Method of optimizing ai model for image recognition

Abstract

For optimizing an Al model for image recognition, a training operation is performed on the Al model with a first group of images for training to obtain a first optimized Al model. The first optimized Al model is loaded into the image module to obtain a working image module, which is used to perform an image recognition test on a group of specific images to obtain a test result data file. The test result data file is automatically analyzed, and a strategy of additionally generating images for training is automatically determined when an analysis result indicates an unsatisfactory result. A second group of images for training is automatically generated according to the strategy. The first optimized Al model is further trained with the second group of images for training to obtain a second optimized Al model, which is loaded into the image module to obtain another working image module.

Description

METHOD OF OPTIMIZING AI MODEL FOR IMAGE RECOGNITIONFIELD OF THE INVENTION

[0001] The present invention relates to optimization a onf AI model for image recognition, and more particularly to a trainingth moed to optimize an AI model for image recognition.BACKGROUND OF THE INVENTION

[0002] Up to now, Artificial Intelligence (AI) technoloyg has entered a matured stage, so it can be introduced into industries a fpoprlications. However, current AI software and hardware architectures are complex in adnivdidual training cost is high. Thus it is adverse to introduce AI technology i snmtoall and medium-sized enterprises without always available AI professional enginee Frso.r example, an image recognition engine constructed by artificial initgeellnce technology can continuously enhance the ability of vehicle recognition throu dgehep learning. However, currentlyavailable image data that have been previouslye c toeldl and preprocessed may not besuitable for some specific and niche purpos Wehs.en the application scenario is an event hardly occurring in real life, it is diffictul to obtain enough varieties and quantities of images in connection with the raren et.v Therefore, a general AI model trained with the above-mentioned ready-made imagtae i ds not suitable for all image recognition hardware modules. Especially when t iot i bse used in a specific scenario or the requirements on image styles or qualitiersy, v dairect application of the ready- made high-definition AI model will lead to misjudgemnt. As such, it is one of the main1objects of the disclosure to ameliorate the drawksba ocf the conventional means as mentioned above.SUMMARY OF THE INVENTION

[0003] An aspect of the present invention relatese tloate rs to a method of optimizing an AI model for image recognition, whic ish adapted to be used in a training process of an image module. The methodlud inecs steps of: performing atraining operation on the AI model for image reciotigon with a first group of imagesfor training to obtain a first optimized AI modeolr f image recognition; loading the first optimized AI model for image recognition in tthoe image module to obtain a first working image module; using the first working ima mgoedule to perform a first imagerecognition test on a group of specific images to obtain at f tiers t result data file;automatically analyzing the first test result data file to obtain an analyseis u rlt;automatically determining a first strategy of aidodniatlly generating images for training when the analysis result indicates an t uisnfsaactory result; automatically generating a second group i omfages for training according to the first strategy of additionally generating images for train;in pgerforming a training operation on the first optimized AI model for image recognition with e second group of images for training to obtain a second optimized AI model im foarge recognition; and loading the second optimized AI model for image recognitiono in thte image module to obtain a second working image module.

[0004] Another aspect of the present invention rel taote as method of optimizing an AI model for image recognition, which is adapted be to used in a training process of2an image module. The method includes steps pr of c:essing an initial image data setto realize respective coordinates of all image dpaotiants included in the initial image data set in a multi-dimensional scatter diagramto;m auatically analyzing the multi-dimensional scatter diagram to locate a specifeica, a wr here the coordinates of all image data points are insufficient or lacking, a cnodrresponding the specific area to a specific image category; supplementing image deatloan bging to the specific image category, and adding the supplemented image dtaota th ine initial image data set to form a first expanded image data set; and trai tnhineg AI model for image recognition with the first expanded image data set to obtain a firsitm oipzetd AI model for image recognition, which is to be loaded into the imagoed mule to form a first working image module.

[0005] A further aspect of the present invention resla tote a method of optimizing an AI model for image recognition, which is adap ttoed be used in a training processof an image module. The method includes steps pr of c:essing an initial image dataset to realize respective coordinates of all im daagtea points included in the initial image data set in a multi-dimensional table; autoicmalaly a statistical analysis of the multi-dimensional table to locate a specific pahrotw sing a relatively few quantity of images, and corresponding the specific part to aecif sicp image category; supplementing image data belonging to the spe icmifaicge category, and adding thesupplemented image data into the initial image dseata to form a first expanded imagedata set; and training th AeI model for image recognition with th fierst expanded imagedata set to obtain a first optimize AdI model for image recognition, which is to beloaded into the image module to form a first wogrki mnage module.3BRIEF DESCRIPTION OF DRAWINGS

[0006] The invention will become more readily appareon tth tose ordinarily skilled in the art after reviewing the following detailedes dcription and accompanying drawings, in which:FIG. 1 is a functional block diagram schematica ill ulystrating a method of using animage database to train an AI model for image rneictiog n according to the presentinvention;FIG. 2 is a flowchart schematically illustrating m aethod of optimizing an AI model for image recognition according to the presentn intivoen;FIG. 3A is a schemati tcwo-dimensional scatter diagram generated by lnogca atill the image data included in an initial image data; seFIG. 3B is a flowchart schematically illustratintgep ss for automatically generating images according to an embodiment of the presevnetn itnion; andFIG. 4 is a flowchart schematically illustratingep sst for improving image recognition capability of an optimized AI model.DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS

[0007] The invention will now be described more specailfliyc with reference to the following embodiments. It is to be noted that thoello fwing descriptions of preferred embodiments of this invention are presented he froerin purpose of illustration and description only. It is not intended to be exhavues otir to be limited to the precise form disclosed.4

[0008] It should be noted that the terms "first" and "snedc"o in the specification, claims and drawings of the present disclosure asered u to distinguish similar items, and are not necessarily used to describe a spe ocrdifiecr or sequence. It should be understood that the materials so used are integrcehaabnle under appropriate circumstances, so that the embodiments of the t iniovnen described herein can be implemented in other orders than those illustra otred escribed herein.

[0009] Please refer to FIG.1. An image database 10 ids f uosre training an artificial intelligence (AI) model 11 for image recognition. a I training stage, numerous and diverse image data are retrieved from the imagaeb daaste 10 and fed into the AI model 11 for image recognition to result in an optimiz AeId model 110 for image recognition. The optimized AI model 110 for image recognition th isen loaded into an image module 12 to form a working image module 13. I utn isderstood that if data amount and diversity of the image database 10 are insieunfftic and monotonous, it would takea lot of time to conduct training of the AI mode1l t 1o obtain the optimized AI model110 for image recognition. Without well training o thfe AI model for image recognition, misjudgement might occur at a highb parboility and the resulting working image module 13 would not have satisfying perforcmea.n

[0010] For enhancing accuracy of the AI model for ima regceognition, a training method for optimizing an AI model for image recotgionni according to an embodiment of the present invention is schemaytic ilalullstrated in a flowchart shown in FIG. 2. In Step 21, a program is first execu ftoerd analyzing an initial image data set in the image database 10 to realize respec otiovredinates of all the image data points included in the initial image data set. T imheage data points constitute a scatter5diagram formed with the coordinates. In an embodnitm, tehe program for analysis can be executed by a hardware system installed with AI th meodel 11 or any other suitable information system to perform numerical analysis al olf the image data in the initial image data set based on at least two parameterr es.xa Fmople, in a case that the initialimage data set includes images of a variety ofc vle shi in a variety of scenes, one ofthe parameters, e.g., a first parameter, may b aect aunal height of a vehicle, and another one of the parameters, e.g., a second parametyer b,e ma n actual length of a vehicle. The image data associated with the first param aentder the image data associated with the second parameter can be combined to obtaino a-di tmwensional coordinate indicating a specified imag.e As such, when analysis of all the image datau idnecdl in the initial image data set is complete twd,o-dimensional coordinates of respective images are obtained, thereby tw ao-dimensional scatter diagram as exemplified in FIG. 3A can be obtained. It is to be noted that p twaorameters, e.g., height and length, of a vehicle are considered in this embodiment, a b tuhtird parameter or further may be additionally considered to define a three-dimioennasl scatter diagram or a more-dimensional correlation table. For example, therd th piarameter may be brightness of the surrounding, angle of the camera relativee to s tchene, field of view (FOV) of the camera, angle of the objects relative to the ca,m deisratance of the objects relative tothe camera, contract level of the image, rainfcaall e s on site, mist scale on site, smokescale on site, windy scale on site, etc.

[0011] Afterwards, image analysis is further performend t ohe scatter diagram shown in FIG. 3A. For example, from the scatterg draiam, it can be seen that point density is relatively low in some area. That is,in ptso in such specific area are6insufficient or lacking. Therefore, in Step 22,a lotec such specific area first. It is to be noted that the above-mentioned scatter diagramus its o jne of the examples for realizing what kinds of image data are insufficieonrt lacking, and alternatively, any other suitable algorithm or any other suitabler daiamg or table may be executed or presented to locate specific area where image a draeta insufficient or lacking. For example, a data file may be presented as a mumltei-ndsiional diagram or a multi-dimensional table for classification and storag teh oef image data. The data file is then statistically analyzed to reveal specific partsw sihnog relatively few quantities. It is understood that the specific parts inherently iantdeic what kinds of image data are insufficient or lacking in the initial image dateat s based on the parameters used for establishing the multi-dimensional diagram or a t mi-duiml ensional table. As mentioned above, the images may be classified a in vtaoriety of categories according to length of vehicles, height of vehicles, brighstsne or other factors of the surroundings, etc. With regards to insufficien lta ocrking categories of image data, theimage data can be supplemented by executing St.e Inp 2 an3 embodiment, images canbe supplementally collected on site. In another o edmimbent, images can besupplemented by image generation instead of imoalgle c ction. In either way or bothways, the number of images in the insufficienta ocrk ling image categories can beincreased. Nevertheless, when the multi-dimensio dnial gram or the multi-dimensional table includes parameters in connec wtiiothn infrequent events in real life, e.g., typhoons, fires, volcanic eruptions,rn taodoes, etc., and image data corresponding to these events are insufficienatc okrin lg, it is preferable to supplement image data by image generation. Subsequently,u tphpele smental image data are added7into the initial image data set to form a firsta exnpded image data set.

[0012] In a case that Step 23 is executed by image a gteionne,r a method of automatically generating images for supplementinnsguf ificient or lacking images according to an embodiment of the present inven istio enxemplified with reference to the flowchart shown in FIG.3B. First, a specifeicy kword corresponding to one of the insufficient or lacking image categories is autoimcaallty generated (Step 31). It is understood for more than one insufficient or lacgk iimnage categories, more than one corresponding keywords are automatically genera rteesdp,ectively. Examples of the keyword may include, but not limited to, typhooinre,, f volcano eruption, tornado, etc. Next, in Step 32, the specific keyword is fed i ant loarge language model (LLM), such as ChatGPT, to generate a prompt required for aauttiocm image generation. Then the prompt is inputted into an AI tool for image genteiorna, such as Stable Diffusion, to automatically generate a variety of images belogng toin the corresponding image category (Step 33). For example, the keyword “tyopnh”o can result in automatic generation of various images in connection withh toyopn; the keyword “fire” canresult in automatic generation of various image cso in nection with fire; the keyword“volcano eruption” can result in automatic geneornat oif various images in connection with volcano eruption; and the keyword “tornado”n c raesult in automatic generation of various images in connection with tornado. Thuetom aatic image generation isadvantageous in quickly collecting images of ravre n ets. Meanwhile, the multi-dimensional diagram or a multi-dimensional tabl uep isdated with the generated image data. The image generation Step 33 is optionapllyea reted if it is determined in Step 34 that the multi-dimensional diagram or a multmi-deinsional table has not yet shown8a complete level of image database. In other wo irtd is, desirable that the imagegeneration is repetitively executed until the ima dgaetabase contains sufficient kinds and quantities of image data in the multi-dimenaslio sncatter diagram or the multi-dimensional table (Step 35).

[0013] After the first expanded image data set is b uupilt in the processes illustrated in FIG. 2 and FIG. 3B, the image data in the f eirxsptanded image data set can be fed into the AI model 11 for image recognition at ain tirnag stage, serving as a first group of images for training. Then the AI model 11 fora igme recognition is installed into ahardware for simulation test. For example, thedw har e for simulation test may bean edge image module, a general computer system an,y or other suitable hardware device for executing a simulation test of an AI melo.d After a training task of the AI model 11 for image recognition has been completiethd t whe first group of images fortraining, a first optimized AI model 110 for imag re cognition can be obtained.Subsequently, the first optimized AI model 110 i fmorage recognition can be loaded into the image module 12 to provide the first wonrgki mage module 13.

[0014] Please refer to FIG. 4, in which a flowchart o mf aethod for determining whether the optimized AI model 110 complies with re aquirement for imagerecognition is schematically illustrated. In Ste0p, a 4 simulation test is first performedon the optimized AI model 110 for image recogni,tio an d a first test result data fileis obtained. For example, the simulation test meay a b first image recognition test procedure performed by the working image module T 1h3e. working image module 13 replies to a set of questions relating to ap gr oofu specific images, and the first test result data file is obtained based on the answaebresll led in the group of specific9images. The first test result data file containfsor imnation about whether or not each image in the group is answered correctly and inafotiromn about distribution of the categories of images that are answered correcdtly in acnorrectly, respectively. Next, in Step 41, the first test result data file is aynzeadl, and based on the analysis result, afirst strategy of additionally generating imagesr f toraining is determined. Forexample, by analyzing the information about whet ohre nrot each image in the groupwas answered correctly, there might be some imagte go cry considered not wellrecognized as answering accuracy of images inclu ind tehde image catergory is lower than a preset threshold, for example but not lidm tiote, 50%. As such, the first strategy of additionally generating images for training arcdcinog to an embodiment of the present invention may include addition of a sec gornodup of images belonging to such image categories for training. For example, conoartdeis of the images belonging to such image categories are located and marked in m tuhleti-dimensional scatterdiagram as shown in FIG. 3A or a multi-dimension ta blle in an alternativeembodiment, and it aims to increase the image a dmatoaunt of the marked images. Furthermore, when the above-described first styrat oefg additionally generating images for training is performed to obtain the snedc goroup of images for training, it is preferable to take the specification of the hwaarrde platform for image recognition into consideration, so that the added second g orofu impages for training can be the effective ones complying with the requirements h oef t hardware platform. For example, the hardware platform may be the imageu mleo,d and the specification of the hardware platform may include instruction prsoscineg speed, processor architecture type, model framework, or model impelenmtation method of the image10module.

[0015] In an embodiment of the first strategy of addniatiloly generating images for training, the second group of images for traininagy m be added to increase and / or diversify the image data in the first expanded iema dgata set in Step 42 by, for example, extensive image collection and / or autocm imataige generation in a manner as illustrated in FIG.3B. The first expanded ima dgaeta set incorporating therein the second group of images for training forms a sec eoxnpdanded image data set (Step 43), which can be used to further improve the qiutyan atnd diversity of images for training in the training image database 10. Subesnetqlyu, the first optimized AI model 110 for image recognition can be further trainetdh w thie second expanded image data set (Step 44) so as to obtain a second optimize mdo AdIel for image recognition (Step 45). The second optimized AI model for image rencitoiogn is then loaded into the image module 12 to form a second working image mleo.d Luikewise, another simulation test is performed on the second work imingage module to obtain a testresult data file. Based on the test result daeta, w filhether to further expand imagedatabase or if an on-site test is ready to be prmerefdo is determined. In an embodiment, questions used for the simulation test may be aauttiocmally generated or supplemented by automatically generating images as illustrante FdIG i .3B.

[0016] For performing the on-site test, the first, sedco onr further optimized AImodel for image recognition is installed into ang ed hardware, e.g., the image module12 or another network camera module, which is altleodc at the work site for training and test. For example, when the images to be reizceodgn are vehicle-related images, the worksite may be a traffic junction or an enctrea / nexit of a construction site. The11edge hardware then collects images and performgse im reacognition on site. During the on-site test, if there still are images thantn coat be well recognized, enhanced collection of such images is performed. Alternaltyiv,e the method as illustrated in FIG.3B is executed again to automatically generatee im daagta of the required image categories. The additionally collected or genera imteadge data are then added to the training image database 10 to enhance the qua anntdity diversity of images containedtherein. Repetitively, the image database 10 cnoint gai the additionally generatedimage data is used for training and further optiinmgiz the AI model for image recognition in order to evolve the AI model for igmea recognition.

[0017] While the invention has been described in termf s wh oat is presently considered to be the most practical and prefermrebdo ediments, it is to be understood that the invention needs not be limited to thel doissecd embodiment. On the contrary, it is intended to cover various modifications animdil sar arrangements included within the spirit and scope of the appended claims whriech to a be accorded with the broadest interpretation so as to encompass all such modtiiofincsa and similar structures.12

Claims

WHAT IS CLAIMED IS:

1. A method of optimizing an AI model for image recoitgionn, which is adapted to be used in a training process of an image modhuele m, tethod comprising steps of: performing a training operation on the AI model i fmorage recognition with a first group of images for training to obtain at fi orsptimized AI model for image recognition;loading the first optimized AI model for image regcnoition into the image module to obtain a first working image module;using the first working image module to performir ast f image recognition test on a group of specific images to obtain a first testu rlt data file;automatically analyzing th feirst test result data file to obtain an analyseis u rlt;automatically determining a first strategy of aidodnitally generating images for training when the analysis result indicates ant uisnfsaactory result;automatically generating a second group im oafges for training according to the first strategy of additionally generating imagers t froaining;performing a training operation on the first optzimedi AI model for image recognition with the second group of images foirn tinrag to obtain a second optimized AI model for image recognition; andloading the second optimized AI model for imageo rgencition into the image module to obtain a second working image module.

2. The method according to claim 1, wherein the f imirsatge recognition test includes: the first working image module replying to a set of questions relating to images1comprised in the group of specific images, wher theein first test result data file is obtained based on correct answers labelled inm thaege is comprised in the group ofspecific images, and the first test result datea c fiol mprises information aboutwhether or not each of the images comprised in g trhoeup of specific images is answered correctly.

3. The method according to claim 2, where thine information about whether or not each of the images comprised in the group of sipce icmifages is answered correctly is analyzed, and the analysis result indicates un thseatisfactory result if there is at least one unsatisfactory image category, whosee imsa agre recognized with an accuracy lower than a preset threshold in fi trhset image recognition test.

4. The method according to claim 3, wherein t fhirest strategy of additionally generating images for training includes: automallytic laocating the at least one unsatisfactory image category, and automaticallnye graeting images belonging to the at least one unsatisfactory image category.

5. The method according to claim 4, wherein s theecond group o ifmages for training is automatically generated b aydding the automatically generated images belogngin to the at least one unsatisfactory image categnotory th ie first group o ifmages for training.

6. The method according to claim 5, wherein t fhirest strategy of additionally generating images for training further includes:em expting any one of the2automatically generated images, which does not clyom wpith a requirement of ahardware specification, from being added into tirhset g f roup of images for training.

7. The method according to claim 6, wherein h thaerdware specification includes an instruction processing speed of the image modu plero,c aessor architecture type of the image module, a model framework of the imagedu mleo, or a model implementation method of the image module.

8. The method according to claim 4, wherein i tmheages belonging to the at least one unsatisfactory image category are automaticallye graetned by:automatically generatin agt least one specific keyword corresponding to a theleast one unsatisfactory image categ;oryfeeding the at least one specific keyword into a large langu magoedel to generate at least one prompt required for automatic imagneer gaetion;inputting the at least one prompt into an AI tool for generat iimngages to automatically generate a plurality of images beilnogng to the at least one unsatisfactory image category; andcontinuing generating images belonging to thea astt le one unsatisfactory image category until an image database correspondinhge to se tcond group of images for training reaches a predetermined complete level.

9. A method of optimizing an AI model for image recoitgionn, which is adapted to be used in a training process of an image modhuele m, tethod comprising steps of:3processing an initial image data set to realizepe rcetsive coordinates of all image data points included in the initial image data in se at multi-dimensional scatter diagram;automatically analyzing the multi-dimensional secra dttiagram to locate a specific area, where the coordinates of all image datas po ainret insufficient or lacking, andcorresponding the specific area to a specific im caagte gory;supplementing image data belonging to the spe icmifaicge category, and adding the supplemented image data into the initial im daagtea set to form a first expanded image data set; andtraining the AI model for image recognition with th feirst expanded image data set to obtain a first optimize AdI model for image recognition, which is to be loeadd into the image module to form a first working ima mgeodule.

10. The method according to claim 9, wherein t imheage data belonging to the specificimage category are supplemented by automatic im geange ration.

11. The method according to claim 10, wherein timheage data belonging to the specific image category are automatically gener bayte:dautomatically generatin ag specific keyword corresponding to the specificag ime category;feeding the specific keyword into a large language model tnoe graete a prompt required for automatic image generation; andinputting the prompt into an AI tool for generating images too amuattically generate a plurality of images belonging to theci sfipce image category.

412. A method of optimizing an AI model for image recnoitgion, which is adapted to be used in a training process of an image modhuele m, tethod comprising steps of: processing an initial image data set to realizepe rcetsive coordinates of all imagedata points included in the initial image data in se at multi-dimensional table;automatically a statistical analysis of the muilmti-ednsional table to locate a specific part showing a relatively few quantity i omfages, and corresponding the specific part to a specific image category;supplementing image data belonging to the spe icmifaicge category, and adding the supplemented image data into the initial im daagtea set to form a first expanded image data set; andtraining the AI model for image recognition with th feirst expanded image data set to obtain a first optimize AdI model for image recognition, which is to be loeadd into the image module to form a first working ima mgeodule.

13. The method according to claim 12, wherein i tmheage data belonging to the specificimage category are supplemented by automatic im geange ration.

14. The method according to claim 13, wherein timheage data belonging to the specific image category are automatically gener bayte:dautomatically generatin ag specific keyword corresponding to the specificag ime category;5feeding the specific keyword into a large language model tnoe graete a prompt required for automatic image generation; andinputting the prompt into an AI tool for generating images too amuattically generate a plurality of images belonging to theci sfipce image category.6

Images