CN-115565053-B - Construction method of ballastless track apparent damage sample database

CN115565053BCN 115565053 BCN115565053 BCN 115565053BCN-115565053-B

Abstract

The invention discloses a construction method of a ballastless track apparent damage sample database, which comprises the steps of collecting ballastless track apparent images, constructing ballastless track apparent damage coding standards, marking the ballastless track apparent images, coding damage in the ballastless track apparent images based on the ballastless track apparent damage coding standards, automatically generating ballastless track apparent damage sample legends and attribute files thereof, constructing the ballastless track apparent damage sample database based on the ballastless track apparent damage sample legends and the attribute files thereof, unifying and standardizing ballastless track damage descriptions, efficiently marking ballastless track injury samples and centrally managing the ballastless track injury samples, and facilitating industrial application.

Inventors

CHAI XUESONG
YANG GUOTAO
MAO YULIN
LING LIEPENG
LI JIANCHAO
WANG NING
LIU HAITAO
JIANG ZIQING
ZHAO YONG

Assignees

中国国家铁路集团有限公司
中国铁道科学研究院集团有限公司
中国铁道科学研究院集团有限公司铁道建筑研究所
中铁科学技术开发有限公司

Dates

Publication Date: 20260505
Application Date: 20221013

Claims (7)

1. The construction method of the ballastless track apparent damage sample database is characterized by comprising the following steps of: s100, collecting an apparent image of the ballastless track; S200, constructing an apparent damage coding specification of the ballastless track; s300, automatically pre-identifying the damage in the ballastless track apparent image to obtain the relative position of the damage in the ballastless track apparent image, marking the ballastless track apparent image, and simultaneously encoding the damage in the ballastless track apparent image based on the ballastless track apparent damage encoding specification so as to automatically generate a ballastless track apparent damage sample legend and an attribute file thereof, wherein the automatically pre-identifying the damage in the ballastless track apparent image comprises the following substeps: S311, constructing an image block-based damage detection model, wherein the image block-based damage detection model is constructed by dividing an image without damage into small image blocks, inputting the small image blocks into a deep convolutional neural network model, enabling the model to learn the characteristic distribution of an image without damage background, and training to form the image block-based damage detection model; s312, filtering out background blocks in the ballastless track apparent image based on the damage detection model; S313, performing multi-scale fusion segmentation on the apparent image of the ballastless track based on the image blocks containing the flaws in different scales, so as to obtain the relative positions of the flaws in the apparent image of the ballastless track; s400, constructing a ballastless track apparent damage sample database based on the ballastless track apparent damage sample legend and the attribute files thereof.
2. The construction method according to claim 1, wherein the step S100 further comprises performing an image segmentation preprocessing on the acquired ballastless track apparent image.
3. The construction method according to claim 1, wherein the ballastless track apparent damage coding specification in the step S200 includes four groups including a dictionary code, a feature code, an information code and an image code, eight stages including a part name, a damage type, positioning information, an attribute feature, a damage level, cis-position information, detection information and image information, and eight stages including a plurality of fields.
4. The construction method according to claim 1, wherein labeling the ballastless track apparent image in the step S300 includes: Marking the accurate position of the damage on the apparent image of the ballastless track, thereby obtaining the marked apparent image of the ballastless track.
5. The method of claim 4, wherein the marking of the exact location of the lesion is accomplished by adding a linear lesion marking function to marking tool labelme that adjusts the width of the brush.
6. The method of claim 1, wherein the profile in step S300 includes a unique code identifier for each individual lesion generated according to the ballastless track apparent lesion code specification, the code identifier being associated with a name of a ballastless track apparent lesion sample legend.
7. The construction method according to claim 1, wherein the step S400 further comprises synchronizing the newly generated ballastless track apparent damage sample legend and the attribute files thereof by regularly refreshing the files of the designated positions, entering the ballastless track apparent damage sample database, and displaying the ballastless track apparent damage sample legend and the attribute files thereof on a Web page end.

Description

Construction method of ballastless track apparent damage sample database Technical Field The invention relates to the technical field of railway track detection, in particular to a construction method of a ballastless track apparent damage sample database. Background Ballastless track detection and monitoring equipment is increasingly perfect, comprehensive detection trains, comprehensive inspection vehicles, professional detection vehicles and carrying type monitoring equipment are increasingly perfect, detection and monitoring data are large in scale and various in variety, and characteristics such as multiple sources and mass are presented. With the increase of operation mileage and operation time, the structural damage of the ballastless track line is developed gradually, and the influence of structural damage inspection and maintenance work on operation is also increasingly prominent; the damage names, damage descriptions and the like of the circuit structure in operation are not clearly defined, and checking and recording results is confusing and eight-door, so that great difficulty is brought to subsequent informatization processing, centralized unified management data and statistical analysis. In a railway traffic scene, linear damage images with relatively fine relative sizes are often required to be marked in large-size and low-quality damage images, and the linear damage images often do not have the clear and consistent structure and apparent characteristics as natural object images, so that the special knowledge of labeling people is also required to a certain extent; The existing image marking tools labelme, CVAT and the like can meet the basic requirements of pixel level marking, but the purely manual marking mode is poor in efficiency for massive ballastless track apparent image data, and the ballastless track common flaw cracks, gaps and the like usually show linear structures, are fuzzy in boundary, but are not closed areas formed by a boundary line, but the marking tools adopt closed type and non-loose fuzzy marking methods and are not completely suitable for marking work of ballastless track scenes, and the analysis shows that the pixel level marking in the professional scenes is a very time-consuming and labor-consuming work, and the research on efficient professional marking methods and processes has extremely high practical application value. The deep convolutional neural network based on the artificial intelligence technology increasingly shows the advantages in the image field, aims at the rapid detection and intelligent semantic recognition of the apparent crack, the gap, the defect and other damage states of the ballastless track, has developed a plurality of research works in the industry, accumulates rich image data, does not formally establish a high-quality and large-scale damage sample database aiming at the ballastless track for unified and centralized management, and provides support for the site. Disclosure of Invention Aiming at the problems, the invention aims to provide a construction method of an apparent damage sample database of a ballastless track, which unifies and standardizes the description of the damage of the ballastless track, efficiently marks the damage sample of the ballastless track and centrally manages the sample, and is convenient for industry application. The technical scheme adopted by the invention is that the construction method of the ballastless track apparent damage sample database comprises the following steps: s100, collecting an apparent image of the ballastless track; S200, constructing an apparent damage coding specification of the ballastless track; s300, marking the ballastless track apparent image, and simultaneously encoding the damage in the ballastless track apparent image based on the ballastless track apparent damage encoding specification, so as to automatically generate a ballastless track apparent damage sample legend and an attribute file thereof; s400, constructing a ballastless track apparent damage sample database based on the ballastless track apparent damage sample legend and the attribute files thereof. Preferably, the step S100 further includes performing an image segmentation pretreatment on the acquired ballastless track apparent image. Preferably, the ballastless track apparent damage coding specification in the step S200 comprises four groups of dictionary codes, feature codes, information codes and image codes, wherein the four groups comprise eight stages of part names, damage types, positioning information, attribute features, damage levels, order information, detection information and image information, and the eight stages comprise a plurality of fields. Preferably, labeling the ballastless track apparent image in the step S300 includes: Marking the accurate position of the damage on the apparent image of the ballastless track, thereby obtaining the marked apparent image of the ballastless track. Preferably, the m