CN110738254A - microscopic image target detection method and system based on depth geometric characteristic spectrum - Google Patents

Abstract

The invention discloses a microscopic image target detection method and system based on a deep geometric feature spectrum. The method includes: firstly constructing a training set, secondly training a funnel network with parameters according to the training set; The microscopic image is input into the funnel network with parameters, and four extreme point heatmaps and center point heatmaps corresponding to the microscopic image to be detected are output; then according to the four extreme point heatmaps and all The center point heat map is used to determine the candidate detection target and the adjacent characteristic spectrum; finally, the detection target is determined according to the candidate detection target and the adjacent characteristic spectrum. The invention adopts the depth geometric feature spectrum for detection, and improves the accuracy of the detection target.

Description

A method and system for microscopic image target detection based on depth geometric feature spectrum

technical field

The invention relates to the technical field of target detection in microscopic image analysis, in particular to a microscopic image target detection method and system based on depth geometric characteristic spectrum.

Background technique

Parasites are cells with various forms, and parasitic diseases are a difficult public health problem faced by most poor countries. Due to the backward medical conditions and technology, the people of these countries have suffered from parasitic diseases for a long time. According to the application value and object characteristics, the present invention selects Toxoplasma, Trypanosoma, Babesia and normal cells as research objects. Most of the microscopic biological analysis relies on the assistance of microscope images, therefore, the microscope images bring convenience for cell detection and statistics. However, the method of manual microscope observation is time-consuming and there are differences in the knowledge of researchers. In addition, the color of the smear, the intensity of light, and the size of the target are quite different in different experimental scenarios. Therefore, the present invention attempts to find a computer-aided method for the microscope. The parasite images below are automatically analyzed.

Deep learning techniques are widely used in the field of object detection, but they are mainly based on candidate regions. Recently, detection methods based on key points have been proposed. There are algorithms based on corner points, center points, extreme points, etc. These methods achieve good results in target detection, but they all ignore the close relationship between key points. , especially in microscopic images with complex and numerous target morphology, the cell geometry is crucial to the detection process.

The existing proposed point-based target detection algorithms cannot be directly transferred to microscopic target detection, and most of them ignore the geometric correlation between each two key points, resulting in the combination of different target points in the key point combination process. Thus, a wrong combination result is formed, which affects the correct result of target detection.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a microscopic image target detection method and system based on the depth geometric characteristic spectrum, which adopts the deep geometric characteristic spectrum for inspection and improves the accuracy of the detection target.

In order to achieve the above objects, the present invention provides a method for detecting objects in microscopic images based on depth geometric feature spectrum, the method comprising:

Step S1: constructing a training set, the training set includes a plurality of labeled microscopic images;

Step S2: obtain a funnel network with parameters according to the training set;

Step S3: Input the microscopic image to be detected into the funnel network with parameters, and output four extreme point heatmaps and center point heatmaps corresponding to the microscopic image to be detected; The heatmaps are up, down, left and right direction heatmaps respectively;

Step S4: Determine the target to be detected and the adjacent feature spectrum according to the four extreme point heatmaps and the center point heatmap;

Step S5: Determine a detection target according to the candidate detection target and the adjacent feature spectrum.

Optionally, the funnel network with parameters obtained by training according to the training set specifically includes:

Step S21: constructing a funnel network based on a fully convolutional network;

Step S22: Select a set number of labeled microscopic images from the training set and input them into the funnel network for training. Under the supervision of local loss and correction loss, use the back-propagation algorithm to update the data in the funnel network. parameter;

Step S23: Determine whether the number of iterations is greater than or equal to the threshold of the number of iterations; if the number of iterations is greater than or equal to the threshold of the number of iterations, output the parameters of the funnel network; if the number of iterations is less than the threshold of the number of iterations, return to step S22.

Optionally, determining the target to be detected and the adjacent feature spectrum according to the four extreme point heatmaps and the center point heatmap specifically includes:

Step S41: extracting all extreme points in each extreme point heat map, and placing them in the set corresponding to each extreme point heat map;

Step S42: Choose an extreme point from each set to form a key point;

Step S43: Calculate the center point according to the key point;

Step S44: determining the pixel value corresponding to the center point on the center point heat map;

Step S45: determine whether the pixel value is greater than or equal to the pixel setting threshold; if the pixel value is less than the pixel setting threshold, remove the key point and return to step S42; if the pixel value is greater than or equal to the pixel setting threshold, then the key point and the center point are used as the detection target to be selected;

Step S46: constructing a weighted graph according to the key points;

Step S47: determining an adjacency matrix according to the weighted graph;

Step S48: Determine the adjacency feature spectrum according to the adjacency matrix.

Optionally, determining the detection target according to the to-be-selected detection target and the adjacent feature spectrum specifically includes:

Step S51: Determine the Euclidean distance according to the adjacent feature spectrum;

Step S52: judge whether the Euclidean distance is within the distance setting threshold; if the Euclidean distance is within the distance setting threshold, then the key point corresponding to the candidate detection target is the detection target; if the Euclidean distance is not within the distance Within the set threshold, the key point corresponding to the candidate detection target is released.

Optionally, the extreme point is a maximum value in an n×n sliding window, where n is the selected window width pixel value.

The present invention also provides a microscopic image target detection system based on the depth geometric characteristic spectrum, the system comprising:

a training set building module for constructing a training set, the training set including a plurality of labeled microscopic images;

a funnel network determination module, used for obtaining a funnel network with parameters according to the training set;

The heat map output module is used to input the microscopic image to be detected into the funnel network with parameters, and output four extreme point heatmaps and center point heatmaps corresponding to the microscopic image to be detected; The extreme point heat map is the heat map in four directions: up, down, left, and right;

a parameter determination module, configured to determine the target to be detected and the adjacent feature spectrum according to the four extreme point heatmaps and the center point heatmap;

A detection target determination module, configured to determine the detection target according to the candidate detection target and the adjacent feature spectrum.

Optionally, the funnel network determination module specifically includes:

The funnel network construction unit is used to build a funnel network based on a fully convolutional network;

A parameter updating unit, for selecting a set number of labeled microscopic images from the training set and inputting them to the funnel network for training, and under the supervision of local loss and correction loss, the back propagation algorithm is used to update the funnel parameters in the network;

The first judgment unit is used to judge whether the number of iterations is greater than or equal to the threshold of the number of iterations; if the number of iterations is greater than or equal to the threshold of the number of iterations, output the parameters of the funnel network; if the number of iterations is less than the threshold of the number of iterations, return "Parameter Update Unit".

Optionally, the parameter determination module specifically includes:

The extraction unit is used to extract all extreme points in each extreme point heat map, and place them in the set corresponding to each extreme point heat map;

The key point determination unit is used to select an extreme point from each set to form a key point;

a center point determination unit, used for calculating the center point according to the key point;

a pixel value determination unit, configured to determine the pixel value corresponding to the center point on the center point heat map;

The second judgment unit is used to judge whether the pixel value is greater than or equal to the pixel setting threshold; if the pixel value is less than the pixel setting threshold, remove the key point and return to the "key point determination unit"; If the pixel value is greater than or equal to the pixel setting threshold, the key point and the center point are used as the detection target to be selected;

a weighted graph construction unit for constructing a weighted graph according to the key points;

an adjacency matrix determining unit for determining an adjacency matrix according to the weighted graph;

An adjacency feature spectrum determining unit, configured to determine an adjacency feature spectrum according to the adjacency matrix.

Optionally, the detection target determination module specifically includes:

an Euclidean distance determining unit for determining the Euclidean distance according to the adjacent feature spectrum;

The third judging unit is used to judge whether the Euclidean distance is within the distance setting threshold; if the Euclidean distance is within the distance setting threshold, the key point corresponding to the candidate detection target is the detection target; if the Euclidean distance is within the distance setting threshold If the Euclidean distance is not within the distance setting threshold, the key point corresponding to the candidate detection target is released.

Optionally, the extreme point is a maximum value in an n×n sliding window, where n is the selected window width pixel value.

According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

The invention discloses a microscopic image target detection method and system based on a deep geometric feature spectrum. The method includes: firstly constructing a training set, secondly training a funnel network with parameters according to the training set; The microscopic image is input into the funnel network with parameters, and four extreme point heatmaps and center point heatmaps corresponding to the microscopic image to be detected are output; then according to the four extreme point heatmaps and all The center point heat map is used to determine the candidate detection target and the adjacent characteristic spectrum; finally, the detection target is determined according to the candidate detection target and the adjacent characteristic spectrum. The invention adopts the depth geometric feature spectrum for detection, and improves the accuracy of the detection target.

Description of drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings required in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some of the present invention. In the embodiments, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without any creative effort.

1 is a flow chart of a method for detecting a target in a microscopic image according to an embodiment of the present invention;

2 is a structural diagram of a microscopic image target detection system according to an embodiment of the present invention;

FIG. 3 is a detection diagram in which the detection target is the nucleus according to the embodiment of the present invention;

Fig. 4 is the detection diagram that the detection target of the embodiment of the present invention is trypanosome;

Fig. 5 is the detection diagram that the detection target of the embodiment of the present invention is Toxoplasma gondii;

FIG. 6 is a detection diagram of Babesia as a detection target according to an embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

The purpose of the present invention is to provide a microscopic image target detection method and system based on the depth geometric feature spectrum, which adopts the depth geometric feature spectrum for detection and improves the accuracy of the detection target.

In order to make the above objects, features and advantages of the present invention more clearly understood, the present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.

FIG. 1 is a flowchart of a method for detecting objects in a microscopic image according to an embodiment of the present invention. As shown in FIG. 1 , the present invention provides a method for detecting objects in a microscopic image based on a depth geometric characteristic spectrum, characterized in that the method includes:

Step S1: constructing a training set, the training set includes a plurality of labeled microscopic images;

Step S2: obtain a funnel network with parameters according to the training set;

Step S3: Input the microscopic image to be detected into the funnel network with parameters, and output four extreme point heatmaps and center point heatmaps corresponding to the microscopic image to be detected; The heatmaps are up, down, left and right direction heatmaps respectively;

Step S4: Determine the target to be detected and the adjacent feature spectrum according to the four extreme point heatmaps and the center point heatmap;

Step S5: Determine a detection target according to the candidate detection target and the adjacent feature spectrum.

Each step is discussed in detail below:

Step S2: Train a funnel network with parameters according to the training set, which specifically includes:

Step S21: constructing a funnel network based on a fully convolutional network;

Step S22: Select a set number of labeled microscopic images from the training set and input them into the funnel network for training. Under the supervision of local loss and correction loss, use the back-propagation algorithm to update the data in the funnel network. parameters; the set number in the embodiment of the present invention is 28, and a total of 4 graphics processing units (GPUs) are used to train the funnel network.

The local loss L _det set by the present invention is to perform weighted point-by-point logistic regression on each heat map. The purpose of weighting is to reduce the false alarm penalty around the true value; the offset correction loss L _off is to improve the accuracy of target detection. proposed to compensate for the resolution loss caused during downsampling.

The formula for the local loss is:

为通过网络计算输出热图中位置(i,j)的预测得分；本实施例中α为2、β为4。Among them, H is the long pixel of the microscopic image, W is the wide pixel of the microscopic image, N is the number of objects in the microscopic image, α and β are hyperparameters respectively, and Y _ij is the value of the microscopic image in the training set. true pixel value,

is the prediction score of the position (i, j) in the output heat map through the network calculation; in this embodiment, α is 2, and β is 4.

分别表示偏移量的实际值和预测值。where N is the number of objects in the microscopic image, SL ₁ is the smoothing loss, _ok and

represent the actual and predicted values of the offset, respectively.

Step S23: Determine whether the number of iterations is greater than or equal to the threshold of the number of iterations; if the number of iterations is greater than or equal to the threshold of the number of iterations, output the parameters of the funnel network; if the number of iterations is less than the threshold of the number of iterations, return to step S22. The threshold for the number of iterations in the embodiment of the present invention is 100,000.

Step S4: determining the target to be detected and the adjacent feature spectrum according to the four extreme point heatmaps and the center point heatmap, specifically including:

Step S41 : extract all extreme points in each extreme point heat map, and place them in the set corresponding to each extreme point heat map; the extreme point is the maximum value in the n×n sliding window, where n is the selected window width pixel value.

Step S42: Choose an extreme point from each set to form a key point K; the specific formula is:

K={(x ^(t) ,y ^(t) ),(x ^(l) ,y ^(l) ),(x ^(b) ,y ^(b) ),(x ^(r) ,y ^(r) ) }, where (x ^(t) , y ^(t) ) is the coordinate position of the top extreme point, (x ^(l) , y ^(l) ) is the coordinate position of the left extreme point, (x ^(b) , y ^(b) ) is the coordinate position of the bottom extreme point, (x ^(r) , y ^(r) ) is the coordinate position of the right extreme point.

Step S43: Calculate the center point according to the key point, and the specific formula is:

Among them, x ^(l) is the abscissa of the left extreme point, x ^(r) is the abscissa of the right extreme point, y ^(t) is the ordinate of the top extreme point, and y ^(b) is the ordinate of the bottom extreme point .

Step S44: determining the pixel value corresponding to the center point on the center point heat map;

Step S45: determine whether the pixel value is greater than or equal to the pixel setting threshold; if the pixel value is less than the pixel setting threshold, remove the key point and return to step S42; if the pixel value is greater than or equal to the pixel setting threshold, then the key point and the center point are used as the detection target to be selected;

Step S46: Construct a weighted graph G=(V, E) according to the key points, wherein V={v ₁ , v ₂ ,...v _n } are nodes in the weighted graph G, E={e ₁ , e ₂ ,...e _m } is the set of edges formed by connecting any two nodes in V.

Step S47: Determine the adjacency matrix A(G) according to the weighted graph, and the specific formula is:

n为节点的总数，权重ω_ij用两点间的欧式距离d(v_i,v_j)表示，即ω_ij＝d(v_i,v_j)＝|v_i-v_j|。in:

n is the total number of nodes, and the weight ω _ij is represented by the Euclidean distance d(v _i ,v _j ) between two points, that is, ω _ij =d(vi ,v _j ) _{=|v i -v j | .}

Step S48: Determine the adjacency feature spectrum according to the adjacency matrix. The characteristic polynomial of the adjacency matrix A(G) is: f(G,λ)=|λE-A|, and finding the eigenvalue λ of f(G,λ) is the adjacency eigenspectrum Spec(A) of the weighted graph G, and its process Converted to find the solution of the characteristic equation |λE-A|x=0. which is

Solve n complex roots λ ₁ ,λ ₂ ,...,λ _n , which are the n eigenvalues of the adjacency matrix, that is, the adjacency eigenspectrum Spec(A)=[λ ₁ ,λ ₂ ,...,λ _n ] .

Step S5: Determine the detection target according to the to-be-selected detection target and the adjacent feature spectrum, specifically including:

Step S51: Determine the Euclidean distance according to the adjacent feature spectrum, and the specific formula is:

D _s = |Spec(A)-Spec(0)|;

Wherein, Spec(A) is the adjacent feature spectrum, and Spec(0) is the initial feature spectrum obtained by training, and the initial feature spectrum is determined according to the calibrated parameters.

Step S52: judge whether the Euclidean distance is within the distance setting threshold; if the Euclidean distance is within the distance setting threshold, then the key point corresponding to the candidate detection target is the detection target; if the Euclidean distance is not within the distance Within the set threshold, the key point corresponding to the candidate detection target is released.

FIG. 2 is a structural diagram of a microscopic image target detection system according to an embodiment of the present invention. As shown in FIG. 2 , the present invention also provides a microscopic image target detection system based on depth geometric feature spectrum, and the system includes:

A training set construction module 1, used for constructing a training set, the training set includes a plurality of labeled microscopic images;

The funnel network determination module 2 is used to obtain the funnel network with parameters according to the training set;

The heat map output module 3 is used for inputting the microscopic image to be detected into the funnel network with parameters, and outputting four extreme point heatmaps and center point heatmaps corresponding to the microscopic image to be detected; the The four extreme point heatmaps are up, down, left, and right direction heatmaps;

A parameter determination module 4, configured to determine the target to be detected and the adjacent feature spectrum according to the four extreme point heatmaps and the center point heatmap;

The detection target determination module 5 is configured to determine the detection target according to the candidate detection target and the adjacent feature spectrum.

Each module is discussed in detail below:

The funnel network determination module 2 specifically includes:

The funnel network construction unit is used to build a funnel network based on a fully convolutional network;

A parameter updating unit, used for selecting a set number of labeled microscopic images from the training set and inputting them to the funnel network for training, and under the supervision of local loss and correction loss, the funnel is updated by back-propagation algorithm parameters in the network;

The first judgment unit is used to judge whether the number of iterations is greater than or equal to the threshold of the number of iterations; if the number of iterations is greater than or equal to the threshold of the number of iterations, output the parameters of the funnel network; if the number of iterations is less than the threshold of the number of iterations, return "Parameter Update Unit".

Parameter determination module 4, which specifically includes:

The extraction unit is used to extract all extreme points in each extreme point heat map, and place them in the set corresponding to each extreme point heat map; the extreme point is the maximum value in the n×n sliding window , where n is the selected window width pixel value.

The key point determination unit is used to select an extreme point from each set to form a key point;

a center point determination unit for calculating the center point according to the key point;

a pixel value determination unit, configured to determine the pixel value corresponding to the center point on the center point heat map;

The second judgment unit is used to judge whether the pixel value is greater than or equal to the pixel setting threshold; if the pixel value is less than the pixel setting threshold, remove the key point and return to the "key point determination unit"; If the pixel value is greater than or equal to the pixel setting threshold, the key point and the center point are used as the detection target to be selected;

a weighted graph construction unit for constructing a weighted graph according to the key points;

an adjacency matrix determining unit for determining an adjacency matrix according to the weighted graph;

An adjacency feature spectrum determining unit, configured to determine an adjacency feature spectrum according to the adjacency matrix.

The detection target determination module 5 specifically includes:

an Euclidean distance determining unit for determining the Euclidean distance according to the adjacent feature spectrum;

The third judging unit is used to judge whether the Euclidean distance is within the distance setting threshold; if the Euclidean distance is within the distance setting threshold, the key point corresponding to the candidate detection target is the detection target; if the Euclidean distance is within the distance setting threshold If the Euclidean distance is not within the distance setting threshold, the key point corresponding to the candidate detection target is released.

The present invention uses the three kinds of parasites and one kind of cell nucleus of Toxoplasma gondii, trypanosome and Babesia collected by Olympus IX53 microscope, there are four data sets in total, namely four training sets. The microscopic images are tested, and the test results are as follows: FIG. 3 is a detection diagram of an embodiment of the present invention in which the detection target is a cell nucleus, FIG. 4 is a detection diagram of a trypanosome as the detection target in an embodiment of the present invention, and FIG. 5 is a detection diagram of an embodiment of the present invention. The detection map of the target is Toxoplasma gondii, and FIG. 6 is the detection map of the detection target of Babesia according to the embodiment of the present invention, and obtained 90.2%, 69.3%, 90.2%, 69.3%, The recognition accuracy is 70.2% and 95.9%, and the average test time is 11.5 seconds, which meets the requirements of automatic recognition of microscopic images.

The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same and similar parts between the various embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant part can be referred to the description of the method.

The principles and implementations of the present invention are described herein using specific examples. The descriptions of the above embodiments are only used to help understand the method and the core idea of the present invention; meanwhile, for those skilled in the art, according to the present invention There will be changes in the specific implementation and application scope. In conclusion, the contents of this specification should not be construed as limiting the present invention.

Claims (10)

1, A microscopic image target detection method based on depth geometric characteristic spectrum, characterized in that the method comprises:

step S1: constructing a training set, wherein the training set comprises a plurality of microscopic images with marks;

step S2: training according to the training set to obtain a funnel network with parameters;

step S3: inputting the microscopic image to be detected into the funnel network with parameters, and outputting four extreme point heat maps and a central point heat map corresponding to the microscopic image to be detected; the four extreme point heat maps are respectively an upper direction heat map, a lower direction heat map, a left direction heat map and a right direction heat map;

step S4: determining a detection target to be selected and an adjacent characteristic spectrum according to the four extreme point heat maps and the central point heat map;

step S5: and determining a detection target according to the detection target to be selected and the adjacent characteristic spectrum.

2. The method for detecting the target of the microscopic image based on the depth geometric feature spectrum of claim 1, wherein the training according to the training set to obtain the funnel network with parameters specifically comprises:

step S21: constructing a funnel network based on the full convolution network;

step S22: selecting a set number of marked microscopic images from the training set, inputting the images into the funnel network for training, and updating parameters in the funnel network by adopting a back propagation algorithm under the supervision of local loss and correction loss;

step S23: judging whether the iteration times are larger than or equal to an iteration time threshold value or not; if the iteration times are larger than or equal to the iteration time threshold value, outputting the parameters of the funnel network; if the number of iterations is less than the iteration number threshold, the process returns to step S22.

3. The method for detecting microscopic image targets based on depth geometric feature spectrum according to claim 1, wherein the determining of the target to be detected and the adjacent feature spectrum according to the four extreme point heat maps and the central point heat map specifically comprises:

step S41: extracting all extreme points in each extreme point heat map, and respectively placing the extreme points in a set corresponding to each extreme point heat map;

step S42, selecting extreme points from each set to form key points;

step S43: calculating a central point according to the key points;

step S44: determining pixel values corresponding to the central points on the central point heat map;

step S45: judging whether the pixel value is larger than or equal to a pixel set threshold value; if the pixel value is smaller than the pixel set threshold, removing the key point, and returning to the step S42; if the pixel value is larger than or equal to a pixel setting threshold value, taking the key point and the central point as a detection target to be selected;

step S46: constructing a weighted graph according to the key points;

step S47: determining an adjacency matrix according to the weighted graph;

step S48: determining a adjacency feature spectrum according to the adjacency matrix.

4. The method for detecting the microscopic image target based on the depth geometric feature spectrum of claim 1, wherein the determining the detection target according to the target to be detected and the adjacent feature spectrum specifically comprises:

step S51: determining Euclidean distance according to the adjacent characteristic spectrum;

step S52: judging whether the Euclidean distance is within a distance setting threshold value; if the Euclidean distance is within a distance setting threshold, the key point corresponds to a detection target to be selected and is taken as a detection target; and if the Euclidean distance is not within the distance setting threshold, removing the key point corresponding to the target to be detected.

5. The method of claim 3, wherein the extreme point is a maximum in an n x n sliding window, where n is a selected window width pixel value.

A system for detecting a target in a microscopic image based on a depth geometric characteristic spectrum, the system comprising:

the training set constructing module is used for constructing a training set, and the training set comprises a plurality of microscopic images with marks;

the funnel network determining module is used for obtaining a funnel network with parameters according to the training set;

the heat map output module is used for inputting the microscopic image to be detected into the funnel network with parameters and outputting four extreme point heat maps and a central point heat map corresponding to the microscopic image to be detected; the four extreme point heat maps are respectively an upper direction heat map, a lower direction heat map, a left direction heat map and a right direction heat map;

the parameter determining module is used for determining a detection target to be selected and an adjacent characteristic spectrum according to the four extreme point heat maps and the central point heat map;

and the detection target determining module is used for determining a detection target according to the detection target to be selected and the adjacent characteristic spectrum.

7. The system for detecting the target of the microscopic image based on the depth geometric feature spectrum of claim 6, wherein the funnel network determination module specifically comprises:

the funnel network construction unit is used for constructing a funnel network based on the full convolution network;

the parameter updating unit is used for selecting a set number of marked microscopic images from the training set, inputting the microscopic images into the funnel network for training, and updating parameters in the funnel network by adopting a back propagation algorithm under the supervision of local loss and correction loss;

and the th judging unit is used for judging whether the iteration times is more than or equal to the iteration time threshold, outputting the parameters of the funnel network if the iteration times is more than or equal to the iteration time threshold, and returning to the parameter updating unit if the iteration times is less than the iteration time threshold.

8. The system for detecting the target of the microscopic image based on the depth geometric feature spectrum of claim 1, wherein the parameter determining module specifically comprises:

the extraction unit is used for extracting all extreme points in each extreme point heat map and respectively placing the extreme points in a set corresponding to each extreme point heat map;

the key point determining unit is used for forming key points by optionally extreme points in each set;

a central point determining unit for calculating a central point according to the key point;

a pixel value determination unit for determining pixel values corresponding to the central points on the central point heat map;

a second judgment unit configured to judge whether the pixel value is greater than or equal to a pixel setting threshold; if the pixel value is smaller than the pixel set threshold, removing the key point, and returning to a key point determining unit; if the pixel value is larger than or equal to a pixel setting threshold value, taking the key point and the central point as a detection target to be selected;

the weighted graph constructing unit is used for constructing a weighted graph according to the key points;

an adjacency matrix determination unit for determining an adjacency matrix according to the weighted graph;

and the adjacent characteristic spectrum determining unit is used for determining an adjacent characteristic spectrum according to the adjacent matrix.

9. The system for detecting the target of the microscopic image based on the depth geometric feature spectrum of claim 6, wherein the detection target determining module specifically comprises:

the Euclidean distance determining unit is used for determining the Euclidean distance according to the adjacent characteristic spectrum;

a third judging unit, configured to judge whether the euclidean distance is within a distance setting threshold; if the Euclidean distance is within a distance setting threshold, the key point corresponds to a detection target to be selected and is taken as a detection target; and if the Euclidean distance is not within the distance setting threshold, removing the key point corresponding to the target to be detected.

10. A depth geometry feature spectrum based microscopy image target detection system according to claim 8 wherein the extreme point is a maximum in an n x n sliding window, where n is a selected window width pixel value.

Images