CN110097082B

CN110097082B - Splitting method and device of training set

Info

Publication number: CN110097082B
Application number: CN201910251196.4A
Authority: CN
Inventors: 朱喻
Original assignee: Guangzhou Side Medical Technology Co ltd
Current assignee: Guangzhou Side Medical Technology Co ltd
Priority date: 2019-03-29
Filing date: 2019-03-29
Publication date: 2021-08-27
Anticipated expiration: 2039-03-29
Also published as: CN110097082A

Abstract

The embodiment of the invention provides a method and a device for splitting a training set, wherein the method comprises the following steps: acquiring a training set for training a preset model; the training set comprises a first target picture training set based on global structural features; the first target picture training set is a training set determined based on a first target picture with a gastric angle structure and/or a texture structure with special features globally; the first target picture is a picture of which the outer surface of the shot does not contain abnormal features; identifying picture features in the first target picture training set, and if the class to which the picture features belong is judged to be the first class, selecting a corresponding five-level comparison set to compare the pictures corresponding to the first class; and splitting the first target picture training set according to the comparison result. The device performs the above method. The training set splitting method and device provided by the embodiment of the invention can improve the splitting rationality of the training set.

Description

Splitting method and device of training set

Technical Field

The embodiment of the invention relates to the technical field of picture processing, in particular to a method and a device for splitting a training set.

Background

The capsule endoscopy has the advantages of no pain, no injury, large information amount of shot images and the like, and has wide application value.

In the prior art, an original picture shot through a capsule endoscope is identified in a manual mode and is classified, a model needs to be built for identifying the original picture more accurately and efficiently, but the model usually needs to be trained before use, and a training set in the training process needs to be split, so that the model can identify the picture more accurately, but no effective method exists for splitting the training set at present.

Therefore, how to avoid the above-mentioned drawbacks and improve the rationality of splitting the training set becomes a problem that needs to be solved urgently.

Disclosure of Invention

Aiming at the problems in the prior art, the embodiment of the invention provides a method and a device for splitting a training set.

The embodiment of the invention provides a method for splitting a training set, which comprises the following steps:

acquiring a training set for training a preset model; the training set comprises a first target picture training set based on global structural features; the first target picture training set is a training set determined based on a first target picture with a gastric angle structure and/or a texture structure with special features globally; the first target picture is a picture of which the outer surface of the shot does not contain abnormal features; the abnormal features comprise raised features and/or designated color features; the special features comprise dense stripe features; the number of the stripes in the compact stripe characteristics is greater than a preset number;

identifying picture features in the first target picture training set, and if the class to which the picture features belong is judged to be the first class, selecting a corresponding five-level comparison set to compare the pictures corresponding to the first class;

and splitting the first target picture training set according to the comparison result.

The embodiment of the invention provides a device for splitting a training set, which comprises:

the device comprises an acquisition unit, a processing unit and a control unit, wherein the acquisition unit is used for acquiring a training set used for training a preset model; the training set comprises a first target picture training set based on global structural features; the first target picture training set is a training set determined based on a first target picture with a gastric angle structure and/or a texture structure with special features globally; the first target picture is a picture of which the outer surface of the shot does not contain abnormal features; the abnormal features comprise raised features and/or designated color features; the special features comprise dense stripe features; the number of the stripes in the compact stripe characteristics is greater than a preset number;

the selecting unit is used for identifying the picture characteristics in the first target picture training set, and if the class to which the picture characteristics belong is judged to be the first class, selecting a corresponding five-level comparison set to compare the pictures corresponding to the first class;

and the splitting unit is used for splitting the first target picture training set according to the comparison result.

An embodiment of the present invention provides an electronic device, including: a processor, a memory, and a bus, wherein,

the processor and the memory are communicated with each other through the bus;

the memory stores program instructions executable by the processor, the processor invoking the program instructions to perform a method comprising:

An embodiment of the present invention provides a non-transitory computer-readable storage medium, including:

the non-transitory computer readable storage medium stores computer instructions that cause the computer to perform a method comprising:

According to the training set splitting method and device provided by the embodiment of the invention, if the class to which the picture features in the first target picture training set based on the global structure features belong is judged to be the first class, the five-level comparison set is selected to compare the pictures corresponding to the first class, and the first target picture training set is split according to the comparison result, so that the splitting reasonability of the training set can be improved.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and those skilled in the art can also obtain other drawings according to the drawings without creative efforts.

FIG. 1 is a flowchart of an embodiment of a training set splitting method according to the present invention;

FIG. 2 is a schematic structural diagram of a training set splitting apparatus according to an embodiment of the present invention;

fig. 3 is a schematic structural diagram of an electronic device according to an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, but not all, embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

Fig. 1 is a flowchart of an embodiment of a splitting method of a training set according to the present invention, and as shown in fig. 1, the splitting method of the training set according to the embodiment of the present invention includes the following steps:

s101: acquiring a training set for training a preset model; the training set comprises a first target picture training set based on global structural features; the first target picture training set is a training set determined based on a first target picture with a gastric angle structure and/or a texture structure with special features globally; the first target picture is a picture of which the outer surface of the shot does not contain abnormal features; the abnormal features comprise raised features and/or designated color features; the special features comprise dense stripe features; the number of stripes in the dense stripe feature is greater than a preset number.

Specifically, the device acquires a training set for training a preset model; the training set comprises a first target picture training set based on global structural features; the first target picture training set is a training set determined based on a first target picture with a gastric angle structure and/or a texture structure with special features globally; the first target picture is a picture of which the outer surface of the shot does not contain abnormal features; the abnormal features comprise raised features and/or designated color features; the special features comprise dense stripe features; the number of stripes in the dense stripe feature is greater than a preset number. It should be noted that: the original picture is shot by the capsule endoscope, and the working process of the capsule endoscope is explained as follows:

the capsule endoscope enters the digestive tract from the oral cavity and is naturally discharged from the anus.

The battery of capsule endoscopy has limited endurance, and the effective working space is a part of the mouth, esophagus, stomach, duodenum, small intestine and large intestine.

Each activity of the capsule endoscope produces an in-field exam picture and an out-of-field exam picture.

The intra-field examination picture is a result of taking a certain section of the digestive tract.

The out-of-field inspection picture is a picture taken by the capsule endoscope in addition to the in-field inspection picture.

All pictures can be automatically identified without any human intervention (including image pre-processing).

After the images are identified, the images taken by the capsule endoscopy are divided into six major categories (125 minor categories) and automatically saved in 125 image folders, wherein the six major categories can be:

the first major category: one class of out-of-domain category labels (10 classes).

The second major category: class two out-of-domain category labels (13 categories).

The third major category: the tags (14 classes) are classified based on the first target picture of the local structural features.

The fourth major category: hole-structured first target picture classification tags (8 classes).

The fifth main category: the tags (24 classes) are classified based on the first target picture of the global structural features.

The sixth major class: the second target picture category label (56 categories).

It is possible to automatically recognize different parts of the digestive tract such as the oral cavity, the esophagus, the stomach, the duodenum, the small intestine, and the large intestine.

The number of the original pictures which can be shot by each capsule endoscope at each time can be 2000-3000, namely the number of the pictures which are acquired by the capsule endoscopes and concentrated.

Raw pictures taken of the capsule endoscopy (JPG format) can be derived from the hospital information system without any processing. The first target picture training set in the embodiment of the present invention corresponds to the first target picture classification label based on the global structural feature, and is specifically described as follows: the first target picture classification label based on the global structural feature may include a stomach corner structure picture, a texture structure picture with a special feature, and the like, each subclass (for example, one of the subclasses is the stomach corner structure picture) includes a certain number of sample pictures that can be used as training samples, and the first target picture training set based on the global structural feature may be understood as a training set including all sample pictures corresponding to all subclasses; the raised features may include swollen, particulate matter raised. The designated color characteristics may include red and white, and are not particularly limited. It should be noted that: the abnormal feature can be used as an intermediate reference feature in some disease diagnosis processes, and the abnormal feature is not enough to diagnose the disease only by relying on the abnormal feature. The pictures in the first target picture set can be understood as standard pictures for evaluating the target part to be detected, and the target part to be detected can be the stomach. The specific numerical value of the preset quantity can be set independently according to the actual situation.

It should be noted that: the first target picture with the gastric angle structure in the whole situation can specifically comprise a half-moon-shaped gastric angle structure picture, a hole-shaped gastric angle structure picture, a double-hole gastric angle structure picture and a partial gastric angle structure picture; wherein, the edge part of the half-moon-shaped stomach angle structure picture is in a half-moon shape; the center of the hole-shaped stomach corner structure picture is a non-hollow hole-shaped structure; the two opposite positions on the two sides of the outer surface of the shot in the double-hole gastric angle structure picture are respectively provided with a non-hollow hole-shaped structure; the partial stomach angle structure pictures comprise shape structures corresponding to stomach angles in a contraction state and a relaxation state, each subdivision classification (for example, one type of the partial stomach angle structure pictures is a half-moon-shaped stomach angle structure picture) comprises a certain number of sample pictures which can be used as training samples, and each subsequent subdivision classification also comprises a certain number of sample pictures which can be used as training samples, and the details are not repeated.

The first target picture of the texture structure with the overall dense stripe feature can specifically comprise a dense thick stripe picture, a dense fine stripe picture, a dense superficial stripe picture and a dense stripe semilunar structure picture, wherein the dense thick stripe picture does not have a hole-shaped structure and the drift diameter of the stripe is larger than a first preset drift diameter value; the compact fine stripe picture has no hole-shaped structure, and the drift diameter of the stripe is smaller than a first preset drift diameter value; the compact superficial stripe picture has no hole structure, and the protruding height of the stripe along the direction of the outer surface of the shot is smaller than a preset height value; the compact stripe half-moon structure picture is provided with a hole-shaped structure, the drift diameter of the stripe is larger than a first preset drift diameter value, and the whole structure after the hole-shaped structure is removed presents a meniscus-shaped structure. The first preset drift diameter numerical value and the preset height numerical value can be set independently according to actual conditions.

The special features may further include water-covering texture features, and accordingly, the first target picture based on the global texture structure with the water-covering texture features includes: a water covering hole structure picture, a water covering shadow structure picture, a water covering non-intersecting structure picture, a water covering flat structure picture, a complex background waterline picture, a water covering complex structure picture and a transparent structure picture; the water covering hole structure picture is a hole-shaped structure picture which is shot through a long shot and covered by the water surface; the shadow part in the water covering shadow structure picture is a part covered by the water surface; the water-covered non-intersecting structure picture is a non-intersecting structure picture covered by a water surface; the non-intersecting structure picture comprises a plurality of groups of non-intersecting thick and strong prismatic structures and also comprises a partial dark area; the drift diameter of the thick and strong prismatic structure is larger than a first preset drift diameter value; the water covering flat structure picture is a blood vessel flat structure picture covered by the water surface; the complex background waterline picture is a complex shooting background penetrated by a water-vapor interface, and the complex shooting background is a shooting background with a plurality of prismatic structure characteristics; the water covering complex structure picture is a complex shooting background which is completely covered by the water surface; the water vapor interface of the transparent structure picture penetrates through the outer surface of the shot, and one side of the water vapor interface is a transparent water body.

The embodiment of the present invention focuses on the description of the first target picture training set based on the global structural feature corresponding to the fifth main category.

S102: and identifying picture features in the first target picture training set, and if the class to which the picture features belong is judged to be the first class, selecting a corresponding five-level comparison set to compare the pictures corresponding to the first class.

Specifically, the device identifies picture features in the first target picture training set, and selects a corresponding five-level comparison set to compare pictures corresponding to the first class if the class to which the picture features belong is judged to be the first class. The first category can be understood as that the picture features of the picture to be identified are not easily distinguished from the picture features containing abnormal features, a corresponding comparison picture in the five-level comparison set needs to be selected for further comparison, and the first category can include a first target picture with a gastric angle structure overall (namely the first target picture with the gastric angle structure overall and not easily distinguished from the picture features containing the abnormal features), a picture containing the abnormal features related to the gastric angle structure is selected as a comparison picture, and the first target picture with the gastric angle structure overall is compared; an abnormal feature associated with the structure of the stomach horn may be the presence of polyps at a site in the stomach horn; if the comparison result is judged to be consistent, splitting a first target picture with the stomach corner structure to a picture containing abnormal features related to the stomach corner structure; if the comparison result is judged to be inconsistent, splitting a first target picture with a stomach corner structure to the first target picture training set; if the comparison result is consistent, the picture characteristic of the picture to be recognized (the first target picture with the stomach corner structure in the whole situation) is the abnormal characteristic, and therefore, the part of the picture is led into the picture containing the abnormal characteristic related to the stomach corner structure; if the comparison result is inconsistent, the picture characteristics of the picture to be recognized are not abnormal characteristics, so that the part of the picture is kept in the first target picture training set. Each subdivision category (e.g. half-moon gastric angle structure picture) in the first target picture with gastric angle structure globally can be compared with the corresponding comparison picture.

Further, the first category further includes a first target picture of the texture structure with the global dense stripe feature (that is, the first target picture of the texture structure with the global dense stripe feature, which is indistinguishable from the picture feature including the abnormal feature, is specified); correspondingly, the method further comprises the following steps: selecting a picture containing abnormal features related to the texture structure with the dense stripe features as a comparison picture, and comparing a first target picture of the texture structure with the dense stripe features; an abnormal feature associated with the texture structure of a dense streak feature may be the presence of polyps in a certain texture; if the comparison result is judged to be consistent, splitting a first target picture of the texture structure with the dense stripe feature into pictures containing abnormal features related to the texture structure with the dense stripe feature; and if the comparison result is judged to be inconsistent, splitting a first target picture of the texture structure with the dense fringe characteristics into the first target picture training set. If the comparison result is consistent, the picture characteristic of the picture to be identified (the first target picture of the texture structure with the dense stripe characteristic in the whole situation) is the abnormal characteristic, so that the part of the picture is led into the picture containing the abnormal characteristic related to the texture structure with the dense stripe characteristic; if the comparison result is inconsistent, the picture characteristics of the picture to be recognized are not abnormal characteristics, so that the part of the picture is kept in the first target picture training set. Each subdivision category (e.g., dense coarse stripe picture) in the first target picture of the global texture with dense stripe feature can be compared with the corresponding comparison picture.

Further, the first target picture of the global texture structure with the dense stripe feature further comprises a large intestine-shaped texture structure picture; large intestine-like texture: the picture shows a long-range shooting structure like a pig large intestine, which is medically called a stomach and small fovea structure, commonly called a pig large intestine-shaped structure, and the number of the pictures accounts for about 0.8 percent. Correspondingly, the method further comprises the following steps: selecting a picture containing abnormal features related to the large intestine-shaped texture structure as a comparison picture, and comparing the large intestine-shaped texture structure picture; an abnormal feature associated with a large intestine-like texture structure may be the presence of polyps in a certain large intestine-like texture; if the comparison result is judged to be consistent, splitting the large intestine-shaped texture structure picture into a picture containing abnormal features related to the large intestine-shaped texture structure; and if the comparison result is judged to be inconsistent, splitting the large intestine-shaped texture structure picture into the first target picture training set. If the comparison result is consistent, the picture characteristic of the picture to be identified (the large intestine-shaped texture structure picture) is the abnormal characteristic, and therefore, the part of the picture is led into the picture containing the abnormal characteristic related to the large intestine-shaped texture structure; if the comparison result is inconsistent, the picture characteristics of the picture to be recognized are not abnormal characteristics, so that the part of the picture is kept in the first target picture training set.

The large intestine-like texture picture includes:

a sausage-shaped structure picture, a large intestine-shaped structure picture, a small intestine-shaped structure and a hole-shaped structure superposition picture; the sausage-shaped structure picture comprises two to three thick stripes, the drift diameter of each thick stripe is larger than a first preset drift diameter numerical value, and the arrangement of each thick stripe is bent or straight; the large intestine-shaped structure chart sheet comprises four to five channel-shaped structures; the ditch migration structure contained in the small intestine-shaped structure picture is shot through a long-range view; the small intestine-shaped structure and the hole-shaped structure superposed pictures are superposed with a hole-shaped structure at the center of the small intestine-shaped structure pictures, and each subdivision classification (such as sausage-shaped structure pictures) in the large intestine-shaped texture structure pictures can be contrasted with the corresponding contrast pictures respectively.

Further, the first target picture of the global texture structure with the dense stripe feature further comprises a blood vessel-shaped texture picture; vascular texture: the picture looks like a "green-tendon" vein, actually a normal linear bulge of the inner wall of the stomach. Such raised stripes are also relatively sparse compared to dense textures, with a picture count of about 9%. Correspondingly, the method further comprises the following steps: selecting a picture containing abnormal features related to the vascular texture structure as a comparison picture, and comparing the vascular texture picture; an abnormal feature associated with a vascular texture structure may be the presence of polyps in a particular vascular texture; if the comparison result is judged to be consistent, splitting the vascular texture picture into a picture containing abnormal features related to the vascular texture structure; and splitting the vascular texture picture into the first target picture training set if the comparison result is judged to be inconsistent. If the comparison result is consistent, the picture characteristic of the picture to be identified (the vascular texture picture) is the abnormal characteristic, and therefore, the picture is led into the picture containing the abnormal characteristic related to the vascular texture structure; if the comparison result is inconsistent, the picture characteristics of the picture to be recognized are not abnormal characteristics, so that the part of the picture is kept in the first target picture training set.

The vascular texture picture comprises:

a shadow blood vessel structure picture, a blood vessel flat structure picture, a blood vessel superficial structure picture and a thick blood vessel structure picture; the shadow blood vessel structure picture is a half-moon structure overlapped with blood vessel-shaped convex stripes, the half-moon structure is a partial structure corresponding to a bright part in the outer surface of the shot, the partial structure is a half-moon-shaped structure, the drift diameter of the blood vessel-shaped convex stripes is smaller than a second preset drift diameter numerical value, the second preset drift diameter numerical value is smaller than the first preset drift diameter numerical value, and the convex height of the blood vessel-shaped convex stripes in the direction of the outer surface of the shot is larger than a preset height numerical value; the blood vessel flat structure picture is a full moon structure, the blood vessel-shaped convex stripes are superposed, and no other protrusions are left except the blood vessel-shaped convex stripes; the blood vessel superficial structure picture is a shadow blood vessel structure picture with the bulge height smaller than a preset height value, and the bulge height is the bulge height of the blood vessel-shaped bulge stripes along the direction of the outer surface of the shot; thick vascular structure picture is greater than for the latus rectum the second presets latus rectum numerical value, and is less than the shadow vascular structure picture of first preset latus rectum numerical value, the latus rectum is the latus rectum of the protruding stripe of angioid, and every subdivision classification (for example water cover hole structure picture) in the picture of angioid texture can all contrast with corresponding contrast picture respectively.

Further, the first category further includes a first target picture of a texture structure with global water-covering texture features (that is, a first target picture of a texture structure with global water-covering texture features, which is indistinguishable from picture features including abnormal features, is specified); correspondingly, the method further comprises the following steps: selecting a picture containing abnormal features related to the texture structure of the water covering texture features as a comparison picture, and comparing a first target picture of the texture structure with the water covering texture features on the whole; an abnormal feature associated with the texture of a water-covered texture feature may be the presence of polyps in a certain water-covered texture; if the comparison result is judged to be consistent, splitting a first target picture of the texture structure with the water covering texture into pictures containing abnormal features related to the texture structure with the water covering texture; and if the comparison result is judged to be inconsistent, splitting a first target picture of the texture structure with the water covering texture into the first target picture training set. If the comparison result is consistent, the picture characteristic of the picture to be identified (the first target picture of the texture structure with the water covering texture on the whole) is indicated as an abnormal characteristic, and therefore, the part of the picture is led into the picture containing the abnormal characteristic related to the texture structure with the water covering texture; if the comparison result is inconsistent, the picture characteristics of the picture to be recognized are not abnormal characteristics, so that the part of the picture is kept in the first target picture training set. Each subdivision category (e.g. water-covering hole structure picture) in the first target picture of the texture structure with water-covering texture features can be compared with the corresponding comparison picture respectively.

It should be noted that: the five-level comparison set includes the above-mentioned picture including the abnormal feature related to the gastric horn structure, the picture including the abnormal feature related to the texture structure of the dense stripe feature, the picture including the abnormal feature related to the texture structure of the water-covered texture, the picture including the abnormal feature related to the large intestine-shaped texture structure, the picture including the abnormal feature related to the blood vessel-shaped texture structure, and the like, and the corresponding five-level comparison set is selected to compare the pictures corresponding to the first category, which can be understood as: if the first category is a first target picture with a stomach corner structure in the whole situation, the picture corresponding to the first target picture with the stomach corner structure in the whole situation is a picture containing abnormal features related to the stomach corner structure; if the first type is a first target picture of the overall texture structure with the dense stripe feature, a picture corresponding to the first target picture of the overall texture structure with the dense stripe feature is a picture containing abnormal features related to the texture structure with the dense stripe feature; if the first category is a first target picture of the texture structure with the global water covering texture, a picture corresponding to the first target picture of the texture structure with the global water covering texture is a picture containing abnormal features related to the texture structure with the water covering texture; if the first category is a large intestine-shaped texture structure picture, the picture corresponding to the large intestine-shaped texture structure picture is a picture containing abnormal features related to the large intestine-shaped texture structure; if the first category is the vein-like texture picture, the picture corresponding to the vein-like texture picture is a picture including abnormal features related to the vein-like texture picture.

S103: and splitting the first target picture training set according to the comparison result.

Specifically, the device splits the first target picture training set according to the comparison result. Reference is made to the above description and no further description is made.

According to the splitting method of the training set provided by the embodiment of the invention, if the class to which the picture features in the first target picture training set based on the global structural features belong is judged to be the first class, the five-level comparison set is selected to compare the pictures corresponding to the first class, and the first target picture training set is split according to the comparison result, so that the splitting reasonability of the training set can be improved.

On the basis of the above embodiment, the first category includes a first target picture with a gastric angle structure globally; correspondingly, the method comprises the following steps:

and selecting a picture containing abnormal features related to the gastric angle structure as a comparison picture, and comparing the first target picture with the gastric angle structure on the whole.

Specifically, the device selects a picture containing abnormal features related to the gastric angle structure as a comparison picture, and compares the first target picture with the gastric angle structure in the whole situation. Reference may be made to the above embodiments, which are not described in detail.

If the comparison result is judged to be consistent, splitting the first target picture with the stomach corner structure into pictures containing abnormal features related to the stomach corner structure.

Specifically, if the comparison result is judged to be consistent, the device splits the first target picture with the gastric angle structure to the picture containing the abnormal features related to the gastric angle structure. Reference may be made to the above embodiments, which are not described in detail.

And if the comparison result is judged to be inconsistent, splitting the first target picture with the stomach corner structure to the first target picture training set.

Specifically, if the device judges that the comparison result is inconsistent, the device splits the first target picture with the gastric angle structure to the first target picture training set. Reference may be made to the above embodiments, which are not described in detail.

According to the splitting method of the training set provided by the embodiment of the invention, the picture containing the abnormal features related to the gastric angle structure is compared with the first target picture with the gastric angle structure in the whole situation, so that the splitting reasonability of the training set can be further improved.

On the basis of the embodiment, the first category further includes a first target picture of a texture structure with dense stripe features in the whole; correspondingly, the method further comprises the following steps:

and selecting a picture containing abnormal features related to the texture structure with the dense stripe features as a comparison picture, and comparing a first target picture of the texture structure with the dense stripe features on the whole.

Specifically, the device selects a picture containing abnormal features related to the texture structure with the dense stripe features as a comparison picture, and compares the first target picture of the texture structure with the dense stripe features globally. Reference may be made to the above embodiments, which are not described in detail.

If the comparison result is judged to be consistent, splitting the first target picture of the texture structure with the dense stripe feature into pictures containing abnormal features related to the texture structure with the dense stripe feature.

Specifically, if the comparison result is judged to be consistent, the device splits the first target picture of the texture structure with the dense stripe feature to the picture containing the abnormal feature related to the texture structure with the dense stripe feature. Reference may be made to the above embodiments, which are not described in detail.

And if the comparison result is judged to be inconsistent, splitting a first target picture of the texture structure with the dense fringe characteristics into the first target picture training set.

Specifically, if the device judges that the comparison result is inconsistent, the device splits the first target picture of the texture structure with the dense stripe feature to the first target picture training set. Reference may be made to the above embodiments, which are not described in detail.

According to the splitting method of the training set provided by the embodiment of the invention, the picture containing the abnormal features related to the texture structure with the dense stripe features is compared with the first target picture of the texture structure with the dense stripe features, so that the splitting reasonability of the training set can be further improved.

On the basis of the above embodiment, the first target picture of the texture structure with the global dense stripe feature further includes a large intestine-shaped texture structure picture; correspondingly, the method further comprises the following steps:

and selecting a picture containing abnormal features related to the large intestine-shaped texture structure as a comparison picture, and comparing the large intestine-shaped texture structure picture.

Specifically, the device selects a picture containing abnormal features related to the large intestine-shaped texture structure as a comparison picture, and compares the large intestine-shaped texture structure picture. Reference may be made to the above embodiments, which are not described in detail.

And if the comparison result is judged to be consistent, splitting the large intestine-shaped texture structure picture into pictures containing abnormal features related to the large intestine-shaped texture structure.

Specifically, if the device determines that the comparison result is consistent, the device splits the large intestine-shaped texture structure picture into pictures containing abnormal features related to the large intestine-shaped texture structure. Reference may be made to the above embodiments, which are not described in detail.

And if the comparison result is judged to be inconsistent, splitting the large intestine-shaped texture structure picture into the first target picture training set.

Specifically, if the device judges that the comparison result is inconsistent, the device splits the large intestine-shaped texture structure picture into the first target picture training set. Reference may be made to the above embodiments, which are not described in detail.

According to the splitting method of the training set provided by the embodiment of the invention, the picture containing the abnormal features related to the large intestine-shaped texture structure is compared with the large intestine-shaped texture structure picture, so that the splitting reasonability of the training set can be further improved.

On the basis of the above embodiment, the first target picture of the global texture structure with the dense stripe feature further includes a blood vessel-shaped texture picture; correspondingly, the method further comprises the following steps:

and selecting a picture containing abnormal features related to the vascular texture structure as a comparison picture, and comparing the vascular texture picture.

Specifically, the device selects a picture containing abnormal features related to the vascular texture structure as a comparison picture, and compares the vascular texture picture. Reference may be made to the above embodiments, which are not described in detail.

And if the comparison result is judged to be consistent, splitting the vascular texture picture into pictures containing abnormal features related to the vascular texture structure.

Specifically, if the device determines that the comparison result is consistent, the device splits the vascular texture picture into pictures containing abnormal features related to the vascular texture structure. Reference may be made to the above embodiments, which are not described in detail.

And splitting the vascular texture picture into the first target picture training set if the comparison result is judged to be inconsistent.

Specifically, if the device judges that the comparison result is inconsistent, the device splits the vascular texture picture into the first target picture training set. Reference may be made to the above embodiments, which are not described in detail.

According to the splitting method of the training set provided by the embodiment of the invention, the picture containing the abnormal features related to the vascular texture structure is compared with the vascular texture picture, so that the splitting reasonability of the training set can be further improved.

On the basis of the above embodiment, after the step of splitting the first target picture training set according to the comparison result, the method further includes:

and selecting the abnormal characteristic total set as a comparison picture total set, and comparing all pictures in the first target picture training set again.

Specifically, the device selects the total abnormal feature set as a total comparison picture set, and compares all pictures in the first target picture training set again. The abnormal feature collection can be understood as a collection containing all abnormal features, and the comparison is performed for the following purposes: and the abnormal characteristic pictures are prevented from being attracted to the first target picture training set.

The splitting method of the training set provided by the embodiment of the invention can further improve the splitting reasonability of the training set.

On the basis of the above embodiment, after the step of performing the cross-reference again on all the pictures in the first target picture training set, the method further includes:

selecting a first-level contrast set corresponding to a class-out-of-domain training set, a second-level contrast set corresponding to a class-two-out-of-domain training set, a third-level contrast set corresponding to a first target picture training set based on local structural features and a fourth-level contrast set corresponding to a first target picture training set of a cavernous structure, and performing re-contrast on the first target picture training set of the global structural features after re-contrast by adopting the first-level contrast set, the second-level contrast set, the third-level contrast set and the fourth-level contrast set; the first class of out-of-domain training set is a training set determined based on shooting defects of an original picture and a shooting part irrelevant to a target part to be detected, the second class of out-of-domain training set is a training set determined based on an original picture without medical judgment value, an original picture attached with a covering object, an original picture containing a digestive residue object, and the first target picture training set based on local structural features is a training set determined based on a first target picture with local shape change and/or color tone change; the first target picture is a picture of which the outer surface of the shot does not contain abnormal features; the training set of the first target picture of the hole-shaped structure is determined by taking an original picture of which the outer surface does not contain abnormal features and contains the hole-shaped structure.

Specifically, the device selects a first-level contrast set corresponding to a class-out-of-domain training set, a second-level contrast set corresponding to a class-two-level out-of-domain training set, a third-level contrast set corresponding to a first target picture training set based on local structural features, and a fourth-level contrast set corresponding to a first target picture training set of a cavernous structure, and performs re-contrast (sequential contrast in order) on the first target picture training set of the global structural features after re-contrast by using the first-level contrast set, the second-level contrast set, the third-level contrast set, and the fourth-level contrast set; the first class of out-of-domain training set is a training set determined based on shooting defects of an original picture and a shooting part irrelevant to a target part to be detected, the second class of out-of-domain training set is a training set determined based on an original picture without medical judgment value, an original picture attached with a covering object, an original picture containing a digestive residue object, and the first target picture training set based on local structural features is a training set determined based on a first target picture with local shape change and/or color tone change; the first target picture is a picture of which the outer surface of the shot does not contain abnormal features; the training set of the first target picture of the hole-shaped structure is determined by taking an original picture of which the outer surface does not contain abnormal features and contains the hole-shaped structure. The photographing defect may include:

full exposure pictures, full black pictures, half exposure pictures, local exposure pictures, structure fuzzy pictures and detail fuzzy pictures; the full exposure picture, the half exposure picture and the local exposure picture are distinguished according to the area of a picture exposure area; the structure fuzzy picture and the detail fuzzy picture are distinguished according to the area of the picture fuzzy area.

The photographing part may include:

pictures taken before the entrance of the capsule endoscope, pictures taken in the esophagus, pictures of the oral cavity and pictures of the intestinal tract.

The original picture without medical judgment value comprises:

homogenizing the whole picture and the waterline picture; the outer surface of the shot in the homogeneous whole picture is flat and smooth, has no texture and is uniform in color; the waterline picture has an interface line of air and water.

The original picture with the covering attached comprises:

full-overlay pictures, half-overlay pictures; the full-cover picture and the half-cover picture are distinguished according to the size of the area occupied by the cover.

The original picture with the overlay attached further comprises:

bubble cover pictures, cobweb cover pictures; the outer surface of a shot in the bubble covering picture is covered by bubbles and has a light reflection phenomenon; the outer surface of the picture in the webbed covering picture is covered by the webbed covering.

The first target picture with shape change locally may include:

an edge defect picture, a fishtail structure picture, a fracture structure picture and a staggered structure picture; wherein, the edge part of the edge defect picture has a incomplete hole-shaped or semi-radial structure; the edge part of the fishtail structure picture is provided with a fold structure; the crack structure picture comprises two textures which penetrate through the whole picture from inside; the interlaced structure picture includes a plurality of curves that meet together.

The first target picture with local tone variation may include:

the picture comprises a full moon-shaped structure picture and a shadow structure picture, wherein the full moon-shaped structure picture is a homogeneous whole picture with local tone variation; the outer surface of the shot in the homogeneous whole picture is flat and smooth, has no texture and is uniform in color; the shadow structure picture is gradually changed in light and shade, and the areas of the bright area and the dark area respectively account for half.

The first target picture with shape change and color tone change locally may include:

the image processing method comprises the following steps of obtaining a linear structure picture, a cutting structure picture, a linear structure picture and a non-intersecting structure picture, wherein the linear structure picture comprises less than two prismatic structures which do not penetrate through a whole image and also comprises a partial dark area; the cutting structure picture comprises at least one prismatic structure penetrating through the whole picture and a partial dark area; the linear structure picture comprises two thick prismatic structures which penetrate through the whole picture from inside and also comprises a partial dark area; the non-intersecting structure picture comprises a plurality of groups of non-intersecting thick and strong prismatic structures and also comprises a partial dark area. The thick and strong arris-shaped structure can be understood that the arris-shaped drift diameter is larger than a preset drift diameter numerical value, and the preset drift diameter numerical value can be set independently according to actual conditions.

The first target picture of the hole-shaped structure comprises:

big hole structure pictures, small hole structure pictures, flower-shaped structure pictures and radial structure pictures; wherein, the ratio of the area of the hole-shaped opening of the hole-shaped structure in the big hole structure picture to the area of the original picture is larger than a first preset ratio threshold, the ratio of the area of the hole-shaped opening of the hole-shaped structure in the small hole structure picture to the area of the original picture is smaller than the first preset ratio threshold, and the flower-shaped structure picture contains flower-shaped and non-hollow openings; the radial structure picture comprises a plurality of lines converging towards the center, and all the lines converge into a point at the center and/or form an opening.

The primary control set comprises: the image features of the image to be recognized (i.e., the image defects based on the original image and the image features unrelated to the target region to be detected) are not easily distinguished from the image features including the abnormal features, such as a blur degree image (i.e., the image has a part of blur and a part of clear image, and the part of clear content can be used as a reference image for comparison), and the like.

The secondary control set included: the image characteristics of the image to be recognized (i.e. based on non-medical judgment value, attached with a covering, containing a digestive residue) and the image characteristics containing abnormal characteristics are not easily distinguished, such as the whole image containing abnormal characteristics, the image containing abnormal characteristics accompanied with the covering, the image containing abnormal characteristics in the background of the waterline, and the like; further comprising: the image features of the image to be recognized are not easily distinguished from the image features not containing the abnormal features, for example, the reference image is a stomach image, and some intestinal images are similar to the stomach image and are not easily distinguished.

The tertiary control set included: the image characteristics of the image to be recognized (i.e. the first target image with local shape change and color tone change) are not easily distinguished from the image characteristics containing abnormal characteristics, such as an image containing abnormal characteristics related to shape change, an image containing abnormal characteristics related to color tone change, a whole image, and the like; the whole picture is a picture containing abnormal features related to both shape change and color tone change.

Included in the four-level control sets were: the picture features of the picture to be recognized (i.e. the first target picture containing the hole-like structure) and the picture features containing the abnormal features are not easily distinguished from each other, such as a picture containing the abnormal features related to the hole-like structure, a picture containing the abnormal features related to the flower-like structure, a picture containing the abnormal features related to the radial structure, and the like.

It should be noted that: in the splitting process of the first class out-of-domain training set, the second class out-of-domain training set, the first target picture training set based on the local structural features and the first target picture training set of the hole-shaped structure, a small part of pictures which are left (namely, the pictures which are supposed to be kept in the first class out-of-domain training set or the second class out-of-domain training set or the first target picture training set based on the local structural features or the first target picture training set based on the hole-shaped structure need to be split in the first target picture training set based on the global structural features), can be collected in the first target picture training set based on the global structural features through a selectable first-level contrast set, a second-level contrast set, a third-level contrast set and a fourth-level contrast set, and the first-level contrast set, the second-level contrast set, the third-level contrast set and the fourth-level contrast set are not too large, an excessive number of pictures of the first target picture training set that result in many global structural features are guided through.

Splitting pictures in a first target picture training set with global structural characteristics consistent with the comparison result of the target training set into corresponding target training sets; the target training set is one of the class-one out-of-domain training set, the class-two out-of-domain training set, the first target picture training set based on local structural features, and the first target picture training set of the cavernous structure.

Specifically, the device splits pictures in a first target picture training set with global structural features consistent with the comparison result of the target training set into corresponding target training sets; the target training set is one of the class-one out-of-domain training set, the class-two out-of-domain training set, the first target picture training set based on local structural features, and the first target picture training set of the cavernous structure. Referring to the above example, the small part of the missed pictures is split into a class-outside training set or a class-two-outside training set, or a first target picture training set based on local structural features or a first target picture training set of a cavernous structure.

And splitting pictures in the first target picture training set of the global structural features which are inconsistent with the comparison result of the target training set into the first target picture training set of the global structural features.

Specifically, the device splits pictures in a first target picture training set of global structural features which are inconsistent with the comparison result of the target training set into the first target picture training set of global structural features. Referring to the above example, if the missing picture is not the small part, the missing picture is retained in the first target picture training set of the global structural feature.

Fig. 2 is a schematic structural diagram of an embodiment of a splitting device of a training set according to the present invention, and as shown in fig. 2, an embodiment of the present invention provides a splitting device of a training set, which includes an obtaining unit 201, a selecting unit 202, and a splitting unit 203, where:

the obtaining unit 201 is configured to obtain a training set used for training a preset model; the training set comprises a first target picture training set based on global structural features; the first target picture training set is a training set determined based on a first target picture with a gastric angle structure and/or a texture structure with special features globally; the first target picture is a picture of which the outer surface of the shot does not contain abnormal features; the abnormal features comprise raised features and/or designated color features; the special features comprise dense stripe features; the number of the stripes in the compact stripe characteristics is greater than a preset number; the selecting unit 202 is configured to identify picture features in the first target picture training set, and if it is determined that the category to which the picture features belong is the first category, select a corresponding five-level comparison set to compare pictures corresponding to the first category; the splitting unit 203 is configured to split the first target picture training set according to the comparison result.

Specifically, the obtaining unit 201 is configured to obtain a training set used for training a preset model; the training set comprises a first target picture training set based on global structural features; the first target picture training set is a training set determined based on a first target picture with a gastric angle structure and/or a texture structure with special features globally; the first target picture is a picture of which the outer surface of the shot does not contain abnormal features; the abnormal features comprise raised features and/or designated color features; the special features comprise dense stripe features; the number of the stripes in the compact stripe characteristics is greater than a preset number; the selecting unit 202 is configured to identify picture features in the first target picture training set, and if it is determined that the category to which the picture features belong is the first category, select a corresponding five-level comparison set to compare pictures corresponding to the first category; the splitting unit 203 is configured to split the first target picture training set according to the comparison result.

According to the splitting device for the training set provided by the embodiment of the invention, if the class to which the picture features in the first target picture training set based on the global structural features belong is judged to be the first class, the five-level comparison set is selected to compare the pictures corresponding to the first class, and the first target picture training set is split according to the comparison result, so that the splitting rationality of the training set can be improved.

The apparatus for splitting a training set according to the embodiments of the present invention may be specifically configured to execute the processing flows of the above method embodiments, and the functions of the apparatus are not described herein again, and refer to the detailed description of the above method embodiments.

Fig. 3 is a schematic structural diagram of an electronic device according to an embodiment of the present invention, and as shown in fig. 3, the electronic device includes: a processor (processor)301, a memory (memory)302, and a bus 303;

the processor 301 and the memory 302 complete communication with each other through a bus 303;

the processor 301 is configured to call program instructions in the memory 302 to perform the methods provided by the above-mentioned method embodiments, including: acquiring a training set for training a preset model; the training set comprises a first target picture training set based on global structural features; the first target picture training set is a training set determined based on a first target picture with a gastric angle structure and/or a texture structure with special features globally; the first target picture is a picture of which the outer surface of the shot does not contain abnormal features; the abnormal features comprise raised features and/or designated color features; the special features comprise dense stripe features; the number of the stripes in the compact stripe characteristics is greater than a preset number; identifying picture features in the first target picture training set, and if the class to which the picture features belong is judged to be the first class, selecting a corresponding five-level comparison set to compare the pictures corresponding to the first class; and splitting the first target picture training set according to the comparison result.

The present embodiment discloses a computer program product comprising a computer program stored on a non-transitory computer readable storage medium, the computer program comprising program instructions which, when executed by a computer, enable the computer to perform the method provided by the above-mentioned method embodiments, for example, comprising: acquiring a training set for training a preset model; the training set comprises a first target picture training set based on global structural features; the first target picture training set is a training set determined based on a first target picture with a gastric angle structure and/or a texture structure with special features globally; the first target picture is a picture of which the outer surface of the shot does not contain abnormal features; the abnormal features comprise raised features and/or designated color features; the special features comprise dense stripe features; the number of the stripes in the compact stripe characteristics is greater than a preset number; identifying picture features in the first target picture training set, and if the class to which the picture features belong is judged to be the first class, selecting a corresponding five-level comparison set to compare the pictures corresponding to the first class; and splitting the first target picture training set according to the comparison result.

The present embodiments provide a non-transitory computer-readable storage medium storing computer instructions that cause the computer to perform the methods provided by the above method embodiments, for example, including: acquiring a training set for training a preset model; the training set comprises a first target picture training set based on global structural features; the first target picture training set is a training set determined based on a first target picture with a gastric angle structure and/or a texture structure with special features globally; the first target picture is a picture of which the outer surface of the shot does not contain abnormal features; the abnormal features comprise raised features and/or designated color features; the special features comprise dense stripe features; the number of the stripes in the compact stripe characteristics is greater than a preset number; identifying picture features in the first target picture training set, and if the class to which the picture features belong is judged to be the first class, selecting a corresponding five-level comparison set to compare the pictures corresponding to the first class; and splitting the first target picture training set according to the comparison result.

Those of ordinary skill in the art will understand that: all or part of the steps for implementing the method embodiments may be implemented by hardware related to program instructions, and the program may be stored in a computer readable storage medium, and when executed, the program performs the steps including the method embodiments; and the aforementioned storage medium includes: various media that can store program codes, such as ROM, RAM, magnetic or optical disks.

The above-described embodiments of the apparatus are merely illustrative, and the units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of the present embodiment. One of ordinary skill in the art can understand and implement it without inventive effort.

Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented by software plus a necessary general hardware platform, and certainly can also be implemented by hardware. With this understanding in mind, the above-described technical solutions may be embodied in the form of a software product, which can be stored in a computer-readable storage medium such as ROM/RAM, magnetic disk, optical disk, etc., and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the methods described in the embodiments or some parts of the embodiments.

Finally, it should be noted that: the above examples are only intended to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.

Claims

1. A method for splitting a training set is characterized by comprising the following steps:

splitting the first target picture training set according to a comparison result;

after the step of splitting the first target picture training set according to the comparison result, the method further comprises:

selecting an abnormal feature total set as a comparison picture total set, and comparing all pictures in the first target picture training set again;

the first category is that the picture features of the picture to be identified are not easily distinguished from the picture features containing abnormal features, and corresponding comparison pictures in the five-level comparison set are selected for comparison;

the five-level control set includes pictures containing abnormal features related to gastric horn structures, pictures containing abnormal features related to texture structures with dense streak features, pictures containing abnormal features related to texture structures with water-covered texture, pictures containing abnormal features related to large intestine-shaped texture structures, and pictures containing abnormal features related to blood vessel-shaped texture structures.

2. The method of claim 1, wherein the first category includes a first target picture with a gastric corner structure globally; correspondingly, the method comprises the following steps:

selecting a picture containing abnormal features related to the gastric angle structure as a comparison picture, and comparing a first target picture with the gastric angle structure on the whole;

if the comparison result is judged to be consistent, splitting a first target picture with the stomach corner structure to a picture containing abnormal features related to the stomach corner structure;

3. The method of claim 2, wherein the first category further comprises a first target picture of a texture structure with dense stripe features globally; correspondingly, the method further comprises the following steps:

selecting a picture containing abnormal features related to the texture structure with the dense stripe features as a comparison picture, and comparing a first target picture of the texture structure with the dense stripe features;

if the comparison result is judged to be consistent, splitting a first target picture of the texture structure with the dense stripe feature into pictures containing abnormal features related to the texture structure with the dense stripe feature;

4. The method according to claim 3, wherein the first target picture of the global texture with dense stripe features further comprises a large intestine-like texture picture; correspondingly, the method further comprises the following steps:

selecting a picture containing abnormal features related to the large intestine-shaped texture structure as a comparison picture, and comparing the large intestine-shaped texture structure picture;

if the comparison result is judged to be consistent, splitting the large intestine-shaped texture structure picture into a picture containing abnormal features related to the large intestine-shaped texture structure;

5. The method according to claim 3, wherein the first target picture of the global texture with dense stripe features further comprises a vessel-like texture picture; correspondingly, the method further comprises the following steps:

selecting a picture containing abnormal features related to the vascular texture structure as a comparison picture, and comparing the vascular texture picture;

if the comparison result is judged to be consistent, splitting the vascular texture picture into a picture containing abnormal features related to the vascular texture structure;

6. The method of claim 5, wherein after the step of cross-referencing all the pictures in the first target picture training set, the method further comprises:

selecting a first-level contrast set corresponding to a class-out-of-domain training set, a second-level contrast set corresponding to a class-two-out-of-domain training set, a third-level contrast set corresponding to a first target picture training set based on local structural features and a fourth-level contrast set corresponding to a first target picture training set of a cavernous structure, and performing re-contrast on the first target picture training set of the global structural features after re-contrast by adopting the first-level contrast set, the second-level contrast set, the third-level contrast set and the fourth-level contrast set; the first class of out-of-domain training set is a training set determined based on shooting defects of an original picture and a shooting part irrelevant to a target part to be detected, the second class of out-of-domain training set is a training set determined based on an original picture without medical judgment value, an original picture attached with a covering object, an original picture containing a digestive residue object, and the first target picture training set based on local structural features is a training set determined based on a first target picture with local shape change and/or color tone change; the first target picture is a picture of which the outer surface of the shot does not contain abnormal features; the first target picture training set of the hole-shaped structure is a training set determined by an original picture of which the outer surface of a shot does not contain abnormal features and contains the hole-shaped structure;

splitting pictures in a first target picture training set with global structural characteristics consistent with the comparison result of the target training set into corresponding target training sets; the target training set is one of the class-one out-of-domain training set, the class-two out-of-domain training set, the first target picture training set based on local structural features and the first target picture training set of the cavernous structure;

7. A splitting device of a training set is characterized by comprising:

the splitting unit is used for splitting the first target picture training set according to the comparison result;

the secondary selection unit selects the abnormal characteristic total set as a comparison picture total set, and compares all pictures in the first target picture training set again;

8. An electronic device, comprising: a processor, a memory, and a bus, wherein,

the processor and the memory are communicated with each other through the bus;

the memory stores program instructions executable by the processor, the processor invoking the program instructions to perform the method of any of claims 1 to 6.

9. A non-transitory computer-readable storage medium storing computer instructions that cause a computer to perform the method of any one of claims 1 to 6.