Search

CN-114863113-B - Method and equipment for discriminating detection result of two-stage new coronaries pneumonia antigen

CN114863113BCN 114863113 BCN114863113 BCN 114863113BCN-114863113-B

Abstract

The invention discloses a method and equipment for discriminating a detection result of a two-stage new coronal pneumonia antigen, which relate to the field of image processing, in particular to application of an image processing technology in the field of prevention and control of a new coronal pneumonia COVID-19 epidemic situation, and the method comprises the following steps that step 1, a plurality of quadrilateral areas recorded with the detection result are obtained from a target picture, and perspective transformation methods are used for carrying out visual angle adjustment on the quadrilateral areas to obtain rectangular areas with corresponding quantity; and 2, identifying the rectangular area by using a classification network based on depth measurement learning to obtain a detection result. The invention uses a depth measurement learning method based on pairs to pair samples, measures the similarity between samples, promotes similar samples to be close to each other, separates samples of different types from each other, and processes fine-grained images from an instance level. And a memory storage module is added into a fine-granularity image retrieval algorithm to perform optimization so as to help improve the recognition accuracy of the tail class samples.

Inventors

  • LIU JIANNAN
  • ZHAI GUANGTAO
  • ZHANG ZHICHAO
  • JIA JUAN

Assignees

  • 上海交通大学医学院附属第九人民医院

Dates

Publication Date
20260512
Application Date
20220609

Claims (8)

  1. 1. The method for distinguishing the detection result of the two-stage new coronatine pneumonia antigen is characterized by comprising the following steps: Step 1, acquiring a plurality of quadrilateral areas recorded with detection results from a target picture, and performing visual angle adjustment on the quadrilateral areas by using a perspective transformation method to obtain a corresponding number of rectangular areas; Step 2, a classification network based on deep measurement learning is used for identifying the rectangular area, and a detection result is obtained; Wherein, the Said step1 comprises the sub-steps of: step 1.1, determining coordinates of four vertexes of the quadrilateral region recorded with the detection result by adopting a trapezoid detector, and marking the coordinates as follows: , wherein, the trapezoid is provided with shape detector use The network is taken as a basic framework, and the network is taken as a basic framework The output format of the network is improved from a rectangular detection frame to a detection frame capable of identifying quadrangles, and the detection frame is provided with a plurality of detection frames The network uses three different-scale grid division feature graphs, wherein 9 predicted values corresponding to each grid in the grid division feature graphs are respectively , wherein, Is the coordinates of four vertices of the quadrilateral region, s representing the probability of the presence of a target in the bounding box of the quadrilateral region; Step 1.2, performing visual angle adjustment on the quadrilateral region by using the perspective transformation method to obtain a corresponding number of rectangular regions, wherein the sitting marks of the rectangular regions are as follows: ; Wherein, the The said The loss function of the network includes a location loss and a confidence loss; The positioning loss is as follows: Wherein, the Dividing the side length of the feature map for the grid, For each of the grids The number of the pieces of the plastic material, Represents the first The first of the grids Personal (S) Predicted fourth of the quadrilateral region Corner points Coordinate information; Is that Corresponding real labels; represents the first The first of the grids Personal (S) Targets are predicted to exist in the quadrilateral region, Represents the first The first of the grids Personal (S) No target is present in the predicted quadrilateral region; The confidence loss is: Wherein, the Dividing the side length of the feature map for the grid, For each of the grids The number of the pieces of the plastic material, Represents the first The first of the grids Personal (S) There is a confidence level of the object in the list, Is that The corresponding real label is used to identify the real label, Represents the first The first of the grids Personal (S) Targets are predicted to exist in the quadrilateral region, Represents the first The first of the grids Personal (S) No object is present in the predicted quadrilateral region, Is a super parameter.
  2. 2. The method for discriminating between two-stage neocoronal pneumonia antigen detection results according to claim 1, wherein said three different scales of said meshing feature map are 20, 40 and 80.
  3. 3. The method for discriminating a detection result of a two-stage new coronal pneumonia antigen according to claim 1, wherein in said step 1.2, a perspective transformation matrix of said perspective transformation method is: 。
  4. 4. The method for discriminating between two-stage neocoronal pneumonia antigen detection results according to claim 1, wherein in said step 2, said depth metric learning based classification network uses a module comprising a memory queue To store a historical sample dataset during training, defined as: Wherein: represents the first A sample queue of the individual classes is presented, For the number of all categories in the historical sample dataset, Wherein: represents the first The first in the queue of the category A number of samples of the sample were taken, Sample queues for each category Capacity size of (a) is provided.
  5. 5. The method of claim 4, wherein the similarity loss function between pairs of samples in the historical sample dataset is defined as: Wherein, the Represents the first Sample and the first Of individual samples The degree of similarity is determined by the degree of similarity, Representative of Middle (f) Sample number Is used to determine the positive samples of the sample, Representative of Middle (f) Sample number Is used to determine the negative samples of the sample, For one our team columns in the historical sample dataset, A capacity size of one our team columns in the historical sample dataset; The similarity loss function between a sample in one our team column in the historical sample dataset and a stored sample in the memory queue is defined as: Wherein, the Representative of Middle (f) A queue of corresponding homogeneous samples in the memory queue, Representative of Middle (f) A corresponding set of negative samples of the samples in the memory queue, The capacity of the queue representing the corresponding similar samples in the memory queue; Wherein, the , , Is a super parameter; similarity between internal pairs of samples The similarity between the samples in the memory queue and the samples in the memory queue is obtained by the following steps of when iteration is performed Less than In the time-course of which the first and second contact surfaces, Optimization of the classification network not participating in the depth metric-based learning when iterating Greater than or equal to In the time-course of which the first and second contact surfaces, And Together co-optimizing the classification network based on depth metric learning; Thus, the first and second substrates are bonded together, The definition is as follows: Wherein alpha control Loss and internal sample And a weight coefficient between the sample loss of the memory queue.
  6. 6. The method for discriminating between two-stage neocoronal pneumonia antigen detection results according to claim 5, wherein said classification loss of said classification network based on deep metric learning is composed of two parts, respectively And The method comprises the following steps: Wherein, the The classification loss is calculated by cross entropy loss for Softmax functions added after the backbone network and FC layer.
  7. 7. An apparatus for discrimination of results of detection of a two-stage neocoronal pneumonia antigen, wherein said apparatus comprises: Processor, and A memory arranged to store computer executable instructions which, when executed, cause the processor to perform operations according to the method of any one of claims 1 to 6.
  8. 8. A computer readable medium storing instructions that, when executed, cause a system to perform operations of the method of any one of claims 1-6.

Description

Method and equipment for discriminating detection result of two-stage new coronaries pneumonia antigen Technical Field The invention relates to the field of image processing, in particular to application of an image processing technology in the field of prevention and control of epidemic situations of new coronaries COVID-19, and specifically relates to a method and equipment for discriminating detection results of two-stage new coronaries antigens. Background Since COVID-19 viruses spread rapidly and there is a difficult to determine latency, virus screening and control efforts become critical. The internet of things system can play a great role in epidemic prevention due to the characteristics of high-efficiency response speed and non-contact operation. In order to cut off the infectious agent rapidly and prevent the epidemic situation from spreading further, COVID-19 nucleic acid detection and antigen detection are effective schemes. The nucleic acid detection method used in the early stage needs to send the collected sample to a professional detection mechanism and carry out a relatively complex detection process, and the detection time often needs several hours. The antigen detection mode which is recently developed can obtain results more quickly, is simple and convenient to operate, and compared with the nucleic acid detection method with complicated flow, the antigen detection time is greatly shortened, the antigen detection can be completed in a few minutes, and the antigen detection can be directly distributed to risk groups to develop self-tests. The antigen detection method is more flexible and practical when dealing with large-scale risk group screening and medical system pressure. The greater pressure of the medical system and the poor statistical timeliness of the detection result make the control of epidemic spread a great difficult problem. In addition, a lot of manpower is input to the detection work of COVID-19 viruses possibly carried, and the risk of virus transmission is further increased. Therefore, in order to solve the problem of statistics of a huge amount of antigen detection results, the epidemic situation transmission is controlled rapidly and accurately, and a set of efficient and contactless automatic antigen detection result recognition system is designed and is imperative. The purpose of antigen test result statistics is to quickly determine whether a certain tester has been infected with virus, so the automatic antigen result recognition system must have two functions, namely 1, associating the detection result with specific personnel identity information, 2, recognizing the imaging result of antigen test paper, and determining whether the personnel has been infected with virus. If the information in both aspects can be automatically identified, the workflow of the IOT system is closed-loop, and no intervention of a third party is needed, thereby reducing the working pressure of medical staff and reducing the risk of virus infection increased due to contact. The identification information of the determined personnel can be directly matched through the bar code or the two-dimensional code on the antigen kit. At present, many two-dimensional code or bar code decoding technologies based on cameras are mature, and personnel identity information can be directly synchronized into the system through camera scanning. However, automatic discrimination of the results of an antigen test is not currently available. In the antigen detection data acquisition stage, the used antigen detection kit is a special article possibly carrying viruses, so that the detection result is collected in a mode that a detected person directly uploads a picture with the detection result. Due to the fact that the shooting habits of testees are different, the quality of the pictures cannot be guaranteed, the pictures can be affected by various influences such as ambient light, shooting angles and pixels of acquisition equipment, and distortion such as blurring, noise and geometric deformation can exist in a detection box in uploaded pictures. Meanwhile, antigen detection works are carried out by taking families or residences as units, the number of the kits and the types of detection results uploaded by a single picture are various, and a plurality of kits possibly exist on one picture, so that each kit is directly used for identifying the detection result and is affected by various background noises, and the identification accuracy is low. Therefore, the technical staff in the field is dedicated to develop a method for discriminating the antigen detection result of COVID-19 new coronaries, so that the labor cost of medical staff in an epidemic situation prevention and control data center for carrying out secondary discrimination on the uploaded massive antigen detection result is saved, the detection result which is difficult to discriminate manually is accurately judged, the probability of missed detection