CN113806613A

CN113806613A - Training image set generation method and device, computer equipment and storage medium

Info

Publication number: CN113806613A
Application number: CN202111153778.2A
Authority: CN
Inventors: 高预皓; 彭晶
Original assignee: Ping An Life Insurance Company of China Ltd
Current assignee: Ping An Life Insurance Company of China Ltd
Priority date: 2021-09-29
Filing date: 2021-09-29
Publication date: 2021-12-17
Anticipated expiration: 2041-09-29
Also published as: CN113806613B

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a training image set generation method, a training image set generation device, training image set generation equipment and a training image set generation medium, wherein the training image set generation method comprises the following steps: acquiring a to-be-generated image set and a generation quantity; crawling historical tags similar to the target tags and the target description by using a text similarity technology, and determining the category to be migrated; acquiring a corresponding model to be migrated, identifying each image to be generated, acquiring a target area corresponding to each image to be generated, and recording the target area as an image to be processed; carrying out image enhancement processing based on factor enhancement proportion on each image to be processed to generate a training image; and associating each training image; and determining all images to be processed and training images as a training image set. Therefore, the invention realizes the automatic generation of the training image set based on the zero-label small image set to be generated, and improves the accuracy and efficiency of the generation of the training image set. The method is suitable for the field of artificial intelligence, and can further promote the construction of smart cities.

Description

Training image set generation method and device, computer equipment and storage medium

Technical Field

The invention relates to the technical field of data acquisition of data processing, in particular to a training image set generation method and device, computer equipment and a storage medium.

Background

The artificial intelligence technology is a comprehensive subject and relates to the field of extensive technology, namely the technology of a hardware level and the technology of a software level. The artificial intelligence software technology mainly comprises a computer vision technology, a voice processing technology, a natural language processing technology, machine learning/deep learning and the like. Among them, Computer Vision technology (CV) is a science that studies how to "see" a machine, and generally includes technologies such as image processing, image Recognition, image semantic understanding, image retrieval, and Optical Character Recognition (OCR).

As the artificial intelligence technology is gradually mature, the image recognition technology is more and more widely applied to daily life. In order to obtain an image recognition model with higher recognition accuracy, the image recognition model needs to be trained through a large number of labeled samples, and in the prior art, when a training sample is constructed, manual labeling such as manual input is usually adopted, so that not only is the labor cost consumed, but also the labeling efficiency of the sample is greatly reduced, and great difficulty is brought to model training.

Disclosure of Invention

The invention provides a training image set generation method, a training image set generation device, computer equipment and a storage medium, which can automatically generate a training image set based on a small number of image sets to be generated with zero labels, reduce the manual labeling time and the workload of training image collection, improve the labeling efficiency, save the investment cost and improve the accuracy and efficiency of generating the training image set.

A training image set generation method includes:

acquiring a to-be-generated image set and a generation quantity; the image set to be generated comprises a plurality of images to be generated; associating one of the images to be generated with a target label; one said object tag corresponding to one object description; the generated number is the number of training images for training generated by the image set to be generated;

crawling historical labels similar to the target labels and the target description in a historical label library by using a text similarity technology, and determining the category to be migrated according to all the crawled historical labels;

acquiring a model to be migrated corresponding to the category to be migrated from an identification model library, identifying each image to be generated through the model to be migrated to obtain a target area corresponding to each image to be generated, and recording the image to be generated marked with the target area as an image to be processed;

performing image enhancement processing based on factor enhancement proportion on each image to be processed to generate the training images with the generated number;

associating each training image with the target label associated with the image to be generated corresponding to each training image;

and determining all the images to be processed and all the associated training images as a training image set.

An apparatus for training image set generation, comprising:

the acquisition module is used for acquiring the image sets to be generated and the generation quantity; the image set to be generated comprises a plurality of images to be generated; associating one of the images to be generated with a target label; one said object tag corresponding to one object description; the generated number is the number of training images for training generated by the image set to be generated;

the crawling module is used for crawling historical tags similar to the target tags and the target description in a historical tag library by using a text similarity technology, and determining the category to be migrated according to all the crawled historical tags;

the migration module is used for acquiring a model to be migrated corresponding to the category to be migrated from an identification model library, identifying each image to be generated through the model to be migrated to obtain a target area corresponding to each image to be generated, and recording the image to be generated marked with the target area as an image to be processed;

the generating module is used for carrying out image enhancement processing based on factor enhancement proportion on each image to be processed to generate the training images with the generated number;

the association module is used for associating each training image with the target label associated with the image to be generated corresponding to each training image;

and the determining module is used for determining all the images to be processed and all the associated training images as a training image set.

A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, the processor implementing the steps of the training image set generation method when executing the computer program.

A computer-readable storage medium, in which a computer program is stored which, when being executed by a processor, carries out the steps of the above-mentioned training image set generation method.

The invention provides a training image set generation method, a training image set generation device, computer equipment and a storage medium, wherein the training image set generation method comprises the steps of obtaining an image set to be generated and a generation quantity; crawling historical labels similar to the target labels and the target description in a historical label library by using a text similarity technology, and determining the category to be migrated according to all the crawled historical labels; acquiring a model to be migrated corresponding to the category to be migrated from an identification model library, identifying each image to be generated through the model to be migrated to obtain a target area corresponding to each image to be generated, and recording the image to be generated marked with the target area as an image to be processed; performing image enhancement processing based on factor enhancement proportion on each image to be processed to generate the training images with the generated number; associating each training image with the target label associated with the image to be generated corresponding to each training image; all the images to be processed and all the associated training images are determined as training image sets, so that the automatic generation of the training image sets based on a small number of zero-labeled image sets to be generated is realized through a text similarity technology, a crawling technology and a transfer learning technology and by applying image enhancement processing, the manual labeling time and the workload of training image collection are reduced, the labeling efficiency is improved, the investment cost is saved, the accuracy and the efficiency of generating the training image sets are improved, and high-quality training images are provided for subsequent training.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the description of the embodiments of the present invention will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art that other drawings can be obtained according to these drawings without inventive labor.

FIG. 1 is a schematic diagram of an application environment of a training image set generation method according to an embodiment of the present invention;

FIG. 2 is a flowchart of a training image set generation method according to an embodiment of the present invention;

FIG. 3 is a flowchart illustrating a step S20 of a training image set generating method according to an embodiment of the present invention;

FIG. 4 is a flowchart illustrating a step S30 of a training image set generating method according to an embodiment of the present invention;

FIG. 5 is a flowchart illustrating a step S40 of a training image set generating method according to an embodiment of the present invention;

FIG. 6 is a flowchart of step S404 of a training image set generating method according to an embodiment of the present invention;

FIG. 7 is a schematic block diagram of an apparatus for generating training image sets according to an embodiment of the present invention;

FIG. 8 is a schematic diagram of a computer device in an embodiment of the invention.

Detailed Description

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, 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.

The training image set generation method provided by the invention can be applied to the application environment shown in fig. 1, wherein a client (computer equipment or terminal) communicates with a server through a network. The client (computer device or terminal) includes, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers, and portable wearable devices. The server may be an independent server, or may be a cloud server that provides basic cloud computing services such as a cloud service, a cloud database, cloud computing, a cloud function, cloud storage, a Network service, cloud communication, a middleware service, a domain name service, a security service, a Content Delivery Network (CDN), a big data and artificial intelligence platform, and the like.

In an embodiment, as shown in fig. 2, a training image set generating method is provided, and the technical solution thereof mainly includes the following steps S10-S60:

s10, acquiring a to-be-generated image set and a generation number; the image set to be generated comprises a plurality of images to be generated; associating one of the images to be generated with a target label; one said object tag corresponding to one object description; the generated number is the number of training images generated for training through the image set to be generated.

Understandably, the image set to be generated is a set of all the images to be generated, the images to be generated are historically collected images related to the target tags, the target tags are tags assigned to the content of the images to be generated by classification, such as tags for example, a target tag is a copy or non-copy, a human face or non-human face, a certificate or non-certificate, and the like, and the target description is text content describing related key features of the category of the target tag, for example: and aiming at the condition that the target label is the identity document, the target is described as a rectangular document containing a human face head portrait and 18-bit characters, and the generation quantity is the quantity of training images for training generated through the image set to be generated.

S20, crawling historical labels similar to the target labels and the target descriptions in a historical label library by using a text similarity technology, and determining the category to be migrated according to all the crawled historical labels.

Understandably, the crawling process is: firstly, crawling out relevant descriptions related to the target label by using a web crawler technology, collecting the crawled relevant descriptions to extract keywords and the occurrence times of the keywords, using the keywords as a web description, secondly, performing keyword aggregation on the web description and the target description, namely, weighting the keywords consistent with the web description in the target description, performing important comparison on important words in the target description through weighting, and finally, comparing the history description under each history label with the weighted target description in the history label library by using the text similarity technology, and determining the history label with the highest similarity value after comparison as the history label similar to the target label and the target description, for example: the target label is an identity card, the target description is a rectangular certificate containing a human face head portrait and 18-bit characters, the human face head portrait, the 18-bit characters and the rectangle are weighted, through the application of a text similarity technology, the historical description of the historical label, which is a photo frame, is an image of a rectangular photo frame containing a landscape of a person, is crawled from the historical label library, and the similarity value of the image and the weighted target description is the highest, so that the photo frame is determined to be the historical label similar to the identity card.

Wherein, all identifiable historical labels are stored in the historical label library, the historical labels are of categories which can be identified by corresponding trained models, one historical label corresponds to one historical description, the history description is description of relevant characteristics of the history label, the text similarity technology is technology for comparing similarity degree between two texts by using a text similarity algorithm, the text similarity algorithm is to perform word embedding (word embedding) conversion processing on two texts, an algorithm for performing equal weight conversion according to the weighted keywords in the process of word embedding conversion processing, performing similarity calculation on the two processed texts to obtain a similarity value between the two texts, the web crawler technology is a technology for automatically capturing a program or script of web information according to a certain rule, thereby acquiring required information.

In an embodiment, as shown in fig. 3, in the step S20, that is, the crawling, in the history tag library, history tags similar to the target tags and the target descriptions by using a text similarity technique, and determining the category to be migrated according to all the crawled history tags includes:

s201, crawling the network description matched with the target tag by using a web crawler technology.

Understandably, the web crawler technology is a technology for automatically capturing a program or a script of world wide web information according to a certain rule so as to obtain required information, crawling descriptions related to the target labeling category from the internet, summarizing and refining all the crawled related descriptions, extracting keywords by using a TF-IDF algorithm, and determining the occurrence frequency of each keyword of the extracted keywords as network description.

The TF-IDF algorithm is a weighting technique for information retrieval (information retrieval) and text mining (text mining), and is a statistical method for evaluating the importance degree of a word in a text, and the importance of a word or word increases in proportion to the number of times it appears in a document, but decreases in inverse proportion to the frequency of occurrence in a corpus.

S202, according to the network description, carrying out keyword weighting on the target description to obtain a focused description.

Understandably, words which are consistent with the keywords in the network description are found in the target description, and the found words are weighted according to the occurrence frequency of each keyword to obtain the focusing description.

S203, comparing the history description under each history label with the focusing description in the history label library by using a text similarity technology to obtain a similarity value corresponding to each history description.

Understandably, the text similarity technique is a technique for comparing similarity degrees between two texts by using a text similarity algorithm, the text similarity algorithm is an algorithm for performing word embedding (word embedding) conversion processing on the two texts, performing equal weight conversion according to keywords with weights in the word embedding conversion processing process, performing similarity calculation on the two processed texts to obtain a similarity value between the two texts, the word embedding (word embedding) conversion processing is also called word2vec, that is, converting a word (word) into a vector (vector) for representation, giving equal weights to the vectors after converting the keywords with weights in the conversion process, comparing the history description and the focus description under each history label by using the text similarity algorithm in the history label library, and adding equal weights to the vectors corresponding to the keywords with weights in the output similarity values, that is, the influence of the vector corresponding to the weighted keyword on the similarity value as a whole has a higher weight ratio than others.

S204, determining the history label corresponding to the history description corresponding to the maximum similarity value as the category to be migrated.

Understandably, the history label corresponding to the maximum similarity value is recorded as the category to be migrated, and the category to be migrated indicates the category to be migrated.

The invention realizes that the network description matched with the target label is crawled by applying a web crawler technology; according to the network description, carrying out keyword weighting on the target description to obtain a focused description; comparing the historical description and the focusing description under each historical label in the historical label library by using a text similarity technology to obtain a similarity value corresponding to each historical description; the history label corresponding to the history description corresponding to the maximum similarity value is determined as the category to be migrated, so that the category to be migrated is automatically identified from a target classification identification library by applying a web crawler technology, a keyword weighting technology and a text similarity technology, the cost of manual identification is reduced, the category to be migrated can be quickly and accurately found, and the identification accuracy and reliability are improved.

S30, obtaining models to be migrated corresponding to the categories to be migrated from an identification model library, identifying each image to be generated through the models to be migrated to obtain target areas corresponding to the images to be generated, and recording the images to be generated marked with the target areas as images to be processed.

Understandably, the identification model library stores all the various to-be-migrated models corresponding to the history labels, the to-be-migrated models are trained image identification models for identifying the history labels corresponding to the to-be-migrated models, for example, the to-be-migrated models may be deep learning models or target detection models, the network structure of the to-be-migrated models may be selected according to requirements, for example, the network structure of the to-be-migrated models may be a fast R-CNN model, an SSD model, a YOLO model, or the like, further, the to-be-migrated models are set as target detection models based on the fast R-CNN model, because an Anchor (sliding basic box) used for extracting features in the fast R-CNN model has the characteristic of translation invariability, and each to-be-generated image is subjected to target identification through the to-be-migrated models, the target identification process includes extracting migration features corresponding to the history labels in the images to be generated, locating areas corresponding to the migration features through the extracted migration features, namely identifying areas corresponding to the target labels through the features of the similar history labels, determining the areas corresponding to the migration features as the target areas corresponding to the images to be generated, and finely adjusting the identified areas corresponding to the images to be generated in the process, so that the target areas are accurately determined, and the images to be generated marked with the target areas are recorded as images to be processed.

In an embodiment, as shown in fig. 4, in the step S30, that is, the identifying, by the model to be migrated, each of the images to be generated to obtain a target area corresponding to each of the images to be generated includes:

s301, extracting migration features of the images to be generated through the models to be migrated, and performing region identification according to the extracted migration features to obtain identification regions of the images to be generated.

Understandably, a migration feature corresponding to the history label is extracted from the image to be generated, a region corresponding to the migration feature can be located through the extracted migration feature, namely, the region corresponding to the target label is identified through the features of the similar history labels, and the region corresponding to the target label is determined as the identification region corresponding to the image to be generated.

And S302, performing target fine adjustment on the identification area of each image to be generated to obtain a target area corresponding to each image to be generated.

Understandably, the target fine tuning process is to perform edge segmentation on the image in a preset range adjacent to the edge of the identification region output by the model to be migrated corresponding to the image to be generated, identify an edge line, and perform an edge reduction adjustment process on the identification region along the edge line, so as to obtain the target region corresponding to the identification region.

The method and the device realize the migration feature extraction of each image to be generated through the model to be migrated, and perform region identification according to the extracted migration feature to obtain the identification region of each image to be generated; and carrying out target fine adjustment on the identification area of each image to be generated to obtain a target area corresponding to each image to be generated, so that the area to be identified in the image to be generated can be roughly identified in the model to be migrated automatically, and the final target area to be identified can be accurately framed through the target fine adjustment, so that the positioning accuracy and reliability of the target area are improved.

In an embodiment, in the step S302, that is, performing the target fine adjustment on the identification area of each image to be generated to obtain the target area corresponding to each image to be generated, includes:

and performing edge segmentation in a preset range adjacent to the identification region of the image to be generated to identify an edge line.

Understandably, the edge segmentation is a process of identifying the edge in the identification area, identifying the pixel points of which the color difference value of the pixel points between the adjacent pixel points is greater than the preset color difference value, carrying out closed line segmentation on the identified pixel points, namely segmenting the lines which can be enclosed into the closed area, and segmenting the edge line, wherein the edge line is the line which can be enclosed into the closed area by the pixel points which are greater than the preset color difference value.

And according to the edge line, performing edge reduction on the identification area to obtain the target area corresponding to the image to be generated.

Understandably, performing edge reduction on the identification area according to the area surrounded by the edge lines, wherein the edge reduction is a process that the identification area is reduced to be overlapped with the edge lines and can cover the area surrounded by the edge lines, and finally, the reduced area is determined as the target area, and one image to be generated corresponds to one target area.

The invention realizes the purpose of identifying the edge line by performing edge segmentation in the preset range adjacent to the identification region of the image to be generated; and according to the edge line, performing edge reduction on the identification region to obtain the target region corresponding to the image to be generated, so that automatic fine adjustment on the identification region is realized, the accurate real region is drawn closer, and the accuracy of target region identification is improved.

And S40, performing image enhancement processing based on factor enhancement proportion on each image to be processed, and generating the training images with the number of generated images.

Understandably, the factor enhancement proportion is a preset proportion occupied by each image enhancement factor, the factor enhancement proportion is set on the basis of the total number of the generated image set, the proportion occupied by each factor is an enlarged proportion, namely the proportion of the enlarged number according to the total number of the generated image set, and the image enhancement factors comprise a rotation factor, a fuzzy factor, a contrast factor, a brightness factor and a balance factor; the rotation factor corresponds to a rotation angle image enhancement algorithm; the blurring factor corresponds to a blurred image enhancement algorithm; the contrast factor corresponds to a contrast image enhancement algorithm; the brightness factor corresponds to a brightness image enhancement algorithm; the balance factor corresponds to a color balance image enhancement algorithm; the image enhancement processing based on the factor enhancement proportion is a method for processing each image to be processed according to image enhancement algorithms corresponding to different image enhancement factors, and the factor enhancement proportion is continuously adjusted according to the performance index value of the generated test image in the processing process, and the image is continuously drawn to a high performance index value, so that a training image is generated.

Firstly, setting a rotation angle parameter according to the adjusted factor enhancement ratio, and performing image rotation processing on each image to be processed according to the rotation angle image enhancement algorithm and the set rotation angle parameter to generate a plurality of rotation test images corresponding to each image to be processed; setting a fuzziness parameter, and carrying out image fuzziness processing on each image to be processed according to a fuzzy image enhancement algorithm and the set fuzziness parameter to generate a plurality of fuzzy test images corresponding to each image to be processed; setting contrast parameters, and performing color contrast processing on each image to be processed according to a contrast image enhancement algorithm and the set contrast parameters to generate a plurality of contrast test images corresponding to each image to be processed; setting a fuzziness parameter, and performing brightness enhancement processing on each image to be processed according to a brightness image enhancement algorithm and the set brightness parameter to generate a plurality of brightness test images corresponding to each image to be processed; setting color balance parameters, and performing color balance processing on each image to be processed according to a brightness color balance image enhancement algorithm and the set color balance parameters to generate a plurality of balance test images corresponding to each image to be processed; finally, recording all the rotation test image, the blur test image, the contrast test image, the brightness test image and the balance test image as the test images.

Inputting each test image into the model to be migrated, and identifying the test image through the model to be migrated, so as to determine a performance index value, for example: and the performance index value is the accuracy, and the ratio occupied by the corresponding image enhancement factor is adjusted according to the accuracy of each test image, namely the factor enhancement ratio is adjusted, so that the test image with high performance index value can be optimally output, and the training image is an image required by subsequent training.

In an embodiment, as shown in fig. 5, in the step S40, the performing, on each of the images to be processed, an image enhancement process based on a factor enhancement ratio to generate the generated number of training images includes:

s401, determining the ratio of the generated number to the number of the images to be generated in the image set to be generated as a target ratio; and the generated number is greater than the number of the images to be generated in the image set to be generated.

Understandably, dividing the generated number by the number of all the images to be generated, and performing rounding to obtain the target ratio, wherein the rounding may be upward rounding or downward rounding, and the generated number is greater than the number of the images to be generated in the image set to be generated.

S402, performing multi-equal division processing based on the target proportion on all the images to be processed to obtain an image sequence group comprising a plurality of unit image sets which are sequentially ordered.

Understandably, the images to be processed corresponding to all the images to be generated in the image set to be generated are sorted, the sorted images to be processed are subjected to multi-equal division according to the number of the target ratios, a plurality of unit image sets are divided, one equal division corresponds to one unit image set, the remaining images to be processed which are less than one equal division are determined as one unit image set, in the multi-equal division processing process, the sequence number of the image to be processed which is sorted first in the unit image sets is used as the sequence number of the unit image set, all the unit image sets are sequentially sorted according to the respective sequence numbers, and therefore all the unit image sets which are sequentially sorted are determined as the image sequence set.

S403, selecting the unit image set with the first order in the image sequence group.

S404, according to the target ratio and a preset factor enhancement ratio, image enhancement processing is carried out on the selected unit image set, a plurality of test images corresponding to the unit image set are generated, and a performance index value corresponding to the unit image set is determined.

Understandably, the factor enhancement ratio is initially a ratio value between a preset rotation factor dimension, a preset blur factor dimension, a preset contrast factor dimension, a preset brightness factor dimension, and a preset balance factor dimension, the factor enhancement ratio is adjusted according to an effect of a test image output after image enhancement processing is performed on the unit image sets one by one, the factor enhancement ratio is stopped being adjusted until all the unit image sets are subjected to image enhancement processing, and the image enhancement processing includes a process of processing by an image enhancement algorithm corresponding to each factor, for example: the factors comprise a rotation factor, a blurring factor, a contrast factor, a brightness factor and a balance factor, wherein the rotation factor corresponds to a rotation angle image enhancement algorithm, the blurring factor corresponds to an image blurring algorithm, the contrast factor corresponds to a contrast image enhancement algorithm, the brightness factor corresponds to a brightness image enhancement algorithm, the balance factor corresponds to a color balance image enhancement algorithm, and the rotation angle image enhancement algorithm is used for performing image rotation processing on the image to be generated in the unit image set; carrying out image blurring processing on the images to be generated in the unit image set by using an image blurring algorithm; performing color contrast processing on the image to be generated in the unit image set by using a contrast image enhancement algorithm; performing brightness enhancement processing on the images to be generated in the unit image set by using a brightness image enhancement algorithm; and performing color balance processing on the to-be-generated images in the unit image set by using a color balance image enhancement algorithm to obtain a plurality of test images corresponding to the to-be-processed images and enhancement regions corresponding to the test images, wherein the enhancement regions are regions obtained according to synchronous movement or invariance of target regions of the to-be-generated images along with the algorithm in the image enhancement processing, the obtained test images are input into the to-be-migrated model, the test images are identified through the to-be-migrated model to obtain identification results, and performance index values corresponding to the unit image set are determined according to the coincidence ratio between the identification results corresponding to the test images and the enhancement regions, and the performance index values reflect the effect of the generated test images.

In an embodiment, as shown in fig. 6, in the step S404, that is, performing image enhancement processing on the selected unit image set according to the target ratio and a preset factor enhancement ratio to generate a plurality of test images corresponding to the unit image set, and determining a performance index value corresponding to the unit image set, includes:

s4041, determining the generation quantity of each factor corresponding to one image to be processed according to the target ratio and a preset factor enhancement ratio.

Understandably, the number of the target ratios is taken as the total number, the total number is divided according to the factor enhancement ratio to obtain the generation number of each factor required to be generated by one image to be generated, i.e., the number generated for the rotation factor dimension, blur factor dimension, contrast factor dimension, brightness factor dimension, and balance factor dimension, the factor generation numbers include a rotation factor generation number, a blur factor generation number, a contrast factor generation number, a brightness factor generation number, and a balance factor generation number, the number of the generated rotating factors corresponding to the rotating factor dimension, the number of the generated blurring factors corresponding to the blurring factor dimension, the number of the generated contrast factors corresponding to the contrast factor dimension, the number of the generated brightness factors corresponding to the brightness factor dimension, and the number of the generated balance factors corresponding to the balance factor dimension.

S4042, according to the generated number of each factor, performing image enhancement processing on each to-be-processed image in the unit image set to obtain a test image corresponding to each to-be-processed image in the unit image set and an enhancement region corresponding to the test image.

Understandably, according to the generated number of each factor, correspondingly performing image rotation processing, image blurring processing, color contrast processing, brightness enhancement processing and color balance processing on each to-be-processed image in the unit image set, thereby generating a test image corresponding to each to-be-processed image in the unit image set and an enhanced area corresponding to the test image.

The rotation factor generation quantity corresponds to a rotation angle image enhancement algorithm; the fuzzy factor generation quantity corresponds to a fuzzy image enhancement algorithm; the contrast factor generation quantity corresponds to a contrast image enhancement algorithm; the brightness factor generation quantity corresponds to a brightness image enhancement algorithm; the balance factor generation number corresponds to a color balance image enhancement algorithm.

In an embodiment, in step S4042, the performing image enhancement processing on each to-be-processed image in the unit image set according to the generated number of each factor to obtain a test image corresponding to each to-be-processed image in the unit image set and an enhanced region corresponding to the test image includes:

and respectively setting a rotation angle parameter, a fuzzy degree parameter, a contrast ratio parameter, a brightness parameter and a color balance parameter according to the generation quantity of each factor.

Understandably, setting the rotation angle parameter according to the number of generated rotation factors in the number of generated factors, where the rotation angle parameter is a step length or an angle value of the rotation angle of the image to be processed, and the range can be equally divided according to the number of generated rotation factors by taking 360 degrees as a total range, so as to obtain each rotation angle, thereby determining the rotation angle parameter of each rotation, setting the ambiguity parameter according to the number of generated ambiguity factors in the number of generated factors, where the ambiguity parameter is a parameter for radially blurring adjacent pixel points in the step length for each pixel point in the image to be processed, where the step length can be set according to a requirement, such as an increment increased by 2, setting the ambiguity parameter of the step length of each ambiguity according to the number of generated ambiguity factors, and generating the number according to the contrast factors in the number of generated factors, setting the contrast parameter, where the contrast parameter is a parameter related to a step length for performing defogging and sharpening on each pixel point in the image to be processed, where the step length may be set according to a requirement, for example, an increment of 0.2 is used as an increment, a contrast parameter of the step length for each contrast processing is set according to the generated number of the contrast factors, the brightness parameter is a parameter for increasing a pixel value of each pixel point in the image to be processed according to a preset step length, and the step length may be set according to a requirement, for example, an increment of 2 is used as an increment, a parameter of the step length for each brightness enhancement processing is set according to the generated number of the brightness factors, and the color balance parameter is set according to the generated number of balance factors in the generated number of the factors, the color balance parameter is a parameter of step length for correcting the color cast of the image to be processed.

And respectively performing image rotation processing, image blurring processing, color contrast processing, brightness enhancement processing and color balance processing on each image to be processed according to the rotation angle parameter, the blurring degree parameter, the contrast degree parameter, the brightness parameter and the color balance parameter to generate a plurality of test images corresponding to each image to be processed and enhancement regions corresponding to the test images.

Understandably, setting a rotation angle parameter according to the rotation factor generation quantity, and performing image rotation processing on each image to be processed according to the rotation angle image enhancement algorithm and the set rotation angle parameter to generate a plurality of rotation test images corresponding to each image to be processed and enhancement areas corresponding to the rotation test images; one image to be processed corresponds to the rotating test images with the same number as the rotating factor; setting a fuzziness parameter according to the generated number of the fuzziness factors, and performing image fuzziness processing on each image to be processed according to a fuzzy image enhancement algorithm and the set fuzziness parameter to generate a plurality of fuzzy test images corresponding to each image to be processed and enhancement areas corresponding to the fuzzy test images; one to-be-processed image corresponds to the fuzzy test images with the same number as the fuzzy factor generation number; setting contrast parameters according to the generated number of the contrast factors, and performing color contrast processing on each image to be processed according to a contrast image enhancement algorithm and the set contrast parameters to generate a plurality of contrast test images corresponding to each image to be processed and enhancement areas corresponding to the contrast test images; one image to be processed corresponds to the comparison test images with the same number as the comparison factors; setting brightness parameters according to the generated quantity of the brightness factors, and performing brightness enhancement processing on each image to be processed according to a brightness image enhancement algorithm and the set brightness parameters to generate a plurality of brightness test images corresponding to each image to be processed and enhancement areas corresponding to the brightness test images; one image to be processed corresponds to the brightness test images with the same quantity as the brightness factor; setting color balance parameters according to the balance factor generation quantity, and performing color balance processing on each image to be processed according to a brightness color balance image enhancement algorithm and the set color balance parameters to generate a plurality of balance test images corresponding to each image to be processed and enhancement areas corresponding to the balance test images; one image to be processed corresponds to the balance test images with the same number as the balance factor generation number; recording all of the rotated test image, the blurred test image, the contrast test image, the brightness test image, and the balance test image as the test images.

The image rotation processing is a processing process of rotating the rotation angle parameter by taking an image center point as a center, the image blurring processing is a processing process of blurring an image according to the blurring parameter, the color contrast processing is a processing process of making a difference between pixel points in the image to be processed more obvious according to the contrast parameter, the brightness enhancement processing is a processing process of increasing a brightness value of the pixel points in the image to be processed according to the brightness parameter, and the color balance processing is a process of performing equalization processing between the pixel points in the image to be processed according to the color balance parameter.

The invention realizes that the rotation angle parameter, the ambiguity parameter, the contrast parameter, the brightness parameter and the color balance parameter are respectively set according to the generation quantity of each factor; according to the rotation angle parameter, the ambiguity parameter, the contrast parameter, the brightness parameter and the color balance parameter, image rotation processing, image ambiguity processing, color contrast processing, brightness enhancement processing and color balance processing are respectively carried out on each image to be processed, and a plurality of test images corresponding to each image to be processed and enhancement regions corresponding to the test images are generated.

S4043, inputting each test image obtained through image enhancement processing into the model to be migrated, and identifying the test image through the model to be migrated to obtain an identification result.

Understandably, inputting the test image into the model to be migrated, extracting the implicit features of the test image through the model to be migrated, and identifying the region according to the extracted implicit features, thereby identifying the region with the implicit features as the identification result, wherein the implicit features are the migration features corresponding to the migrated model to be migrated.

S4044, determining a performance index value corresponding to the unit image set according to the identification result and the enhancement region corresponding to each test image.

Understandably, the overlap ratio of the area part in the identification result corresponding to one test image and the area part of the enhanced area is calculated, so as to judge the accuracy, and finally the performance index value corresponding to the unit image set and related to each factor is determined.

The generation quantity of each factor corresponding to one image to be processed is determined according to the target ratio and the preset factor enhancement ratio; performing image enhancement processing on each to-be-processed image in the unit image set according to the generated number of each factor to obtain a test image corresponding to each to-be-processed image in the unit image set and an enhancement area corresponding to the test image; inputting each test image obtained by image enhancement processing into the model to be migrated, and identifying the test image through the model to be migrated to obtain an identification result; and determining a performance index value corresponding to the unit image set according to the identification result and the enhancement region corresponding to each test image, so that the identification result of the test image is automatically identified, the performance index value is automatically judged, an adjustment basis is provided for the subsequent factor enhancement proportion, manual identification and adjustment are not needed, the cost is saved, the quality of the subsequent training image set generation is improved, and the high-quality images can be drawn together.

S405, detecting whether unit image sets except the first unit image set in the sequence of images exist.

S406, if it is detected that a unit image set except the first ordered unit image set exists, deleting the first ordered unit image set from the image sequence set, adjusting the factor enhancement proportion based on the performance index value, returning to the step of selecting the first ordered unit image set in the image sequence set until the image sequence set is empty, and recording all the test images as the training images.

Understandably, if it is detected that a unit image set except for the first unit image set is present in the image sequence set, the first unit image set is removed from the image sequence set, values corresponding to the factors in the factor enhancement proportion are adjusted according to the performance index, that is, the proportion values among the factors are adjusted, the factor corresponding to the accuracy higher than the average accuracy is increased correspondingly, the factor corresponding to the accuracy lower than the average accuracy is decreased correspondingly, after the factor enhancement proportion is adjusted, the step of selecting the first unit image set in the image sequence set is returned to be executed, the factor enhancement proportion is continuously adjusted until no unit image set is present in the image sequence set, that is, the image sequence set is empty, thereby recording all of the test images as the training images.

Therefore, the invention realizes that the factor enhancement proportion can be continuously adjusted through the image enhancement processing based on the factor enhancement proportion, thereby generating the high-quality training image.

In an embodiment, after the step S405, that is, after the detecting whether there is a unit image set other than the first-ranked unit image set in the image series group, the method further includes:

if no unit image set except the first unit image set is detected, the unit image sets are processed by image enhancement, and the factor enhancement proportion is not required to be adjusted.

And S50, associating each training image with the target label associated with the image to be generated corresponding to each training image.

Understandably, an association relationship between the training image and the target label associated with the image to be generated corresponding to the image to be processed corresponding to the training image is established.

And S60, determining all the images to be processed and all the associated training images as a training image set.

Understandably, all the images to be processed and all the associated training images are recorded as a training image set, and then the training image set is used as a sample of a subsequent training neural network, so that the time and the cost for collecting the sample are saved, and the accuracy of neural network learning is improved.

The invention realizes the purpose of acquiring the image set to be generated and the generation quantity; crawling historical labels similar to the target labels and the target description in a historical label library by using a text similarity technology, and determining the category to be migrated according to all the crawled historical labels; acquiring a model to be migrated corresponding to the category to be migrated from an identification model library, identifying each image to be generated through the model to be migrated to obtain a target area corresponding to each image to be generated, and recording the image to be generated marked with the target area as an image to be processed; performing image enhancement processing based on factor enhancement proportion on each image to be processed to generate the training images with the generated number; associating each training image with the target label associated with the image to be generated corresponding to each training image; all the images to be processed and all the associated training images are determined as training image sets, so that the automatic generation of the training image sets based on a small number of zero-labeled image sets to be generated is realized through a text similarity technology, a crawling technology and a transfer learning technology and by applying image enhancement processing, the manual labeling time and the workload of training image collection are reduced, the labeling efficiency is improved, the investment cost is saved, the accuracy and the efficiency of generating the training image sets are improved, and high-quality training images are provided for subsequent training.

It should be understood that, the sequence numbers of the steps in the foregoing embodiments do not imply an execution sequence, and the execution sequence of each process should be determined by its function and inherent logic, and should not constitute any limitation to the implementation process of the embodiments of the present invention.

In an embodiment, an apparatus for generating training image sets is provided, and the apparatus for generating training image sets corresponds to the method for generating training image sets in the above embodiments one to one. As shown in fig. 7, the training image set generating apparatus includes an obtaining module 11, a crawling module 12, a migration module 13, a generating module 14, an association module 15, and a determination module 16. The functional modules are explained in detail as follows:

the acquisition module 11 is configured to acquire a set of images to be generated and a generation number; the image set to be generated comprises a plurality of images to be generated; associating one of the images to be generated with a target label; one said object tag corresponding to one object description; the generated number is the number of training images for training generated by the image set to be generated;

the crawling module 12 is configured to crawl historical tags similar to the target tags and the target descriptions in a historical tag library by using a text similarity technology, and determine categories to be migrated according to all the crawled historical tags;

the migration module 13 is configured to obtain a to-be-migrated model corresponding to the to-be-migrated category from an identification model library, identify each to-be-generated image through the to-be-migrated model to obtain a target area corresponding to each to-be-generated image, and record the to-be-generated image marked with the target area as a to-be-processed image;

a generating module 14, configured to perform image enhancement processing based on a factor enhancement ratio on each to-be-processed image, and generate training images of the generated number;

the association module 15 is configured to associate each training image with the target label associated with the to-be-generated image corresponding to each training image;

a determining module 16, configured to determine all the images to be processed and all the associated training images as a training image set.

For specific limitations of the training image set generating apparatus, reference may be made to the above limitations of the training image set generating method, which are not described herein again. The modules in the training image set generation device can be wholly or partially implemented by software, hardware and a combination thereof. The modules can be embedded in a hardware form or independent from a processor in the computer device, and can also be stored in a memory in the computer device in a software form, so that the processor can call and execute operations corresponding to the modules.

In one embodiment, a computer device is provided, which may be a client or a server, and its internal structure diagram may be as shown in fig. 8. The computer device includes a processor, a memory, a network interface, and a database connected by a system bus. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device comprises a readable storage medium and an internal memory. The readable storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of an operating system and computer programs in the readable storage medium. The network interface of the computer device is used for communicating with an external terminal through a network connection. The computer program is executed by a processor to implement a training image set generation method.

In one embodiment, a computer device is provided, which includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor executes the computer program to implement the training image set generating method in the above embodiments.

In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, which, when being executed by a processor, implements the training image set generating method in the above-described embodiments.

It will be understood by those skilled in the art that all or part of the processes of the methods of the embodiments described above can be implemented by hardware instructions of a computer program, which can be stored in a non-volatile computer-readable storage medium, and when executed, can include the processes of the embodiments of the methods described above. Any reference to memory, storage, databases, or other media used in embodiments provided herein may include non-volatile and/or volatile memory. Non-volatile memory can include read-only memory (ROM), Programmable ROM (PROM), Electrically Programmable ROM (EPROM), Electrically Erasable Programmable ROM (EEPROM), or flash memory. Volatile memory can include Random Access Memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDRSDRAM), Enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus Direct RAM (RDRAM), direct bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

It will be apparent to those skilled in the art that, for convenience and brevity of description, only the above-mentioned division of the functional units and modules is illustrated, and in practical applications, the above-mentioned function distribution may be performed by different functional units and modules according to needs, that is, the internal structure of the apparatus is divided into different functional units or modules to perform all or part of the above-mentioned functions.

The above-mentioned embodiments are only used for illustrating the technical solutions of the present invention, and not for limiting the same; 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; such modifications and substitutions do not substantially depart from the spirit and scope of the embodiments of the present invention, and are intended to be included within the scope of the present invention.

Claims

1. A training image set generation method is characterized by comprising the following steps:

2. The training image set generating method of claim 1, wherein the crawling history tags similar to the target tag and the target description in a history tag library by using a text similarity technique, and determining the category to be migrated according to all the crawled history tags comprises:

crawling a network description matched with the target tag by using a web crawler technology;

according to the network description, carrying out keyword weighting on the target description to obtain a focused description;

comparing the historical description and the focusing description under each historical label in the historical label library by using a text similarity technology to obtain a similarity value corresponding to each historical description;

and determining the history label corresponding to the history description corresponding to the maximum similarity value as the category to be migrated.

3. The training image set generation method of claim 1, wherein the identifying each image to be generated through the model to be migrated to obtain a target region corresponding to each image to be generated comprises:

carrying out migration feature extraction on each image to be generated through the model to be migrated, and carrying out region identification according to the extracted migration features to obtain an identification region of each image to be generated;

and performing target fine adjustment on the identification area of each image to be generated to obtain a target area corresponding to each image to be generated.

4. The training image set generating method of claim 3, wherein the performing target fine-tuning on the identified region of each of the images to be generated to obtain a target region corresponding to each of the images to be generated comprises:

performing edge segmentation on the image to be generated in a preset range adjacent to the identification region to identify edge lines;

5. The method as claimed in claim 1, wherein said generating the generated number of training images by performing image enhancement processing based on factor enhancement ratio on each of the images to be processed comprises:

determining the ratio of the generated number to the number of the images to be generated in the image set to be generated as a target ratio; the number of the generated images is larger than the number of the images to be generated in the image set to be generated;

performing multi-equal division processing based on the target ratio on all the images to be processed to obtain an image sequence group comprising a plurality of unit image sets which are sequentially ordered;

selecting the unit image set with the first order in the image sequence group;

performing image enhancement processing on the selected unit image set according to the target ratio and a preset factor enhancement ratio to generate a plurality of test images corresponding to the unit image set, and determining a performance index value corresponding to the unit image set;

detecting whether a unit image set except the unit image set which is the first ordered exists in the image sequence group;

if the unit image sets except the first unit image set are detected to exist, deleting the first unit image set from the image sequence set, adjusting the factor enhancement proportion based on the performance index value, returning to the step of selecting the first unit image set in the image sequence set until the image sequence set is empty, and recording all the test images as the training images.

6. The method as claimed in claim 5, wherein said image enhancement processing the selected unit image set according to the target ratio and the predetermined factor enhancement ratio to generate a plurality of test images corresponding to the unit image set and determine the performance index corresponding to the unit image set comprises:

determining the generation quantity of each factor corresponding to one image to be processed according to the target ratio and a preset factor enhancement ratio;

performing image enhancement processing on each to-be-processed image in the unit image set according to the generated number of each factor to obtain a test image corresponding to each to-be-processed image in the unit image set and an enhancement area corresponding to the test image;

inputting each test image obtained by image enhancement processing into the model to be migrated, and identifying the test image through the model to be migrated to obtain an identification result;

and determining a performance index value corresponding to the unit image set according to the identification result and the enhancement area corresponding to each test image.

7. The training image set generating method of claim 6, wherein the performing image enhancement processing on each to-be-processed image in the unit image set according to the generated number of each factor to obtain a test image corresponding to each to-be-processed image in the unit image set and an enhanced region corresponding to the test image comprises:

respectively setting a rotation angle parameter, a fuzziness parameter, a contrast parameter, a brightness parameter and a color balance parameter according to the generation quantity of each factor;

8. An apparatus for training image set generation, comprising:

9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the training image set generating method according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, in which a computer program is stored which, when being executed by a processor, carries out the training image set generating method according to any one of claims 1 to 7.