CN109543585A - Underwater optics object detection and recognition method based on convolutional neural networks - Google Patents
Underwater optics object detection and recognition method based on convolutional neural networks Download PDFInfo
- Publication number
- CN109543585A CN109543585A CN201811365290.4A CN201811365290A CN109543585A CN 109543585 A CN109543585 A CN 109543585A CN 201811365290 A CN201811365290 A CN 201811365290A CN 109543585 A CN109543585 A CN 109543585A
- Authority
- CN
- China
- Prior art keywords
- image
- target
- detection
- indicate
- value
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computational Linguistics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The underwater optics object detection and recognition method based on convolutional neural networks that the present invention provides a kind of.Firstly, carrying out linear gradation transformation and the enhancing of histogram equalization processing to underwater acquisition image, influence of the water-bed environment to original image, the robustness and detection accuracy of Enhancement Method are reduced;Then, it using multilayer depth convolutional neural networks, by calculating image multilayer depth characteristic, returns to obtain detection position and the generic of different target in output layer, and be trained network using the training method based on supervised learning.The method of the present invention can obtain classifying quality more better than mark feature by hand, and accuracy of identification is higher.
Description
Technical field
The invention belongs to computer vision, computer assisted image processing technical field, specific design is related to a kind of based on convolution mind
Underwater picture object detection and recognition method through network.
Background technique
With the progress of depth learning technology, the development advanced by leaps and bounds also is showed in object detection field in recent years.
The scholar of New York University in 2013 uses deep learning to carry out target detection earliest, relatively previous target detection obtain it is huge into
Step, Ross Girshick is subsequent to have delivered a series of region methods based on convolutional neural networks feature, and using area first is waited
It selects algorithm to extract comprising possible mesh target area, is then finally used using convolutional network in each extracted region convolution feature
Support vector machines classifies to region.Recent research is more by using multiple dimensioned convolution characteristic pattern to reach better
And speed as a result.But there are problems that two for object detection and recognition in underwater scene: first is that underwater light exists seriously
Decaying and scattering effect, underwater picture is presented that scattering is fuzzy and the states such as color regression, prevents color and textural characteristics from making
For the detection feature of underwater optics target.Second is that underwater actual scene is more complicated, so that traditional target detection based on segmentation
Also it is no longer applicable in.Currently, underwater target detection and the research method of identification are mostly the detection methods based on manual features, mainly may be used
It is divided into two kinds.
One is using vision noticing mechanism to form notable figure to obtain possible target area, active profile is then utilized
Method realizes the segmentation of target in image.Christian et al. is in document " A fully automated method to detect
and segment amanufactured object in an underwater color image.Advanced in
0 (0): Signal Processing For Maritime Applications uses color, brightness, side in 1-10,2010 "
To equal three characteristic informations, notable figure to the end is obtained by merging 44 different characteristic figures.But this method feature calculation mistake
Journey time-consuming is more, and time performance is difficult to improve, while when object color and background are close, it is difficult to detect.Another method
It is Bazeille et al. in document " Color-based underwater object recognition using water
Light attenuation, Intelligent Service Robotices, 5 (2): 109-118,2012 " middle use are based on
The object identification of color characteristic preferably reduces the influence of change of scale, while passing through the compensation of the light of active and priori
Colouring information reduces detection error caused by underwater light attenuation.
These methods have certain limitation, comprising: due to detecting the mostly foundation traditional infrastructures feature such as color, brightness,
Cause detection accuracy lower, it is not high to the classification accuracy of different target, and do not have specific aim, complicated multiplicity can not be suitable for
Underwater environment;Detection process is interfered vulnerable to factors such as light, poor for the robustness of environmental change;Based on unsupervised learning
Method training process it is time-consuming, complicated difficult is to reappear.
Summary of the invention
In order to overcome the existing object detection method based on traditional characteristic detection that cannot accurately carry out underwater optics target inspection
Survey the deficiency with identification, the underwater optics object detection and recognition method based on convolutional neural networks that the present invention provides a kind of.
This method leads to the extremely strong characterization ability of optical target using the underwater picture taken as network inputs using neural network
Cross the depth characteristic for calculating underwater photograph technical optical imagery, output layer return to obtain different target detection position and affiliated class
Not.
A kind of underwater optics object detection and recognition method based on convolutional neural networks, it is characterised in that including following step
It is rapid:
Step 1: using following formula image is acquired to each width under water and carries out piecewise linear gray transformation, converted
Image afterwards:
Wherein, (x, y) is image pixel coordinates, and f indicates to acquire image under initial condition, and g indicates image after greyscale transformation, f
(x, y) indicates that the pixel value in image f at (x, y), g (x, y) indicate that the pixel value in image g at (x, y), [a, b] are image f
Intensity value ranges, [c, d] is the intensity value ranges of image g, and the value range of c and d are [0,255], and c < d, F are initial condition
The number of greyscale levels of lower acquisition image.
Step 2: to image progress histogram equalization processing after each width greyscale transformation, image after being enhanced, and will
Image random division is training dataset and test data set after all enhancings, and the training dataset and test data set are wrapped
The ratio of quantity containing image is 3:1.
Step 3: carrying out flip horizontal for image after enhancing, and image and image after its overturning collectively form after all enhancings
New data set is cut from new data set using random image difference mode, obtains the input that size is 300*300*3
Image matrix data collection;
Step 4: with VGG16 network for initial convolution feature extraction network, the input picture square for respectively obtaining step 3
Each Input matrix in battle array data set is repeated the steps of to network, until obtaining trained when Loss value is lower than 0.1
Network:
Image array: being input to 6 layers of convolution fallout predictor of VGG16 network by step a, and each layer obtains a two dimension output
All output matrixes are added by matrix, obtain image characteristic matrix, and the image characteristic matrix contains the inspection of different target
It surveys frame and each detection block belongs to the Forecasting recognition probability of different classifications, the convolution kernel size of the convolution fallout predictor is 3*3*
3;
Step b: it utilizesCalculate target detection position and mark
The Jaccard coefficient of target actual position in image;
Wherein, ApIndicate the position frame of p-th of target in mark image,Indicate the target in correspondence image eigenmatrix
I-th of detection position frame at position,Indicate ApWithThe attribute number that middle attribute is simultaneously 1,Table
Show ApMiddle attribute value be 1 orThe attribute number that middle attribute value is 1, the attribute refer to all pixels point in image for 1
Pixel value is not 0;P=1 ..., P, i=1 ..., Ji, P are total target number, and Ji is target p in image characteristic matrix
Set the detection position frame number at place;
Step c: it utilizesIt is calculated in target detection position and mark image as objective function
Loss value between target actual position;
Wherein, N is the target detection position number that Jaccdard coefficient is greater than threshold value, and the value range of the threshold value is
[0.4,1], α are weight coefficient, α=1, LconfIt is the confidence loss between detection block and callout box, LlocIt is detection block and mark
Positioning loss between frame, calculation formula are respectively as follows:
Wherein, Pos is the set that all detection blocks are constituted in image characteristic matrix, and M is all callout box in mark image
The set of composition,For identifying whether i-th of detection block and j-th of the callout box of target p match, if it does,
Otherwise, Indicate that detection block i belongs to a different category the Forecasting recognition probability of p.Cx indicate target p all detection blocks with
The x-axis position deviation of callout box central point, cy indicate all detection blocks of target p and the y-axis position deviation of callout box central point,
W is the width of callout box, and h is callout box height, smmothL1Indicate smooth L1 norm,Indicate i-th of detection that value is m
Frame,Indicate that value is j-th of callout box of m.
Step 5: the image that the test data that step 2 obtains is concentrated is input to trained network, network output is
All target positions and its affiliated classification, i.e., final recognition result in the image.
The beneficial effects of the present invention are: due to using multilayer depth convolutional neural networks, by calculating image multilayer depth
Feature can obtain classifying quality more better than mark feature by hand, accuracy of identification is more under the premise of being based on pre-training model
It is high;Due to using the network training mode based on supervised learning, sky more significantly more efficient than original image has been arrived by training study
Between structure feature so that the detection accuracy to target position is higher;Due to having carried out linear gradation in advance to underwater acquisition image
Transformation and the enhancing of histogram equalization are handled, and are reduced influence of the water-bed environment to original image, are enhanced the robustness of method
And detection accuracy.
Detailed description of the invention
Fig. 1 is a kind of underwater optics object detection and recognition method flow diagram based on convolutional neural networks of the present invention
Fig. 2 is the result curve schematic diagram classified using the method for the present invention to variety classes object
Specific embodiment
Present invention will be further explained below with reference to the attached drawings and examples, and the present invention includes but are not limited to following implementations
Example.
As shown in Figure 1, the present invention provides a kind of underwater optics object detection and recognition side based on convolutional neural networks
Method, basic process are as follows:
1, piecewise linear gray transformation
In addition to underwater environment light is dark, exposure and equipment self problem etc., can all be caused present in image imaging process
Underwater acquisition picture contrast reduces, and keeps image resolution unclear.Therefore, segmented linear gray is carried out to underwater acquisition image first
Transformation, the dynamic range of expanded images improve the contrast of original image, and linear transformation formula is as follows:
Wherein, (x, y) is image pixel coordinates, and f indicates to acquire image under initial condition, and g indicates image after greyscale transformation, f
(x, y) indicates that the pixel value in image f at (x, y), g (x, y) indicate that the pixel value in image g at (x, y), [a, b] are image f
Intensity value ranges, [c, d] is the intensity value ranges of image g, and the value range of c and d are [0,255], and c < d, F are initial condition
The number of greyscale levels of lower acquisition image.
2, histogram equalization
Histogram equalization processing is carried out to image after greyscale transformation, image detail can be made apparent, improves vision effect
Fruit.Assuming that r ∈ [0,1] is the Normalized Grey Level value before transformation, Pr(r) be original image non-homogeneous probability density function, T
(r) it is grayscale mapping function, enables s=T (r) indicate transformed gray value, be also normalized to s ∈ [0,1], PsIt (s) is equilibrium
Probability density function after change is obtained by probability theory knowledge:
Carry out histogram equalization, Ps(s)=1 it, can be obtained by formula above:
ds=Pr(r)dr (6)
Formula (6) both sides are integrated, the transforming function transformation function of histogram equalization can be obtained:
In digital picture, if the sum of all pixels of image is n, the frequency r for having k gray level to occurkIt indicates, then histogram
Scheme the discrete formula of equalization are as follows:
skIndicate the final gray value that k-th of gray level obtains after carrying out discretization.
After carrying out histogram equalization processing, image random division after all enhancings is training by image after being enhanced
Data set and test data set, wherein the ratio of the quantity of training dataset and the included image of test data set is 3:1.
3, flip horizontal
Using providing method in open source computer vision library OpenCV, image after enhancing is subjected to flip horizontal, Suo Youzeng
Image and image after its overturning collectively form new data set after strong, are carried out from new data set using random image difference mode
It cuts, obtains the input picture matrix data collection that size is 300*300*3.
4, network training
The present invention is with VGG16 network for initial convolution feature extraction network.VGG16 network structure and its each layer are specifically joined
Number scale is loaded in document " K.Simonyan and A.Zisserman, Very Deep Convolutional Networks for
In Large-Scale Image Recognition, arXiv preprint:arXiv:1409.1556,2014. ".
Each Input matrix that the input picture matrix data that step 3 is obtained respectively is concentrated is repeated to VGG16 network
Following steps are trained network, and trained objective function is Loss value, if the pass of output Loss value and the number of iterations
It is that curve is as restrained without obvious fluctuation, specifically 0.1 can be set by objective function Loss value convergence threshold, alternatively, will be repeatedly
Generation number is set as 20,000 times.
(1) image array is input to 6 layers of convolution fallout predictor of VGG16 network, each layer obtains a two dimension output square
Battle array, all output matrixes are added, image characteristic matrix is obtained, the image characteristic matrix contains the detection of different target
Frame and each detection block belong to the Forecasting recognition probability of different classifications, and the convolution kernel size of the convolution fallout predictor is 3*3*3;
(2) it utilizesIt calculates target detection position and mark is schemed
The Jaccard coefficient of target actual position as in.
Wherein, ApIndicate the position frame of p-th of target in mark image,Indicate the target in correspondence image eigenmatrix
I-th of detection position frame at position,Indicate ApWithThe attribute number that middle attribute is simultaneously 1,Table
Show ApMiddle attribute value be 1 orThe attribute number that middle attribute value is 1, the attribute refer to all pixels point in image for 1
Pixel value is not 0;P=1 ..., P, i=1 ..., Ji, P are total target number, and Ji is target p in image characteristic matrix
Set the detection position frame number at place.
(3) objective function is utilizedIt calculates target detection position and target in mark image is true
Loss value between real position.
Wherein, N is the target detection position number that Jaccdard coefficient is greater than threshold value, and the value range of the threshold value is
[0.4,1], α are weight coefficient, α=1, LconfIt is the confidence loss between detection block and callout box, LlocIt is detection block and mark
Positioning loss between frame, calculation formula are respectively as follows:
Wherein, Pos is the set that all detection blocks are constituted in image characteristic matrix, and M is all callout box in mark image
The set of composition,For identifying whether i-th of detection block and j-th of the callout box of target p match, if it does,
Otherwise, Indicate that detection block i belongs to a different category the Forecasting recognition probability of p.All detection blocks of cx expression target p
With the x-axis position deviation of callout box central point, all detection blocks of cy expression target p and the y-axis position of callout box central point are inclined
Difference, w are the width of callout box, and h is callout box height, smmothL1Indicate smooth L1 norm,Indicate i-th of inspection that value is m
Frame is surveyed,Indicate that value is j-th of callout box of m.
Training process optimizes above-mentioned objective function by the way of based on stochastic gradient descent.Based in the present invention
Experimental result, in order to obtain more stable calculated result and more efficient calculating speed, can will training when the number of iterations be arranged
It is 20000 times, 50% procedural learning rate is set as 0.01 before iteration, and rear 50% iterative process learning rate is set as 0.001, weight
Decaying is set as 0.0005, and every batch of training image 64 is opened.
5, evaluation model effect
The present invention carries out input picture using the feedforward convolutional network based on VGG16, using trained network
Processing, it can obtain confidence level different classes of belonging to target in the detection block set and frame of fixed size.It is adopted in the present invention
Recall rate (Recall) is used to carry out the final training pattern of Comprehensive Assessment to the order of accuarcy of data classification as evaluation index.It is calculated
Formula isWherein TP represents classification correct result, and FP represents classification error result.
The present embodiment is in central processing unitI5-4590 3.2GHz CPU, memory 16G, WINDOWS 10 operation
In system, the emulation of MATLAB software progress is utilized.The data used are public data collection PASCAL VOC2007.Fig. 2 is provided
The method of the present invention is on data set respectively to the classification results schematic diagram of animal, plant and stone class.Horizontal axis indicates amount detection,
Unit is thousand times, and the longitudinal axis is percentage shared by variety classes.Wherein, Fig. 2 (a) is the result curve of animal;Fig. 2 (b) is stone
The result curve of class;Fig. 2 (c) is the result curve of plant.What different zones part successively represented from the bottom up is correctly to examine
Survey (Cor), position error (Loc), similar type error (Sim), other wrong (Oth) or background error (BG) accountings differences
In the case where, some accumulated errors that final result generates are influenced.Wherein, when solid line is that Jacarrd coefficient threshold takes 0.7,
The change curve of recall rate (Recall), when dotted line is that Jacarrd coefficient takes 0.5, the change curve of recall rate.It can be seen that this
Invention can effectively classify a variety of different objects, and most of classification results have high confidence level, variety classes
Recall rate result about between 85-90%, while if by Jacarrd coefficient threshold be set as one it is lower as a result,
Nicety of grading of the invention can obtain certain promotion.
Claims (1)
1. a kind of underwater optics object detection and recognition method based on convolutional neural networks, it is characterised in that including following step
It is rapid:
Step 1: using following formula image is acquired to each width under water and carries out piecewise linear gray transformation, obtained transformed
Image:
Wherein, (x, y) is image pixel coordinates, and f indicates to acquire image under initial condition, and g indicates image after greyscale transformation, f (x, y)
Indicate that the pixel value in image f at (x, y), g (x, y) indicate that the pixel value in image g at (x, y), [a, b] are the ash of image f
Angle value range, [c, d] are the intensity value ranges of image g, and the value range of c and d are [0,255], and c < d, F are to adopt under initial condition
Collect the number of greyscale levels of image;
Step 2: histogram equalization processing, image after being enhanced are carried out to image after each width greyscale transformation, and will be owned
Image random division is training dataset and test data set after enhancing, and the training dataset and test data set include figure
The ratio of the quantity of picture is 3:1;
Step 3: carrying out flip horizontal for image after enhancing, and image and image after its overturning collectively form new after all enhancings
Data set is cut from new data set using random image difference mode, obtains the input picture that size is 300*300*3
Matrix data collection;
Step 4: with VGG16 network for initial convolution feature extraction network, the input picture matrix function for respectively obtaining step 3
According to each Input matrix of concentration to network, repeat the steps of, until obtaining trained network when Loss value is lower than 0.1:
Image array: being input to 6 layers of convolution fallout predictor of VGG16 network by step a, and each layer obtains a two dimension output square
Battle array, all output matrixes are added, image characteristic matrix is obtained, the image characteristic matrix contains the detection of different target
Frame and each detection block belong to the Forecasting recognition probability of different classifications, and the convolution kernel size of the convolution fallout predictor is 3*3*3;
Step b: it utilizesCalculate target detection position and mark image
The Jaccard coefficient of middle target actual position;
Wherein, ApIndicate the position frame of p-th of target in mark image,Indicate the target position in correspondence image eigenmatrix
I-th of detection position frame at place,Indicate ApWithThe attribute number that middle attribute is simultaneously 1,Indicate ApIn
Attribute value be 1 orThe attribute number that middle attribute value is 1, the attribute are 1 pixel value for referring to all pixels point in image
It is not 0;P=1 ..., P, i=1 ..., Ji, P are total target number, and Ji is in image characteristic matrix at the position target p
Detect position frame number;
Step c: it utilizesTarget in target detection position and mark image is calculated as objective function
Loss value between actual position;
Wherein, N is the target detection position number that Jaccdard coefficient is greater than threshold value, the value range of the threshold value be [0.4,
1], α is weight coefficient, α=1, LconfIt is the confidence loss between detection block and callout box, LlocBe detection block and callout box it
Between positioning loss, calculation formula is respectively as follows:
Wherein, Pos is the set that all detection blocks are constituted in image characteristic matrix, and M is that all callout box are constituted in mark image
Set,For identifying whether i-th of detection block and j-th of the callout box of target p match, if it does,Otherwise, Indicate that detection block i belongs to a different category the Forecasting recognition probability of p;Cx indicates all detection blocks and mark of target p
The x-axis position deviation of frame central point, cy indicate all detection blocks of target p and the y-axis position deviation of callout box central point, and w is
The width of callout box, h are callout box height, smmothL1Indicate smooth L1 norm,Indicate that value is i-th of detection block of m,Indicate that value is j-th of callout box of m;
Step 5: the image that the test data that step 2 obtains is concentrated is input to trained network, network output is the figure
All target positions and its affiliated classification as in, i.e., final recognition result.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811365290.4A CN109543585A (en) | 2018-11-16 | 2018-11-16 | Underwater optics object detection and recognition method based on convolutional neural networks |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811365290.4A CN109543585A (en) | 2018-11-16 | 2018-11-16 | Underwater optics object detection and recognition method based on convolutional neural networks |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109543585A true CN109543585A (en) | 2019-03-29 |
Family
ID=65847738
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811365290.4A Pending CN109543585A (en) | 2018-11-16 | 2018-11-16 | Underwater optics object detection and recognition method based on convolutional neural networks |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109543585A (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110245602A (en) * | 2019-06-12 | 2019-09-17 | 哈尔滨工程大学 | A kind of underwater quiet target identification method based on depth convolution feature |
CN110706291A (en) * | 2019-09-26 | 2020-01-17 | 哈尔滨工程大学 | Visual measurement method suitable for three-dimensional trajectory of moving object in pool experiment |
CN111445496A (en) * | 2020-02-26 | 2020-07-24 | 沈阳大学 | Underwater image recognition tracking system and method |
CN112597906A (en) * | 2020-12-25 | 2021-04-02 | 杭州电子科技大学 | Underwater target detection method based on degradation prior |
CN112927222A (en) * | 2021-03-29 | 2021-06-08 | 福州大学 | Method for realizing multi-type photovoltaic array hot spot detection based on hybrid improved Faster R-CNN |
CN112926383A (en) * | 2021-01-08 | 2021-06-08 | 浙江大学 | Automatic target identification system based on underwater laser image |
CN114092793A (en) * | 2021-11-12 | 2022-02-25 | 杭州电子科技大学 | End-to-end biological target detection method suitable for complex underwater environment |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103116740A (en) * | 2013-01-24 | 2013-05-22 | 中国科学院声学研究所 | Method and device for identifying underwater targets |
CN107730473A (en) * | 2017-11-03 | 2018-02-23 | 中国矿业大学 | A kind of underground coal mine image processing method based on deep neural network |
CN108564065A (en) * | 2018-04-28 | 2018-09-21 | 广东电网有限责任公司 | A kind of cable tunnel open fire recognition methods based on SSD |
-
2018
- 2018-11-16 CN CN201811365290.4A patent/CN109543585A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103116740A (en) * | 2013-01-24 | 2013-05-22 | 中国科学院声学研究所 | Method and device for identifying underwater targets |
CN107730473A (en) * | 2017-11-03 | 2018-02-23 | 中国矿业大学 | A kind of underground coal mine image processing method based on deep neural network |
CN108564065A (en) * | 2018-04-28 | 2018-09-21 | 广东电网有限责任公司 | A kind of cable tunnel open fire recognition methods based on SSD |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110245602A (en) * | 2019-06-12 | 2019-09-17 | 哈尔滨工程大学 | A kind of underwater quiet target identification method based on depth convolution feature |
CN110706291A (en) * | 2019-09-26 | 2020-01-17 | 哈尔滨工程大学 | Visual measurement method suitable for three-dimensional trajectory of moving object in pool experiment |
CN111445496A (en) * | 2020-02-26 | 2020-07-24 | 沈阳大学 | Underwater image recognition tracking system and method |
CN111445496B (en) * | 2020-02-26 | 2023-06-30 | 沈阳大学 | Underwater image recognition tracking system and method |
CN112597906A (en) * | 2020-12-25 | 2021-04-02 | 杭州电子科技大学 | Underwater target detection method based on degradation prior |
CN112597906B (en) * | 2020-12-25 | 2024-02-02 | 杭州电子科技大学 | Underwater target detection method based on degradation priori |
CN112926383A (en) * | 2021-01-08 | 2021-06-08 | 浙江大学 | Automatic target identification system based on underwater laser image |
CN112927222A (en) * | 2021-03-29 | 2021-06-08 | 福州大学 | Method for realizing multi-type photovoltaic array hot spot detection based on hybrid improved Faster R-CNN |
CN114092793A (en) * | 2021-11-12 | 2022-02-25 | 杭州电子科技大学 | End-to-end biological target detection method suitable for complex underwater environment |
CN114092793B (en) * | 2021-11-12 | 2024-05-17 | 杭州电子科技大学 | End-to-end biological target detection method suitable for complex underwater environment |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109543585A (en) | Underwater optics object detection and recognition method based on convolutional neural networks | |
CN113160192B (en) | Visual sense-based snow pressing vehicle appearance defect detection method and device under complex background | |
CN110472627B (en) | End-to-end SAR image recognition method, device and storage medium | |
CN109584248B (en) | Infrared target instance segmentation method based on feature fusion and dense connection network | |
CN108427920B (en) | Edge-sea defense target detection method based on deep learning | |
EP3614308B1 (en) | Joint deep learning for land cover and land use classification | |
WO2019140767A1 (en) | Recognition system for security check and control method thereof | |
CN103049763B (en) | Context-constraint-based target identification method | |
CN111445488B (en) | Method for automatically identifying and dividing salt body by weak supervision learning | |
Karimpouli et al. | Coal cleat/fracture segmentation using convolutional neural networks | |
CN109684922B (en) | Multi-model finished dish identification method based on convolutional neural network | |
CN110211048B (en) | Complex archive image tilt correction method based on convolutional neural network | |
CN111639587B (en) | Hyperspectral image classification method based on multi-scale spectrum space convolution neural network | |
CN109840483B (en) | Landslide crack detection and identification method and device | |
CN111626993A (en) | Image automatic detection counting method and system based on embedded FEFnet network | |
CN112560675B (en) | Bird visual target detection method combining YOLO and rotation-fusion strategy | |
CN110991257B (en) | Polarized SAR oil spill detection method based on feature fusion and SVM | |
CN112613350A (en) | High-resolution optical remote sensing image airplane target detection method based on deep neural network | |
CN112749621A (en) | Remote sensing image cloud layer detection method based on deep convolutional neural network | |
Yaohua et al. | A SAR oil spill image recognition method based on densenet convolutional neural network | |
Sun et al. | Image recognition technology in texture identification of marine sediment sonar image | |
CN115861409A (en) | Soybean leaf area measuring and calculating method, system, computer equipment and storage medium | |
CN107529647B (en) | Cloud picture cloud amount calculation method based on multilayer unsupervised sparse learning network | |
CN117789037A (en) | Crop growth period prediction method and device | |
CN113469097A (en) | SSD (solid State disk) network-based real-time detection method for water surface floating object multiple cameras |
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 | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190329 |
|
WD01 | Invention patent application deemed withdrawn after publication |