Complex background multi-target identification method based on neural network structure search

Abstract

The invention discloses a complex background multi-target recognition method based on neural network structure search, relates to the technical field of sewer detection, solves the technical problems that complex background multi-target recognition cannot be performed on collected images and pipeline inner wall recognition and detection cannot be performed in the prior art, and particularly relates to sewer image collection and processing. Image acquisition is carried out on the sewer, and acquired image optimization is carried out through image data processing; pipeline complex background analysis: carrying out image complex background identification according to the current image acquisition environment, and carrying out sewer pipeline inner wall surface detection through complex background identification; performing pipeline detection under a complex background, and performing sewer pipeline real-time defect detection in combination with a complex background recognition standard; and pipeline area detection: carrying out pipe network identification detection on the area where the sewer pipe network is located, and inferring whether the current sewer pipe network layout is abnormal or not according to pipe network detection.

Description

Technical Field

[0001] The present invention relates to the technical field of sewer detection, and in particular to a complex background multi-target recognition method based on neural network structure search. Background Art

[0002] Neural network structure search uses algorithms to automatically find well-performing network architectures in a huge search space, thereby reducing the workload of manually designing architectures. Multi-target recognition in complex backgrounds is an important task in the field of computer vision, aiming to accurately identify multiple different types of targets from complex background images containing various interference factors (such as messy textures, lighting changes, occlusions, etc.); urban sewer inspection is a key link in ensuring the normal operation of urban drainage systems.

[0003] However, in the existing technology, sewer detection cannot preprocess the real-time collected images, so that the collected pictures are accurate, and it is impossible to perform complex background multi-target recognition on the collected images, and it is impossible to perform pipe inner wall recognition detection, which reduces the accuracy of sewer detection. In addition, it is impossible to perform layout analysis on the network architecture of the pipeline construction, which reduces the drainage efficiency of the pipeline area.

[0004] In view of the above technical defects, a solution is now proposed. Summary of the Invention

[0005] The purpose of the present invention is to solve the above-mentioned problems and to propose a complex background multi-target recognition method based on neural network structure search.

[0006] The purpose of the present invention can be achieved through the following technical solutions:

[0007] A multi-target recognition method for complex backgrounds based on neural network structure search. The steps of the multi-target recognition method are as follows:

[0008] Sewer image acquisition and processing: collect images of sewers, process image data of various parts of the sewers based on the acquired images, and optimize the acquired images through image data processing;

[0009] Complex background analysis of pipelines: After completing the acquisition and processing of sewer image data, complex background recognition of the image is performed according to the current image acquisition environment, and the inner wall surface of the sewer pipe is detected through complex background recognition;

[0010] Pipeline detection under complex background: After complex background recognition of sewer images, sewer pipeline detection is performed under the current complex background, and real-time defect detection of sewer pipelines is performed in combination with complex background recognition standards;

[0011] Detection of the area where the pipeline is located. After the sewer image detection is completed, the sewer pipes detected in real time are connected to build a sewer network, and the pipe network identification and detection are performed on the area where the sewer network is located. Based on the pipe network detection, it is inferred whether there are any abnormalities in the current sewer network layout.

[0012] As a preferred embodiment of the present invention, the sewer image acquisition and processing process is as follows:

[0013] The sewer area to be inspected is marked as a pipeline area, and image acquisition is performed. The acquired images are divided into several acquired pictures, and the pictures are sorted in the time sequence of the acquisition time. The real-time light intensity of the pipeline position corresponding to the acquired picture is obtained, and the light intensity standard value is set. The light intensity of the corresponding pipeline position is marked according to the acquisition time sequence, and a floating curve of the real-time light intensity of the pipeline position is constructed by connecting all the points.

[0014] The parameters are set based on the real-time light intensity of the captured images obtained in the current sequence. With the same light intensity standard value, the light intensity of the captured images of the corresponding sequence is also collected and the points are marked. The light intensity floating curve on the pipeline location map is constructed based on the marked points.

[0015] As a preferred embodiment of the present invention, any adjacent moments are analyzed, and the coverage area deviation corresponding to the same two adjacent moments of the real-time light intensity floating curve and the light intensity floating curve on the map, and the curve trajectory offset angle deviation corresponding to the same horizontal coordinate point of the real-time light intensity floating curve and the light intensity floating curve on the map are collected and marked as curve numerical information and curve floating information respectively;

[0016] And compare with the coverage area deviation threshold and angle deviation threshold respectively:

[0017] If the numerical information of the curve exceeds the coverage area deviation threshold, or the floating information of the curve exceeds the angle deviation threshold, it is inferred that the current curve segment is marked as a light effective adjustment segment; if the numerical information of the curve does not exceed the coverage area deviation threshold, and the floating information of the curve does not exceed the angle deviation threshold, it is inferred that the current curve segment is marked as a light invalid adjustment segment.

[0018] As a preferred embodiment of the present invention, a standard line is set according to the standard value of light intensity, and the light intensity floating curve on the graph is analyzed. If the effective light adjustment section in the light intensity floating curve on the graph is below the standard line, it is inferred that the light satisfaction of the current image acquisition is poor, and the corresponding acquired picture will be re-captured and the fill light equipment will be replaced to expand the fill light range; if the invalid light adjustment section in the light intensity floating curve on the graph is below the standard line, it is inferred that the light adjustment of the current image acquisition is inefficient, and the corresponding acquired picture will be re-fill light adjusted.

[0019] As a preferred embodiment of the present invention, the pipeline complex background analysis process is as follows:

[0020] The detection area is obtained based on the excerpt of the collected image, the color depth of the pipeline bottom layer is collected, the initial color is obtained based on the real-time image collection, the detection area is divided into several pixel points, and the pipeline bottom layer area and the attached surface area are obtained based on the color depth distribution of the detection area;

[0021] The attachment area data and attachment thickness data are collected, and the attachment area data and attachment thickness data are respectively the increased span of the area ratio of the pipeline bottom area to the attachment surface area, the real-time light intensity floating curve, and the curve trajectory offset angle deviation of the light intensity floating curve on the graph corresponding to the same horizontal coordinate point.

[0022] As a preferred embodiment of the present invention, if the attachment area data does not exceed the area ratio span threshold, or the attachment thickness data exceeds the mean deviation expansion rate threshold, the pipeline position where the corresponding detection area is located is marked as an abnormally attached pipeline; if the attachment area data exceeds the area ratio span threshold and the attachment thickness data does not exceed the mean deviation expansion rate threshold, the pipeline position where the corresponding detection area is located is marked as a normally attached pipeline.

[0023] As a preferred embodiment of the present invention, the pipeline detection process under complex background is as follows:

[0024] The bottom layer texture information and attachment thickness information are collected, and the bottom layer texture information and attachment thickness information are the corresponding values ​​of the frequency of change of the surface texture trend trajectory of the bottom area of ​​the pipeline and the maximum span of the texture trend trajectory floating, and the speed of reduction of the pipeline thickness in the current area during the stage of increasing attachment thickness of the attachment surface area.

[0025] As a preferred embodiment of the present invention, if the underlying texture information exceeds the frequency span and threshold, or the attached thickness information exceeds the thickness reduction rate threshold, the pipeline portion corresponding to the corresponding acquired image detection area is marked as a defect abnormal portion; if the underlying texture information does not exceed the frequency span and threshold, and the attached thickness information does not exceed the thickness reduction rate threshold, the pipeline portion corresponding to the corresponding acquired image detection area is marked as a defect low-risk portion.

[0026] As a preferred embodiment of the present invention, the pipeline area detection process is as follows:

[0027] The pipeline area locations corresponding to the low-risk parts of the pipeline network defects are collected and marked as pipeline network risk locations; the internal impact data and external impact data are obtained, and the internal impact data and external impact data are the proportion of the full-diameter drainage period of the pipeline corresponding to the pipeline network risk location to the total operating period, and the corresponding increase value of the external pressure of the pipeline corresponding to the pipeline network risk location and the increase value of the inner wall surface deformation at the same location.

[0028] As a preferred embodiment of the present invention, if the internal influence data exceeds the time period proportion threshold, or the external influence data exceeds the numerical increase threshold, it is inferred that there is a risk in the pipeline area corresponding to the pipeline network risk position; if the internal influence data does not exceed the time period proportion threshold, and the external influence data does not exceed the numerical increase threshold, it is inferred that there is temporarily no risk in the pipeline area corresponding to the pipeline network risk position.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] 1. In the present invention, images of the sewer are collected, and image data processing is performed on various parts of the sewer based on the collected images. The collected images are optimized through image data processing to avoid the situation in which the collected images during the sewer inspection process cannot meet the actual inspection requirements, thereby increasing the difficulty of complex background recognition and failing to accurately perform sewer inspection;

[0031] After completing the acquisition and processing of sewer image data, complex background recognition of the image is performed according to the current image acquisition environment. The inner wall surface of the sewer pipe is detected through complex background recognition to improve the sewer detection efficiency, avoid surface anomalies in the sewer pipe, reduce the operation quality of the sewer pipe and make it impossible to discharge sewage in time. The inner wall surface detection of the sewer pipe can timely issue pipeline early warning, facilitate timely early warning and maintenance when abnormalities are found, and effectively reduce the risk of sewer pipe blockage or damage.

[0032] 2. In the present invention, after complex background recognition of the sewer image, sewer pipe inspection is performed under the current complex background. In combination with the complex background recognition standard, real-time sewer pipe defect detection is performed to ensure the efficiency of sewer pipe defect detection, ensure the stability of pipeline operation, and prevent pipeline damage from reducing pipeline availability, hindering the sewage discharge progress in the current area, and affecting the transportation balance of the entire sewer network;

[0033] Based on the pipe network detection, it is inferred whether there are any abnormalities in the current sewer network layout, so as to facilitate real-time pipe network planning based on the sewer network analysis, to avoid some locations in the pipe network being unsuitable for pipe laying, which will cause the drainage efficiency of the sewer network to decrease. Blockage of any location will cause the drainage balance of the entire sewer network. Therefore, through pipe network detection, the drainage efficiency of the current area can be guaranteed in real time. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] To facilitate understanding by those skilled in the art, the present invention is further described below with reference to the accompanying drawings.

[0035] Figure 1 Schematic diagram of the overall method flow of the present invention;

[0036] Figure 2 The figure is a flow chart of the image acquisition and processing method of the present invention. DETAILED DESCRIPTION

[0037] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0038] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present invention. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute a separate or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0039] This application is used for sewer detection. It performs pipeline identification and analysis in complex background scenes based on sewer image recognition and analysis, and performs sewer layout detection based on the real-time construction of the pipeline network. Based on the neural network structure search technology, it constructs the sewer network and performs detection based on the current network. Please refer to Figure 1 As shown in the figure, the complex background multi-target recognition method based on neural network structure search, the specific steps of the multi-target recognition method are as follows:

[0040] Sewer image acquisition and processing: collect images of sewers, process image data of various parts of the sewers based on the collected images, and optimize the collected images through image data processing to avoid the collected images failing to meet the actual detection requirements during the sewer inspection process, thereby increasing the difficulty of complex background recognition and failing to accurately perform sewer inspections;

[0041] Complex background analysis of pipelines: After completing the acquisition and processing of sewer image data, complex background recognition of images is performed according to the current image acquisition environment. The inner wall surface of the sewer pipe is detected through complex background recognition to improve the efficiency of sewer detection, avoid surface anomalies in the sewer pipe, reduce the operation quality of the sewer pipe and make it impossible to discharge sewage in time. The inner wall surface detection of the sewer pipe can timely issue pipeline early warning, facilitate timely early warning and maintenance when abnormalities are found, and effectively reduce the risk of sewer pipe blockage or damage;

[0042] Pipeline detection under complex backgrounds: After complex background recognition of sewer images, sewer pipeline inspection is performed under the current complex background. Real-time sewer pipeline defect detection is performed in combination with complex background recognition standards to ensure the efficiency of sewer pipeline defect detection, ensure pipeline operation stability, and prevent pipeline damage from reducing pipeline availability, hindering the sewage discharge progress in the current area, and affecting the transportation balance of the entire sewer network;

[0043] Detection of the area where the pipeline is located. After the sewer image detection is completed, the sewer pipes detected in real time are connected to build a sewer network. The network identification and detection are performed on the area where the sewer network is located. Based on the network detection, it is inferred whether there are any abnormalities in the current sewer network layout, so as to facilitate real-time network planning based on the sewer network analysis, and avoid some locations in the network being unsuitable for pipe laying, which will cause the drainage efficiency of the sewer network to decrease. Blockage of any location will cause the drainage balance of the entire sewer network. Therefore, the drainage efficiency of the current area can be guaranteed in real time through network detection.

[0044] See also Figure 2 As shown in the figure, the sewer image acquisition and processing process is as follows:

[0045] The sewer area to be inspected is marked as a pipeline area, and images of the pipeline area are collected. The collected images are divided into several collected pictures, and the pictures are sorted in the time sequence of the collection time. The real-time light intensity of the pipeline position corresponding to the collected pictures is obtained, and the light intensity standard value is set. The light intensity of the corresponding pipeline position is marked according to the collection time sequence, and a floating curve of the real-time light intensity of the pipeline position is constructed by connecting all the points.

[0046] At the same time, the real-time light intensity setting parameters of the collected pictures are obtained according to the current sequence, that is, the light intensity after the fill light is applied before the pictures are collected; with the same light intensity standard value, the light intensity of the pictures collected in the corresponding sequence is also collected and the points are marked, and the light intensity floating curve on the pipeline location map is constructed according to the marked points;

[0047] Compare the real-time light intensity floating curve with the light intensity floating curve on the map: set an interval threshold for the corresponding horizontal axis during the comparison process, where the pipeline position order is consistent with the image acquisition order;

[0048] Analyze any adjacent moments, collect the coverage area deviation corresponding to the same two adjacent moments of the real-time light intensity floating curve and the light intensity floating curve on the map, and the curve trajectory offset angle deviation corresponding to the same horizontal coordinate point of the real-time light intensity floating curve and the light intensity floating curve on the map, and mark them as curve value information and curve floating information, respectively, and compare them with the coverage area deviation threshold and angle deviation threshold respectively:

[0049] If the coverage area deviation between the real-time light intensity floating curve and the light intensity floating curve on the map at the same two adjacent moments exceeds the coverage area deviation threshold, or the angle deviation between the real-time light intensity floating curve and the light intensity floating curve on the map at the same horizontal coordinate point exceeds the angle deviation threshold, it is inferred that the current curve segment is marked as an effective light adjustment segment;

[0050] If the coverage area deviation between the real-time light intensity floating curve and the light intensity floating curve on the map at the same two adjacent moments does not exceed the coverage area deviation threshold, and the curve trajectory offset angle deviation between the real-time light intensity floating curve and the light intensity floating curve on the map at the same horizontal coordinate point does not exceed the angle deviation threshold, then it is inferred that the current curve segment is marked as an invalid light adjustment segment;

[0051] A standard line is set according to the standard value of light intensity, and the light intensity floating curve on the graph is analyzed. If the effective light adjustment section in the light intensity floating curve on the graph is below the standard line, it is inferred that the light satisfaction of the current image acquisition is poor, and the corresponding captured image will be recaptured and the fill light equipment will be replaced to expand the fill light range; if the invalid light adjustment section in the light intensity floating curve on the graph is below the standard line, it is inferred that the light adjustment of the current image acquisition is inefficient, and the corresponding captured image will be re-filled with light.

[0052] After completing the light intensity adjustment of the captured image, the real-time captured image will be used for complex background analysis of the pipeline;

[0053] The process of analyzing the complex background of the pipeline is as follows:

[0054] The collected image is used to identify the complex background of the pipeline. The interior of the sewer is excerpted from the collected image and marked as the detection area. The color depth of the bottom layer of the sewer pipe in the current collected image is collected, and the corresponding bottom layer color depth is marked as the initial color under the supplementary light intensity of the current collected image. The detection area is divided into several pixels, and the color depth distribution of each pixel in the detection area in the collected image is obtained. Through initial color comparison, the area covered by the initial color and color depth is marked as the pipeline bottom layer area; and the area covered by the non-initial color and color depth is marked as the attachment surface area;

[0055] The increase span of the area ratio of the detection area in the acquired image to the pipeline bottom area and the attachment surface area is obtained, and the expansion speed of the mean deviation of the thickness between the attachment surface area and the pipeline bottom area in the acquired image is obtained. The increase span of the area ratio of the detection area in the acquired image to the pipeline bottom area and the attachment surface area and the expansion speed of the mean deviation of the thickness between the attachment surface area and the pipeline bottom area are marked as attachment area data and attachment thickness data, respectively, and compared with the area ratio span threshold and the mean deviation expansion speed threshold, respectively: wherein, the acquired image is extracted from the acquired image at the inner wall surface shooting angle and the pipeline transportation direction shooting angle, respectively used for intuitive detection of inner wall surface defects and inner wall attachment thickness detection;

[0056] If the increase span of the area ratio between the bottom layer of the pipeline and the attachment surface area in the detection area corresponding to the collected image does not exceed the area ratio span threshold, or the expansion speed of the mean deviation of the thickness between the attachment surface area and the bottom layer of the pipeline in the collected image exceeds the mean deviation expansion speed threshold, it is inferred that there is an attachment abnormality in the detection area corresponding to the collected image, and the position of the pipeline corresponding to the detection area is marked as an attachment abnormality pipeline, and the attachment abnormality pipeline is sent to the maintenance personnel terminal for inspection and dredging;

[0057] If the increase span of the area ratio of the pipeline bottom area to the attachment surface area corresponding to the detection area in the collected image exceeds the area ratio span threshold, and the expansion speed of the mean deviation of the thickness between the attachment surface area and the pipeline bottom area in the collected image does not exceed the mean deviation expansion speed threshold, it is inferred that the detection area corresponding to the collected image is normally attached, and the pipeline position where the corresponding detection area is located is marked as a pipeline with normal attachment, and the current pipeline maintenance cycle is not changed;

[0058] The pipeline detection process under complex background is as follows:

[0059] The frequency of changes in the surface texture trend trajectory of the pipeline bottom area and the corresponding numerical sum of the maximum floating span of the texture trend trajectory are obtained in the detection area of ​​the collected image. The numerical sum is expressed as the numerical sum of the frequency and span data. The influence of the numerical fluctuation is inferred, and the numerical sum of the frequency of changes in the surface texture trend trajectory of the pipeline bottom area and the corresponding numerical sum of the minimum floating span of the texture trend trajectory are marked as the bottom texture information;

[0060] The decreasing speed of the pipe thickness in the current area during the stage of increasing thickness of the attachment surface area is obtained in the detection area of ​​the collected image. The corrosion effect of the attachment is inferred through the internal thickness fluctuation, and the decreasing speed of the pipe thickness in the current area during the stage of increasing thickness of the attachment surface area is marked as attachment thickness information.

[0061] Compare the underlying texture information and the attached thickness information with the frequency span and threshold, and the thickness reduction speed threshold respectively:

[0062] If the frequency of the surface texture trend trajectory change in the pipeline bottom area and the corresponding value of the maximum span of the texture trend trajectory fluctuation exceed the frequency span and threshold, or the speed of the pipe thickness reduction in the current area during the thickness increase phase of the attached surface area exceeds the thickness reduction speed threshold, it is inferred that the sewer pipe detection in the image collected under the complex background is abnormal, and the pipe part corresponding to the detection area of ​​the corresponding collected image is marked as a defective abnormal part. The maintenance personnel will count and inspect and maintain the defective abnormal part;

[0063] If the sum of the corresponding values ​​of the frequency of changes in the surface texture trend trajectory of the pipeline bottom area and the maximum span of the texture trend trajectory fluctuation does not exceed the frequency span and threshold, and the speed of decrease in the thickness of the pipeline in the current area during the stage of increasing adhesion thickness in the adhesion surface area does not exceed the thickness decrease speed threshold, then it is inferred that the sewer pipeline inspection corresponding to the image collected under the complex background is normal, and the pipeline part corresponding to the detection area of ​​the corresponding collected image is marked as a low-risk defect part;

[0064] The detection process of the pipeline area is as follows:

[0065] After the sewer pipe inspection, the current pipe network is analyzed, and the locations of the pipe areas corresponding to the low-risk areas of the pipe network defects are collected and marked as risky locations of the pipe network;

[0066] During the operation period of the pipeline network, the ratio of the full-diameter drainage period of the pipeline corresponding to the risk position of the pipeline network to the total operation period is obtained, and the ratio of the full-diameter drainage period of the pipeline corresponding to the risk position of the pipeline network to the total operation period is marked as the intrinsic impact data; the full-diameter drainage is represented by the real-time flow rate during the sewage discharge process and the maximum flow rate of the pipeline diameter;

[0067] During the period when the pipeline is not drained, the sum of the increase in external pressure of the pipeline corresponding to the risk position of the pipeline network and the increase in the deformation of the inner wall surface at the same position is obtained, and the sum of the increase in external pressure of the pipeline corresponding to the risk position of the pipeline network and the increase in the deformation of the inner wall surface at the same position is marked as external impact data;

[0068] Compare the internal impact data and external impact data with the time period proportion threshold and value increase threshold respectively:

[0069] If the internal impact data exceeds the time period proportion threshold, or the external impact data exceeds the value increase threshold, the pipeline specifications at the corresponding pipeline network risk location in the pipeline area need to be replaced. If the difference between the pipeline specifications at that location and those at other locations is too large, it is inferred that there is a risk at the corresponding pipeline network risk location in the pipeline area. The pipeline network path is planned for the current pipeline network risk location, and the pipeline network is rebuilt. The pipeline specifications are expressed as the pressure bearing value of the pipeline material, the external impact force, the minimum temperature, etc.

[0070] If the internal impact data does not exceed the time period proportion threshold, and the external impact data does not exceed the value increase threshold, it is inferred that the corresponding pipeline network risk location in the pipeline area is temporarily risk-free;

[0071] When the present invention is used, the sewer image acquisition and processing is performed, the sewer image is acquired, image data processing is performed on various parts of the sewer according to the acquired image, and the acquired image is optimized through image data processing; the pipeline complex background analysis is completed, and after the sewer image data acquisition and processing is completed, the image complex background recognition is performed according to the current image acquisition environment, and the inner wall surface of the sewer pipe is detected through complex background recognition;

[0072] Pipeline detection under complex background: After complex background recognition of the sewer image, sewer pipeline detection is performed under the current complex background, and real-time defect detection of sewer pipelines is performed in combination with complex background recognition standards; pipeline area detection: after the sewer image detection is completed, the sewer pipelines after real-time detection are connected to construct the sewer network, and pipeline network identification and detection are performed on the area where the sewer network is located. Based on the pipeline network detection, it is inferred whether there are any abnormalities in the current sewer network layout.

[0073] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the present invention to specific embodiments. Obviously, many modifications and variations are possible based on the contents of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A complex background multi-target recognition method based on neural network structure search, characterized in that: The steps of multi-target recognition method are as follows: Sewer image acquisition and processing: collect images of sewers, process image data of various parts of sewers according to the acquired images, and optimize the acquired images through image data processing; Complex background analysis of pipelines: After completing the acquisition and processing of sewer image data, complex background recognition of images is performed according to the current image acquisition environment, and the inner wall surface of the sewer pipe is detected through complex background recognition; Pipeline detection under complex background: after complex background recognition of sewer images, sewer pipeline detection is performed under the current complex background, and real-time defect detection of sewer pipelines is performed in combination with complex background recognition standards; Detection of the area where the pipeline is located. After the sewer image detection is completed, the sewer pipes detected in real time are connected to build the sewer network, and the network identification and detection are performed on the area where the sewer network is located. Based on the network detection, it is inferred whether there is any abnormality in the current sewer network layout.

2. The complex background multi-target recognition method based on neural network structure search according to claim 1 is characterized in that: The sewer image acquisition and processing process is as follows: The area where the sewer needs to be inspected is marked as the pipeline area, and image acquisition is performed. The acquired images are divided into several acquired pictures, and the pictures are sorted in the time sequence of the acquisition time. The real-time light intensity of the pipeline position corresponding to the acquired picture is obtained, and the light intensity standard value is set. According to the acquisition time sequence, the light intensity of the corresponding pipeline position is marked at the point, and the real-time light intensity floating curve of the pipeline position is constructed by connecting all the points; The parameters are set according to the real-time light intensity of the collected pictures obtained in the current sequence; with the same light intensity standard value, the light intensity of the collected pictures of the corresponding sequence is also collected and the points are marked, and the light intensity floating curve on the pipeline location map is constructed according to the marked points.

3. The complex background multi-target recognition method based on neural network structure search according to claim 2 is characterized in that: Analyze any adjacent moments, collect the coverage area deviation of the real-time light intensity floating curve and the light intensity floating curve on the map corresponding to the same two adjacent moments, and the curve trajectory offset angle deviation of the real-time light intensity floating curve and the light intensity floating curve on the map corresponding to the same horizontal coordinate point, and mark them as curve value information and curve floating information respectively; And compared with the coverage area deviation threshold and angle deviation threshold respectively: If the numerical information of the curve exceeds the coverage area deviation threshold, or the floating information of the curve exceeds the angle deviation threshold, it is inferred that the current curve segment is marked as a light effective adjustment segment; if the numerical information of the curve does not exceed the coverage area deviation threshold, and the floating information of the curve does not exceed the angle deviation threshold, it is inferred that the current curve segment is marked as a light invalid adjustment segment.

4. The complex background multi-target recognition method based on neural network structure search according to claim 3 is characterized in that: A standard line is set according to the standard value of light intensity, and the light intensity floating curve on the graph is analyzed. If the effective light adjustment section in the light intensity floating curve on the graph is below the standard line, it is inferred that the light satisfaction of the current image acquisition is poor, and the corresponding acquired picture will be re-captured and the fill light equipment will be replaced to expand the fill light range; if the invalid light adjustment section in the light intensity floating curve on the graph is below the standard line, it is inferred that the light adjustment of the current image acquisition is inefficient, and the corresponding acquired picture will be re-fill light adjusted.

5. The complex background multi-target recognition method based on neural network structure search according to claim 1 is characterized in that: The process of complex pipeline background analysis is as follows: The detection area is obtained according to the excerpt of the collected image, the color depth of the bottom layer of the pipeline is collected, the initial color is obtained according to the real-time image collection, the detection area is divided into several pixel points, and the bottom layer area of ​​the pipeline and the attached surface area are obtained according to the color depth distribution of the detection area; The attachment area data and the attachment thickness data are collected, and the attachment area data and the attachment thickness data are respectively the increased span of the area ratio of the pipeline bottom area to the attachment surface area, the real-time light intensity floating curve and the curve trajectory offset angle deviation of the light intensity floating curve on the graph corresponding to the same horizontal coordinate point.

6. The complex background multi-target recognition method based on neural network structure search according to claim 5 is characterized in that: If the attachment area data does not exceed the area ratio span threshold, or the attachment thickness data exceeds the mean deviation expansion speed threshold, the pipeline position of the corresponding detection area will be marked as an abnormally attached pipeline; if the attachment area data exceeds the area ratio span threshold, and the attachment thickness data does not exceed the mean deviation expansion speed threshold, the pipeline position of the corresponding detection area will be marked as a normally attached pipeline.

7. The complex background multi-target recognition method based on neural network structure search according to claim 1 is characterized in that The pipeline detection process under complex background is as follows: The bottom layer texture information and attachment thickness information are collected, and the bottom layer texture information and attachment thickness information are respectively the frequency of change of the surface texture trend trajectory of the bottom layer of the pipeline and the corresponding values ​​of the maximum floating span of the texture trend trajectory, and the speed of reduction of the pipeline thickness in the current area during the stage of increasing attachment thickness in the attachment surface area.

8. According to the complex background multi-target recognition method based on neural network structure search in claim 7, if the underlying texture information exceeds the frequency span and threshold, or the attached thickness information exceeds the thickness reduction speed threshold, the pipeline part corresponding to the corresponding acquisition image detection area is marked as a defect abnormal part; if the underlying texture information does not exceed the frequency span and threshold, and the attached thickness information does not exceed the thickness reduction speed threshold, the pipeline part corresponding to the corresponding acquisition image detection area is marked as a defect low risk part.

9. According to the complex background multi-target recognition method based on neural network structure search in claim 1, the pipeline area detection process is as follows: The pipeline area locations corresponding to the low-risk parts of pipeline network defects are collected and marked as pipeline network risk locations; the internal impact data and external impact data are obtained, and the internal impact data and external impact data are the proportion of the full-diameter drainage period of the pipeline corresponding to the pipeline network risk location to the total operating period, the increase value of the external pressure of the pipeline corresponding to the pipeline network risk location and the increase value of the inner wall surface deformation at the same location.

10. According to the complex background multi-target recognition method based on neural network structure search described in claim 9, if the internal influence data exceeds the time period proportion threshold, or the external influence data exceeds the numerical increase threshold, it is inferred that there is a risk in the pipeline area corresponding to the pipeline network risk position; if the internal influence data does not exceed the time period proportion threshold, and the external influence data does not exceed the numerical increase threshold, it is inferred that the pipeline area corresponding to the pipeline network risk position is temporarily risk-free.

Images