varies less in this 30 frame time, so as to take a frame every 30 frames
As data are judged, so the result of determination in the case where size is 30 sliding window is exactly just by network to fish mesh in 30 second time
Mark the judgement of survival condition.
In judging target detection, in order to be made to the target fish activity under toxic chemical substance in real time effectively
Detection and classification, respectively to 1,3,5 fish in the case of plus people's agricultural chemicals after there is the dead situation of fish target and detected.Should
Test in shooting video using moving about under first 30 minutes normal water quality, latter 30 minutes are to be separately added into insecticide, poison with poison
Three kinds of concentration of tick and Rogor are all 0.2mg/L common farming medicine, total duration 1 hour, and fish go out in latter 30 minutes
Existing abnormal growth phenomenon.
Fig. 8-3 shows that in the environment of wall scroll fish fish target is by one section after the reduction of medicine action activity and death
During the entire process of time, the model can still identify to its Accurate classification.But in the case of 3 fishes, 5 fishes, fish
When target occurs dead, there is obvious reduction in the accuracy for monitoring judgement.When occurring fish target death in the live fish shoal of fish,
Live fish target can collide during travelling with dead fish target, so that the network model input picture extracted is dead fish
The common information with live fish, and the training during training to dead fish is simple target, other situations are as live fish sample
Classification is trained, and so as to cause that accurately classification judgement can not be carried out to dead fish target, ultimately results in monitoring accuracy reduction.But
When occurring dead to fish, in the case that no live fish is hit, dead fish still can be detected accurately, so that monitoring judgement rate
Still more than 80%.Pass through the variance analysis (F (1,3,5)=6.48, p=0.0124 to 1,3,5 movement pattern of fish conditions<0.05),
It was found that notable difference be present is judged to the dead classification for fish later occur under Different Individual, and (F (agricultural chemicals 1, agricultural chemicals 2, agricultural chemicals
3)=0.34, p=0.7169>0.05) the difference unobvious under different agriculture drug conditions are shown.
Individually analysis is made further in judgement for this to 1,3,5 movement pattern of fish condition lower network model, dead in single goal dosing
During front and rear analysis, by obtaining the video of 30 minutes before and after ten groups of wall scroll fish death under the conditions of different agricultural chemicals, to it
Detection accuracy is counted and makees variance analysis, such as Fig. 8-4.There is fish target death condition in addition agricultural chemicals in the as shown by data
Front and rear, the network model can be to fish activated state Accurate classification.
In the judgement that 3 fishes add before and after fish target death condition occurs in agricultural chemicals, when having fish target death, relative to list
The dead judgement of bar fish, the classification judging nicety rate of fish existing state decrease.When fish freely move about, between fish target
The generation of meeting not timing is staggeredly collided or social foraging behavior, the presence of this behavior will bring in the extraction of fish target to go out
Existing dead fish is with live fish simultaneously as a sample data as input so that misjudgment occurs in the network after training.It is but logical
Cross Fig. 8-5 analysis of experimental data to understand, generation of the average about 5 minutes time interaction phenomena in 30 minutes, that is to say, that
The monitoring result for having about 80% period in dead later 30 minutes of fish target is accurate.So as to draw follow-up
Although the dead accuracy rate in judgement that continues to monitor has declined, on having monitored that it is little that individual death influences.
Similarly, in the judgement that 5 fishes add before and after agricultural chemicals death, the situation in also being tested similar to 3 fishes, it is small to have one
The collision phenomenon of section time occurs.And in the training process to training dataset, the present invention touches without selection dead fish and live fish
Training data of the samples pictures as a certain classification when hitting, ultimately result in dead fish live fish while as a detection classification pair
As when, there is false judgment, as shown in Fig. 8-6.
Being indicated in above-mentioned analysis after death occurs in wall scroll fish, it is still very accurate that follow-up monitoring judges, and a plurality of
In the case of fish, due to there is interaction phenomena, judge from the classification for monitoring to monitor in dead backward lasting a period of time
Error, but judge dead influence less to being continued to monitor in practical application.
Now, fish goal behavior is predicted extremely by the different of drug concentration, and then analyzes drug concentration to monitoring
It is larger whether pre-warning time postpones.10 groups of the Process Design is the solution that concentration is 0.1mg/L and 10 groups of concentration are 0.2mg/
L solution, the analysis that analysis network model changes to fish goal behavior under various concentrations, and then further verify network model
Maturity.Analysis result is as shown in Fig. 8-7.
For in the data histograms of 1,3,5 fish, two, left side post represents drug concentration and made as 0.1mg/L in upper figure
Under, network monitor death time and actual environment Mesichthyes death time.Two, the right post represents drug concentration as 0.2mg/L
Under effect, network monitor death time and actual environment Mesichthyes death time.Judge selection due to the monitoring is that the time is big
Small is 30s sliding windows, poor by the average death time in the data graph discovery monitoring process and actual average death time
Other very little, show quickly to detect it when the network model occurs dead to fish, and the reality of the higher fish of concentration
Death time and the time difference that network detects are smaller.
3) the fish activity monitoring in practical application
Although it is above-mentioned used must sliding window pig's tongue fish activity is made in real time monitoring judge, actually should
The most important accuracy for being ensuring that monitoring in.Then under conditions of sliding window judgement, this experiment is sentenced to monitoring again
Disconnected result employs the interval time interval of 2 minutes and is acquired judgement, as whether making the Rule of judgment of early warning to water quality.
As shown in Fig. 8-8, respectively to monitored for the first time in the case of 1,3 and 5 live fish the death time take statistics it is (every kind of
The lower ten groups of experiments of fish number), 2 minutes intervals of death time and predetermined threshold value of wall scroll fish judge to coincide substantially, to 3,5 fishes
In the case of due to the situation together with the interaction of dead fish and live fish occurs, may result in and judge when being spaced 2 minutes and judging
Do not go out fish death, the delay of certain time finally occur, it can also be seen that its average delay time is in certain mistake in figure
In poor scope.Analyzed by the root-mean-square error of the population sample to 1,3,5 fish, without significant difference (F between them
(1,3,5)=2.99, p=0.081>0.05).
Summary, test result indicates that the model can accurately be classified to fish activated state, so as to water quality
In toxic pollutant that may be present make real-time early warning.When noxious material in water sample be present, Fish behavior has obvious change
Change, even in dead state., can be quickly to whether there is dead fish in water sample by the neural network model of design
Quickly and accurately judged.Especially, when the bar number that a plurality of fish in water sample be present and be in dead state fish is less than in water sample
Total number when, the neural network model of design can also be accurately judged to wherein whether have fish to be in dead state.In a word, if
The neural network model of meter can be quickly and accurately pre- by monitoring toxicity progress of the activated state of fish in water sample to water sample
It is alert.
Claims (1)
1. a kind of fish activity test method based on neutral net, step are as follows:
1) it is monitored in real time using crucian as biological monitoring object, it is special according to realtime graphic of the crucian in cylinder is detected
Sign, judge that it is in dead or existing state with this, so as to realize to whether being polluted in monitoring water sample by noxious material
Carry out real-time early warning;
2) characteristics of image of target fish is gathered and extracted, fish objective contour is extracted using the method for background difference;By identifying,
Split the target image taken by monitor video, target fish is tracked and demarcates its real time kinematics position, setting is fixed
Time interval obtains the characteristics of image of the cycle shoal of fish as measurement period;
3) state that monitored fish is dead or survives is judged by neutral net;Define dead fish characteristic data set and live fish feature
Data set, by importing the dead fish collected and live fish data set, neural network model is trained and judges to be supervised so as to obtain
Survey the characteristic information that fish is dead or survives;Detailed process is as follows:
3.1) training data of fish death or existing state is built, as the foundation for judging monitored fish survival state
So as to detect whether water quality occurs abnormal index;
By carrying out real-time tracking fish target to the target shoal of fish in video image, target fish in each time interval cycle is extracted
The contour area information of individual, then is pre-processed to information in target fish region, by the area information after processing preserve and incite somebody to action
Training input data of this regional image information as training neutral net;The extracting method of training sample data is as follows:
S1) the fish target arrived according to track and localization, target fish edge contour information is found;
S2) according to marginal information, the background information of fish target information and surrounding is intercepted in a manner of minimum enclosed rectangle;
S3 thresholding processing) is carried out to the information in interception rectangular area, by the background process outside target fish profile in region
For invalid information;Retain fish target information not deal with;
S4 rectangle picture) is subjected to gray processing processing, carries out the training of neutral net as input sample data afterwards;
3.2) generation classification structure model is trained to data based on neural network algorithm;
The neutral net is using a kind of Multi-layered Feedforward Networks trained according to error backward propagation algorithm, i.e. back propagation network
Network (Backpropagation, BP);The topological structure of BP neural network model includes input layer, hidden layer and output layer, in it
The connection that portion is established between current layer neuron and next layer of neuron by certain weight matrix, each layer neuron only with phase
Mutually full connection between adjacent bed neuron, with connectionless between neuron in layer, feedback-less connection between each layer neuron, composition
Feed-forward type nerve network system with hierarchical structure, is exported after sample learning, when reality output and desired output are not inconsistent
When, into the back-propagation phase of error;Error is made by output layer in the way of error gradient declines to each connection weight
Dynamic adjusts, and to the successively anti-pass of hidden layer, input layer.The information forward-propagating to go round and begin again and error back propagation process, it is
The process that each layer weights constantly adjust, and the process of neural network learning training, during continuous repetitive exercise, work as net
The error sum of squares of network is minimum, and training terminates and completes the training process of network model;
The network model number of plies of BP neural network is four layers, including an input layer, two hidden layers and an output layer;
BP neural network algorithmic procedure is as follows:
A) input layer information is defined by vector x=[x1,x2,…,xm] represent, it is big by the way that samples pictures are adjusted into uniform sizes
It is small, and make local binarization processing (LBP) to every width picture and obtain characteristic vector, as network inputs sample data, it is designated as x.
Each xiCharacteristic information represents the number of input layer as an input neuron, m.Wherein LBP characteristic vectors are carried
The step of taking is as follows:
A1) Target Photo is divided into 16 × 16 zonule (cell);
A2) for a pixel in each cell, by the gray value of 8 adjacent pixels compared with it, if surrounding
Pixel value is more than center pixel value, then the position of the pixel is marked as 1, is otherwise 0.So, 8 points in 3*3 neighborhoods
8 bits can be produced through comparing, that is, obtain the LBP values of the window center pixel;
A3 each cell histogram, i.e., the frequency that each digital (it is assumed that decimal number LBP values) occurs) are calculated;Then it is right
The histogram is normalized;
A4) finally obtained each cell statistic histogram is attached as a characteristic vector, that is, view picture figure
LBP texture feature vectors;
B) defining network connection weights isIt by weights size is w to have one between layer and layer neuronijConnecting quantity,
And between the neuron of current layer independently of each other.WhereinRepresent i-th of neuron of preceding layer and j-th of neuron of current layer
Connection weight, p value takes 1,2,3, represents input layer and hidden layer, hidden layer and hidden layer, hidden layer and output layer respectively
Between weight matrix, as shown in formula (1);
C) target activation primitive f is defined, as shown in formula (2), for normalizing the output result of hidden layer and output layer so that
Output result is the value between 0 to 1;The output of given current layer be Y, and the output of preceding layer be v, produce current layer export and
The output relation expression formula of last layer, see formula (3).
<mrow>
<mi>f</mi>
<mrow>
<mo>(</mo>
<mi>z</mi>
<mo>)</mo>
</mrow>
<mo>&equiv;</mo>
<mfrac>
<mn>1</mn>
<mrow>
<mn>1</mn>
<mo>+</mo>
<msup>
<mi>e</mi>
<mrow>
<mo>-</mo>
<mi>z</mi>
</mrow>
</msup>
</mrow>
</mfrac>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>2</mn>
<mo>)</mo>
</mrow>
</mrow>
Y=f (W × v) (3)
D) it is n to define hidden layer neuron nodes, and the output of j-th of neuron of l layers is usedRepresent, j-th of l layers
The error amount of neuron is usedRepresent;Using two hidden layers in network structure, therefore l values are 1 and 2;Wherein first is hidden
Shown in output containing layer such as formula (5), shown in the output such as formula (6) of the second hidden layer;
<mrow>
<msubsup>
<mi>v</mi>
<mi>j</mi>
<mn>1</mn>
</msubsup>
<mo>=</mo>
<mi>f</mi>
<mrow>
<mo>(</mo>
<msubsup>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>n</mi>
</msubsup>
<msubsup>
<mi>w</mi>
<mrow>
<mi>i</mi>
<mi>j</mi>
</mrow>
<mn>1</mn>
</msubsup>
<msub>
<mi>x</mi>
<mi>i</mi>
</msub>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>5</mn>
<mo>)</mo>
</mrow>
</mrow>
<mrow>
<msubsup>
<mi>v</mi>
<mi>j</mi>
<mn>2</mn>
</msubsup>
<mo>=</mo>
<mi>f</mi>
<mrow>
<mo>(</mo>
<msubsup>
<mi>&Sigma;</mi>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>n</mi>
</msubsup>
<msubsup>
<mi>w</mi>
<mrow>
<mi>i</mi>
<mi>j</mi>
</mrow>
<mn>2</mn>
</msubsup>
<msubsup>
<mi>v</mi>
<mi>i</mi>
<mn>1</mn>
</msubsup>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>6</mn>
<mo>)</mo>
</mrow>
</mrow>
E) output of the second hidden layer is again by exporting weight matrixThe output layer of a neuron node is merged into, by successively
The output for finally giving this neuron is calculated, what this output valve represented is fish target in current time death or survival shape
State;
F) for n given training sample, its desired output ynIt is known, then for input sample by training network
The value of output and desired output will produce error e, define shown in error function expression formula such as formula (9)
<mrow>
<mi>R</mi>
<mo>=</mo>
<msub>
<mfenced open = "[" close = "]">
<mtable>
<mtr>
<mtd>
<msub>
<mi>r</mi>
<mn>1</mn>
</msub>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<msub>
<mi>r</mi>
<mi>n</mi>
</msub>
</mtd>
</mtr>
</mtable>
</mfenced>
<mrow>
<mi>O</mi>
<mi>u</mi>
<mi>t</mi>
<mi>p</mi>
<mi>u</mi>
<mi>t</mi>
<mi>l</mi>
<mi>a</mi>
<mi>y</mi>
<mi>e</mi>
<mi>r</mi>
</mrow>
</msub>
<mo>-</mo>
<msub>
<mfenced open = "[" close = "]">
<mtable>
<mtr>
<mtd>
<msub>
<mi>y</mi>
<mn>1</mn>
</msub>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<msub>
<mi>y</mi>
<mi>n</mi>
</msub>
</mtd>
</mtr>
</mtable>
</mfenced>
<mrow>
<mi>r</mi>
<mi>e</mi>
<mi>a</mi>
<mi>l</mi>
<mi>l</mi>
<mi>a</mi>
<mi>b</mi>
<mi>e</mi>
<mi>l</mi>
</mrow>
</msub>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>8</mn>
<mo>)</mo>
</mrow>
</mrow>
<mrow>
<mi>e</mi>
<mo>=</mo>
<mfrac>
<mn>1</mn>
<mn>2</mn>
</mfrac>
<mo>|</mo>
<mo>|</mo>
<mi>R</mi>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>9</mn>
<mo>)</mo>
</mrow>
</mrow>
G) the partial derivative δ of the error function e calculation error function pair output layer neurons obtained by output layero, recycle δoWith
Connection weight is corrected in the output of each neuron of preceding layer hidden layerAs shown in formula (10);
<mrow>
<msubsup>
<mi>&Delta;w</mi>
<mrow>
<mi>i</mi>
<mi>j</mi>
</mrow>
<mn>3</mn>
</msubsup>
<mo>=</mo>
<mo>-</mo>
<mi>&mu;</mi>
<mfrac>
<mrow>
<mo>&part;</mo>
<mi>e</mi>
</mrow>
<msubsup>
<mi>w</mi>
<mrow>
<mi>i</mi>
<mi>j</mi>
</mrow>
<mn>3</mn>
</msubsup>
</mfrac>
<mo>=</mo>
<msub>
<mi>&mu;&delta;</mi>
<mi>o</mi>
</msub>
<msubsup>
<mi>v</mi>
<mrow>
<mi>i</mi>
<mi>j</mi>
</mrow>
<mn>2</mn>
</msubsup>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>10</mn>
<mo>)</mo>
</mrow>
</mrow>
Partial derivatives of the following calculation error function e to the second hidden layer neuronRecycleWith each god of preceding layer hidden layer
Connection weight is corrected through first outputAs shown in formula (11);
<mrow>
<msubsup>
<mi>&Delta;w</mi>
<mrow>
<mi>i</mi>
<mi>j</mi>
</mrow>
<mn>2</mn>
</msubsup>
<mo>=</mo>
<mo>-</mo>
<mi>&mu;</mi>
<mfrac>
<mrow>
<mo>&part;</mo>
<mi>e</mi>
</mrow>
<msubsup>
<mi>w</mi>
<mrow>
<mi>i</mi>
<mi>j</mi>
</mrow>
<mn>2</mn>
</msubsup>
</mfrac>
<mo>=</mo>
<msubsup>
<mi>&mu;&delta;</mi>
<mi>j</mi>
<mn>2</mn>
</msubsup>
<msubsup>
<mi>v</mi>
<mrow>
<mi>i</mi>
<mi>j</mi>
</mrow>
<mn>1</mn>
</msubsup>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>11</mn>
<mo>)</mo>
</mrow>
</mrow>
Partial derivatives of the last calculation error function e to the first hidden layer neuronUtilize each neuron of the first hidden layerWith
The input of each neuron of input layer is to connection weightAmendment, as shown in formula (12);
<mrow>
<msubsup>
<mi>&Delta;w</mi>
<mrow>
<mi>i</mi>
<mi>j</mi>
</mrow>
<mn>1</mn>
</msubsup>
<mo>=</mo>
<mo>-</mo>
<mi>&mu;</mi>
<mfrac>
<mrow>
<mo>&part;</mo>
<mi>e</mi>
</mrow>
<msubsup>
<mi>w</mi>
<mrow>
<mi>i</mi>
<mi>j</mi>
</mrow>
<mn>1</mn>
</msubsup>
</mfrac>
<mo>=</mo>
<msubsup>
<mi>&mu;&delta;</mi>
<mi>j</mi>
<mn>1</mn>
</msubsup>
<msub>
<mi>x</mi>
<mi>j</mi>
</msub>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>12</mn>
<mo>)</mo>
</mrow>
</mrow>
The reverse of coupled later layer is multiplied by the output that the gradient of each weight is equal to coupled previous node layer
The output of propagation;
When error is more than zero to the partial derivatives of weights, weighed value adjusting is negative, and reality output is more than desired output, and weights are to reduction
Direction adjusts so that reality output and the subtractive of desired output are few.When error it is small to the partial derivative of weights zero when, weighed value adjusting is
Just, reality output is less than desired output, and weights adjust to augment direction so that reality output and the subtractive of desired output;
Judge whether network error meets to require.When error reach default precision or learn number be more than setting maximum times,
Then terminate algorithm.Otherwise, next learning sample and corresponding desired output are chosen, returns c), continues to learn;
3.3) generalization ability of the detection model is judged using new data to detect;It is, the image of input target fish is special
Sign, by the neural network model trained, output result between zero and one, i.e., is made by its value size to dead or survival
Go out classification;It is x, output result y to calculate current input image characteristic vector according to formula (13)0Value;
y0=f (x) (13)
Work as y0Value be less than or equal to 0.1 when, then judge that target fish is in existing state, otherwise in dead state.Pass through new fish
Group's video is tested the network after training, examines it that live fish is judged into dead fish and dead fish during whole T time
Judge into the time t of live fish, if t exceed must threshold value, the grader it is not mature enough, it is necessary in regulating networks some ginseng
Number re-starts study, until t is less than the threshold value, the grader ripe can be applied in real-time early warning water quality;
3.4) early warning and monitoring is carried out to water quality according to the neural network model after test;
Obtained ripe network model is applied into camera to monitor in picture in real time, carried so as to be monitored to each moment in picture
The fish targeted activity state taken carries out accurately classification and judged, so as to make real-time early warning to water quality toxicity that may be present;When
Target fish is denoted as once dead when time length is to judge the ratio of dead state more than 50% in 30s time windows in detection cylinder
Die;But during the monitoring to fish activity judges, in order to ensure the Stability and veracity of monitoring, interval takes one in every two minutes
Secondary monitoring result carries out final early warning and judged, determines whether water quality by toxic pollutant and send pre-warning signal.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710187860.4A CN107423745A (en) | 2017-03-27 | 2017-03-27 | Fish activity classification method based on neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710187860.4A CN107423745A (en) | 2017-03-27 | 2017-03-27 | Fish activity classification method based on neural network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107423745A true CN107423745A (en) | 2017-12-01 |
Family
ID=60423617
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710187860.4A Pending CN107423745A (en) | 2017-03-27 | 2017-03-27 | Fish activity classification method based on neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107423745A (en) |
Cited By (20)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108132099A (en) * | 2017-12-20 | 2018-06-08 | 佛山市洛克威特科技有限公司 | A kind of fish monitoring system |
CN108363970A (en) * | 2018-02-02 | 2018-08-03 | 鲁东大学 | A kind of recognition methods of fingerling class and system |
CN109543679A (en) * | 2018-11-16 | 2019-03-29 | 南京师范大学 | A kind of dead fish recognition methods and early warning system based on depth convolutional neural networks |
CN110069745A (en) * | 2019-04-12 | 2019-07-30 | 山东省科学院海洋仪器仪表研究所 | A kind of active calculation method of bivalve shellfish biological life |
CN110260884A (en) * | 2019-05-24 | 2019-09-20 | 中国联合网络通信集团有限公司 | A kind of biosurveillance method, terminal and server |
CN110633530A (en) * | 2019-09-18 | 2019-12-31 | 南通大学 | Fishway design method based on computational fluid dynamics and convolutional neural network |
CN110704619A (en) * | 2019-09-24 | 2020-01-17 | 支付宝(杭州)信息技术有限公司 | Text classification method and device and electronic equipment |
CN111597937A (en) * | 2020-05-06 | 2020-08-28 | 北京海益同展信息科技有限公司 | Fish gesture recognition method, device, equipment and storage medium |
CN111868472A (en) * | 2018-03-20 | 2020-10-30 | 吉利海洋科技有限公司 | System and method for extracting statistical samples of multiple moving objects |
CN112204674A (en) * | 2018-03-07 | 2021-01-08 | 维迪控股私人有限公司 | Method for identifying biological material by microscopy |
CN112446872A (en) * | 2020-12-07 | 2021-03-05 | 吉首大学 | Giant salamander abnormal behavior identification method based on image identification |
CN112690823A (en) * | 2020-12-22 | 2021-04-23 | 海南力维科贸有限公司 | Method and system for identifying physiological sounds of lungs |
CN112766317A (en) * | 2020-12-31 | 2021-05-07 | 清华大学 | Neural network weight training method based on memory playback and computer equipment |
CN113038118A (en) * | 2020-12-08 | 2021-06-25 | 重庆邮电大学 | Video QoE prediction system and method based on big data environment |
CN113191222A (en) * | 2021-04-15 | 2021-07-30 | 中国农业大学 | Underwater fish target detection method and device |
CN114463675A (en) * | 2022-01-11 | 2022-05-10 | 北京市农林科学院信息技术研究中心 | Underwater fish group activity intensity identification method and device |
EP3909424A4 (en) * | 2019-01-11 | 2022-10-19 | Fulldepth Co., Ltd. | Fish monitoring system and camera unit |
CN115676935A (en) * | 2023-01-03 | 2023-02-03 | 南京博知源环境科技有限公司 | River channel purification treatment method based on nano water treatment technology |
CN116911640A (en) * | 2023-09-11 | 2023-10-20 | 中国地质大学(北京) | Shale reservoir gas content prediction method based on machine deep learning |
US20230368538A1 (en) * | 2018-05-03 | 2023-11-16 | X Development Llc | Fish measurement station keeping |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20030234218A1 (en) * | 2002-06-21 | 2003-12-25 | H2L Co., Ltd. | System and method for AI controlling waste-water treatment by neural network and back-propagation algorithm |
US20070233397A1 (en) * | 2006-03-20 | 2007-10-04 | Sensis Corporation | System for detection and prediction of water quality events |
CN102559822A (en) * | 2012-01-07 | 2012-07-11 | 中国海洋大学 | Method for monitoring preparation of codfish immunoactive peptide on line |
CN106526112A (en) * | 2016-10-25 | 2017-03-22 | 浙江工业大学 | Water toxicity detection method based on fish activity analysis |
-
2017
- 2017-03-27 CN CN201710187860.4A patent/CN107423745A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20030234218A1 (en) * | 2002-06-21 | 2003-12-25 | H2L Co., Ltd. | System and method for AI controlling waste-water treatment by neural network and back-propagation algorithm |
US20070233397A1 (en) * | 2006-03-20 | 2007-10-04 | Sensis Corporation | System for detection and prediction of water quality events |
CN102559822A (en) * | 2012-01-07 | 2012-07-11 | 中国海洋大学 | Method for monitoring preparation of codfish immunoactive peptide on line |
CN106526112A (en) * | 2016-10-25 | 2017-03-22 | 浙江工业大学 | Water toxicity detection method based on fish activity analysis |
Non-Patent Citations (1)
Title |
---|
孙君顶等: "基于BP神经网络的医学图像分类", 《计算机系统应用》 * |
Cited By (27)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108132099A (en) * | 2017-12-20 | 2018-06-08 | 佛山市洛克威特科技有限公司 | A kind of fish monitoring system |
CN108363970A (en) * | 2018-02-02 | 2018-08-03 | 鲁东大学 | A kind of recognition methods of fingerling class and system |
CN108363970B (en) * | 2018-02-02 | 2021-03-23 | 鲁东大学 | Method and system for identifying fish species |
CN112204674A (en) * | 2018-03-07 | 2021-01-08 | 维迪控股私人有限公司 | Method for identifying biological material by microscopy |
CN111868472A (en) * | 2018-03-20 | 2020-10-30 | 吉利海洋科技有限公司 | System and method for extracting statistical samples of multiple moving objects |
US20230368538A1 (en) * | 2018-05-03 | 2023-11-16 | X Development Llc | Fish measurement station keeping |
CN109543679A (en) * | 2018-11-16 | 2019-03-29 | 南京师范大学 | A kind of dead fish recognition methods and early warning system based on depth convolutional neural networks |
EP3909424A4 (en) * | 2019-01-11 | 2022-10-19 | Fulldepth Co., Ltd. | Fish monitoring system and camera unit |
CN110069745A (en) * | 2019-04-12 | 2019-07-30 | 山东省科学院海洋仪器仪表研究所 | A kind of active calculation method of bivalve shellfish biological life |
CN110069745B (en) * | 2019-04-12 | 2023-03-03 | 山东省科学院海洋仪器仪表研究所 | Method for calculating biological life activity of bivalve shellfish |
CN110260884A (en) * | 2019-05-24 | 2019-09-20 | 中国联合网络通信集团有限公司 | A kind of biosurveillance method, terminal and server |
CN110633530A (en) * | 2019-09-18 | 2019-12-31 | 南通大学 | Fishway design method based on computational fluid dynamics and convolutional neural network |
CN110633530B (en) * | 2019-09-18 | 2023-05-05 | 南通大学 | Fishway design method based on computational fluid dynamics and convolutional neural network |
CN110704619A (en) * | 2019-09-24 | 2020-01-17 | 支付宝(杭州)信息技术有限公司 | Text classification method and device and electronic equipment |
CN111597937A (en) * | 2020-05-06 | 2020-08-28 | 北京海益同展信息科技有限公司 | Fish gesture recognition method, device, equipment and storage medium |
CN111597937B (en) * | 2020-05-06 | 2023-08-08 | 京东科技信息技术有限公司 | Fish gesture recognition method, device, equipment and storage medium |
CN112446872A (en) * | 2020-12-07 | 2021-03-05 | 吉首大学 | Giant salamander abnormal behavior identification method based on image identification |
CN113038118A (en) * | 2020-12-08 | 2021-06-25 | 重庆邮电大学 | Video QoE prediction system and method based on big data environment |
CN112690823A (en) * | 2020-12-22 | 2021-04-23 | 海南力维科贸有限公司 | Method and system for identifying physiological sounds of lungs |
CN112766317B (en) * | 2020-12-31 | 2022-10-04 | 清华大学 | Neural network weight training method based on memory playback and computer equipment |
CN112766317A (en) * | 2020-12-31 | 2021-05-07 | 清华大学 | Neural network weight training method based on memory playback and computer equipment |
CN113191222A (en) * | 2021-04-15 | 2021-07-30 | 中国农业大学 | Underwater fish target detection method and device |
CN113191222B (en) * | 2021-04-15 | 2024-05-03 | 中国农业大学 | Underwater fish target detection method and device |
CN114463675A (en) * | 2022-01-11 | 2022-05-10 | 北京市农林科学院信息技术研究中心 | Underwater fish group activity intensity identification method and device |
CN115676935A (en) * | 2023-01-03 | 2023-02-03 | 南京博知源环境科技有限公司 | River channel purification treatment method based on nano water treatment technology |
CN116911640A (en) * | 2023-09-11 | 2023-10-20 | 中国地质大学(北京) | Shale reservoir gas content prediction method based on machine deep learning |
CN116911640B (en) * | 2023-09-11 | 2023-12-26 | 中国地质大学(北京) | Shale reservoir gas content prediction method based on machine learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107423745A (en) | Fish activity classification method based on neural network | |
Sabzi et al. | A fast and accurate expert system for weed identification in potato crops using metaheuristic algorithms | |
Gandhi et al. | Rice crop yield prediction using artificial neural networks | |
Bhatt et al. | Automatic apple grading model development based on back propagation neural network and machine vision, and its performance evaluation | |
CN106526112B (en) | Water toxicity detection method based on fish activity analysis | |
Lek et al. | Predicting stream nitrogen concentration from watershed features using neural networks | |
CN108509976A (en) | The identification device and method of animal | |
CN109190665A (en) | A kind of general image classification method and device based on semi-supervised generation confrontation network | |
Sabzi et al. | Machine vision system for the automatic segmentation of plants under different lighting conditions | |
Ennadifi et al. | Wheat diseases classification and localization using convolutional neural networks and GradCAM visualization | |
CN109919241A (en) | Hyperspectral unknown class target detection method based on probability model and deep learning | |
CN106568907A (en) | Chinese mitten crab freshness damage-free detection method based on semi-supervised identification projection | |
CN116310519A (en) | Surface defect classification method for semi-supervised deep learning | |
Dinata et al. | Classification of strawberry plant diseases with leaf image using CNN | |
Devi et al. | IoT-deep learning based prediction of amount of pesticides and diseases in fruits | |
Rajasekaran et al. | Turmeric plant diseases detection and classification using artificial intelligence | |
Prasher et al. | Analysis of DenseNet201 with SGD optimizer for diagnosis of multiple rice leaf diseases | |
Ambashtha et al. | Leaf disease detection in crops based on single-hidden layer feed-forward neural network and hierarchal temporary memory | |
Sabancı et al. | Classification of different forest types with machine learning algorithms | |
Abdullah et al. | A characterization of watermelon leaf diseases using Fuzzy Logic | |
Lüy et al. | Investigation of the effect of hectoliter and thousand grain weight on variety identification in wheat using deep learning method | |
Yu | Research progress of crop disease image recognition based on wireless network communication and deep learning | |
CN105675807A (en) | Evaluation method of atrazine residue based on BP neural network | |
Sahu et al. | Deep learning models for crop quality and diseases detection | |
Adke et al. | Instance segmentation to estimate consumption of corn ears by wild animals for GMO preference tests |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
WD01 | Invention patent application deemed withdrawn after publication | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20171201 |