Algorithm development method and system for image detection and recognition, and electronic equipment

Abstract

The present invention provides an algorithm development method and system, and electronic equipment for image detection and recognition, and relates to the technical field of image detection and recognition. It includes: determining the data volume of the collected data of the target algorithm, the collected data at least including image data; obtaining the collected data volume based on the first preset collection requirement to obtain the algorithm development data set; obtaining the algorithm annotation data set according to the algorithm development data set; performing algorithm development based on the algorithm annotation data set to obtain the initial development algorithm; when the development accuracy of the initial development algorithm does not meet the preset accuracy requirement, determining the iterative data volume of the iterative data of the target algorithm; obtaining the iterative data volume based on the second preset collection requirement to obtain the algorithm iteration data set; performing algorithm iteration on the initial development algorithm according to the algorithm iteration data set to generate a target algorithm whose development accuracy meets the preset accuracy requirement. The present invention has the characteristics of low algorithm development cost.

Description

Technical Field

[0001] The present invention relates to the technical field of image detection and recognition, and in particular to an algorithm development method and system and electronic equipment for image detection and recognition. Background Art

[0002] Algorithm customization development refers to providing users with tailor-made algorithm solutions based on specific needs and application scenarios. This development method usually involves adjusting and optimizing existing algorithms to meet the unique needs of customers, thereby improving the applicability and effectiveness of the algorithm.

[0003] Taking the algorithm development method for image detection and recognition as an example, the amount of data required in the algorithm development process generally depends on the algorithm type, application scenario and specific task requirements.

[0004] However, the inventors have discovered that the amount of data for current algorithm development is generally determined based on manual experience, but the amount of data determined based on manual experience deviates greatly from the actual data volume demand, resulting in a large deviation between the algorithm development cost and the estimated cost. In addition, during the algorithm development process, the iteration of the current algorithm requires advance data investment or frequent iterations, and there is also the problem of high cost of algorithm model iteration. Summary of the invention

[0005] The present invention provides an algorithm development method and system, and electronic equipment for image detection and recognition, which are used to solve the problems that the algorithm development cost currently exists in the algorithm development process and the expected estimated cost has a large deviation, and the cost of algorithm model iteration is high.

[0006] One aspect of the present invention provides an algorithm development method for image detection and recognition, comprising: determining the data volume of collected data of a target algorithm, the collected data including at least image data; obtaining the collected data volume collected based on a first preset collection requirement to obtain an algorithm development data set; obtaining an algorithm annotation data set based on the algorithm development data set; performing algorithm development based on the algorithm annotation data set to obtain an initial development algorithm; generating a target algorithm when the development accuracy of the initial development algorithm meets the preset accuracy requirement; determining the iterative data volume of iterative data of the target algorithm when the development accuracy of the initial development algorithm does not meet the preset accuracy requirement; obtaining iterative data volume collected based on a second preset collection requirement to obtain an algorithm iteration data set; performing algorithm iteration on the initial development algorithm according to the algorithm iteration data set to generate a target algorithm whose development accuracy meets the preset accuracy requirement.

[0007] According to some embodiments of the present invention, the amount of collected data is:

[0008] ;

[0009] ;

[0010] ;

[0011] ;

[0012] Among them, N is the amount of collected data, N1 is the amount of positive sample data, and N2 is the amount of negative sample data; A is the number of target object categories, and A is the sum of subcategories of all target object categories; B is the number of camera points in the scene that are not repeated; C is the number of target object categories that appear simultaneously in each scene; K1 is the percentage of non-simulated data for each point; D1 is the number of non-simulated collections for each point; K2 is the percentage of simulated data for each point; D2 is the number of simulated collections for each point; K1+K2=1, and K1≥0, K2≥0; Q1 is the number of target object categories that appear simultaneously for each point, and Q2 is the number of collections for each target category that appears simultaneously; Q3 is the number of target object categories that can be simulated simultaneously; Q4 is the number of collections for each target category that can be simulated simultaneously.

[0013] According to some embodiments of the present invention, the first preset acquisition requirement includes a general acquisition requirement, including: for the acquisition data of the analog acquisition and the non-analog acquisition, data is collected for each target object category, and there is at least one target object in each target object category; and

[0014] ;

[0015] ;

[0016] Among them, F1 is the average number of each target category; F2 is the deviation ratio of the number of each target category, 0<F2<1.

[0017] According to some embodiments of the present invention, for the number of non-simulated acquisitions at each point, the first preset acquisition requirement also includes a detailed acquisition requirement, including: the number of acquisitions of each target category appearing simultaneously includes at least the number of perspective acquisitions for acquisitions at different perspectives, the number of light acquisitions for acquisitions at different light rays, the number of distance acquisitions for acquisitions at different distances, the number of orientation acquisitions for acquisitions at different orientations, the number of quality acquisitions for acquisitions at different qualities, and the number of posture acquisitions for acquisitions at different postures; wherein: ; =The number of perspective collections*the number of light collections*the number of distance collections*the number of orientation collections*the number of quality collections*the number of posture collections; for the number of simulated collections at each point, the first preset collection requirement also includes a detailed collection requirement, including: the number of collections of each target category that can be simulated simultaneously includes at least the number of perspective collections for collections at different perspectives, the number of light collections for collections at different light rays, the number of distance collections for collections at different distances, the number of orientation collections for collections at different orientations, the number of quality collections for collections at different qualities, and the number of posture collections for collections at different postures; wherein:

[0018] ;or

[0019]

[0020] It is the number of viewing angle collections * the number of ray collections * the number of distance collections * the number of direction collections * the number of quality collections * the number of posture collections.

[0021] According to some embodiments of the present invention, for the amount of negative sample data, the first preset collection requirement also includes a fine collection requirement, including: the amount of negative sample data at least includes the number of light collections of different light rays; wherein, in the application scenario of classification tasks, the ratio of the amount of positive sample data to the amount of negative sample data at each point is 1:1; in the application scenario of non-classification tasks, the ratio of the amount of positive sample data to the amount of negative sample data at each point is 8:2.

[0022] According to some embodiments of the present invention, in the case of data misreporting, the iterative data volume of the iterative data is: ; Wherein, M1 is the iterative data volume of the iterative data in the case of data misreporting; P1 is the number of false alarm categories; P2 is the number of false alarm points; P3 is the number of false alarm categories that appear at each point at the same time; P4 is the number of samples collected at each point; In the case of data omission, the iterative data volume of the iterative data is:

[0023] ;

[0024] ;

[0025] ;

[0026] Among them, M2 is the amount of iterative data in the case of data omission; R1 is the number of omission target categories, R1 is the sum of omission subcategories of all target categories; R2 is the number of omission points; R3 is the number of omission target categories that appear at each point at the same time; D3 is the number of non-simulated collections at each point in the case of data omission; Q5 is the number of omission target categories that appear at the same time for each scene or point; Q6 is the number of collections for each omission target category that appears at the same time; D4 is the number of simulated collections at each point in the case of data omission; Q7 is the number of omission target categories that can be simulated at the same time; Q8 is the number of collections for each omission target category that can be simulated at the same time.

[0027] According to some embodiments of the present invention, in the case of data false alarm, the second preset collection requirement includes a general collection requirement, including: for the collected data collected by non-simulated collection, each false alarm target object category collects data, and each false alarm target object category has at least one target object; and

[0028] ;

[0029] ;

[0030] Among them, W1 is the average number of false alarms in each category; W2 is the deviation ratio of the number of false alarms in each category, 0<W2<1.

[0031] In the case of data underreporting, the second preset collection requirement includes a general collection requirement, including: for the collected data of the simulated collection and the non-simulated collection, data is collected for each underreported target object category, and there is at least one target object in each underreported target object category; and

[0032] ;

[0033] ;

[0034] Wherein, W3 is the average number of each category of missed target objects; W4 is the deviation ratio of the number of each category of missed target objects, 0<W4<1.

[0035] According to some embodiments of the present invention, in the case of data omission, for the number of non-simulated collections at each point, the second preset collection requirement also includes a fine collection requirement, including: the number of collections of each category of target objects that are simultaneously omit includes at least the number of perspective collections for collections at different perspectives, the number of light collections for collections at different light rays, the number of distance collections for collections at different distances, the number of orientation collections for collections at different orientations, the number of quality collections for collections at different qualities, and the number of posture collections for collections at different postures; wherein, in the case of data omission: ; =The number of viewing angle collections*the number of light collections*the number of distance collections*the number of orientation collections*the number of quality collections*the number of posture collections; in the case of data omission, for the number of simulated collections at each point, the second preset collection requirement also includes a detailed collection requirement, including: the number of collections of each category of omission target objects that can be simulated simultaneously includes at least the number of viewing angle collections for collections at different viewing angles, the number of light collections for collections at different light rays, the number of distance collections for collections at different distances, the number of orientation collections for collections at different orientations, the number of quality collections for collections at different qualities, and the number of posture collections for collections at different postures; wherein:

[0036] ;

[0037] It is the number of viewing angle collections * the number of ray collections * the number of distance collections * the number of direction collections * the number of quality collections * the number of posture collections.

[0038] Another aspect of the present invention provides an algorithm development system for image detection and recognition, including a data processing module and an algorithm development module. The data processing module determines the data volume of the collected data of the target algorithm; obtains the collected data volume collected based on the first preset collection requirement to obtain the algorithm development data set; and obtains the algorithm annotation data set according to the algorithm development data set; the algorithm development module performs algorithm development based on the algorithm annotation data set to obtain the initial development algorithm; generates the target algorithm when the development accuracy of the initial development algorithm meets the preset accuracy requirement; the data processing module determines the iterative data volume of the collected data of the target algorithm when the development accuracy of the initial development algorithm does not meet the preset accuracy requirement; and obtains the iterative data volume of the iterative data collected based on the second preset collection requirement to obtain the algorithm iteration data set; the algorithm development module performs algorithm iteration on the initial development algorithm according to the algorithm iteration data set to generate the target algorithm whose development accuracy meets the preset accuracy requirement.

[0039] Another aspect of the present invention provides an electronic device, which includes: one or more processors; and a storage device for storing one or more programs, which, when executed by the one or more processors, enables the one or more processors to implement the algorithm development method described above.

[0040] Another aspect of the present invention provides a non-volatile computer-readable storage medium having a computer program stored thereon, which can implement the algorithm development method described above when executed by a processor.

[0041] The present invention obtains the collected data of the amount of collected data of the target algorithm based on the first preset collection requirement to obtain an algorithm development data set. The present invention can develop an initial development algorithm through the algorithm development data set. When the development accuracy of the initial development algorithm does not meet the preset accuracy requirement, iterative data of the amount of iterative data collected based on the second preset collection requirement is obtained. Multiple iterations of the initial development algorithm can be completed through the iterative data of the amount of iterative data to obtain a target algorithm that meets the preset accuracy requirement, thereby completing the algorithm development.

[0042] The present invention can accurately obtain an algorithm development data set that meets the algorithm development needs through the precise setting of the data volume and the first preset collection requirement, and can make the actual algorithm development cost closer to the estimated cost to meet user expectations; and the present invention can accurately obtain an algorithm iteration data set that meets the algorithm iteration through the precise setting of the iterative data volume and the second preset collection requirement when the development accuracy of the initial development algorithm does not meet the preset accuracy requirement, so that for the initial development algorithm that does not meet the preset accuracy requirement, the algorithm customization or iteration will not be started until the iterative data meets certain requirements, which can avoid the investment of advance workload or frequent algorithm iterations, and can save the algorithm development cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0044] Figure 1 A flow chart showing an algorithm development method according to an embodiment of the present invention;

[0045] Figure 2 Another flow chart showing the algorithm development method according to an embodiment of the present invention;

[0046] Figure 3 A schematic diagram showing the structure of an algorithm development system according to an embodiment of the present invention is shown.

[0047] Description of reference numerals:

[0048] Algorithm development system 1; data processing module 10; algorithm development module 20. DETAILED DESCRIPTION

[0049] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that the present invention will be comprehensive and complete and fully convey the concepts of the example embodiments to those skilled in the art. The same reference numerals in the figures represent the same or similar parts, and thus their repeated description will be omitted.

[0050] The described features, structures or characteristics may be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present disclosure. However, those skilled in the art will appreciate that the technical solutions of the present disclosure may be practiced without one or more of these specific details, or other modes, components, materials, devices, etc. may be adopted. In these cases, known structures, methods, devices, implementations, materials or operations will not be shown or described in detail.

[0051] In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or devices.

[0052] The terms "first", "second" and the like in the specification, claims and the above drawings of the present invention are used to distinguish different objects rather than to describe a specific order.

[0053] The following is a clear and complete description of the technical solution of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0054] According to one aspect of the present invention, the present invention provides an algorithm development method for generating a specific algorithm model according to a user's specific development requirements. According to an example embodiment, the application scenario of the algorithm development method provided by the present invention can be algorithm development for object detection in image data.

[0055] Optionally, the algorithm development method provided by the present invention can be applied to the development of various deep learning AI algorithms, and the present invention is not limited to this.

[0056] Figure 1 A schematic diagram of a process flow of the algorithm development method according to an embodiment of the present invention is shown. Figure 1As shown, the algorithm development method includes steps S100-S700 (wherein step S500 may include step S500A and step S500B). Exemplarily, the algorithm development method may be executed by an algorithm development system having computing capabilities.

[0057] According to an example embodiment, in step S100 , the algorithm development system determines the data volume of collected data of a target algorithm.

[0058] For example, the target algorithm is determined based on the user's algorithm customization requirements. For example, the algorithm customization requirements include but are not limited to the algorithm name, requirement description (such as business alarm logic), sample data, accuracy requirements, data acquisition difficulty, camera type, camera angle, number of cameras, and number of non-repetitive scenes.

[0059] Exemplarily, the person in charge of algorithm development evaluates the algorithm customization requirements. If the evaluation result is passed, the person in charge of algorithm development outputs information such as algorithm boundary conditions, implementation plan, customization manpower, delivery time, data volume and collection requirements, acceptance plan and risks to determine the target algorithm. The present invention can avoid certain algorithm customization risks by evaluating the algorithm customization requirements.

[0060] After the target algorithm is determined, the algorithm development system responds to user instructions and determines the amount of data required to be collected for the target algorithm.

[0061] For example, the collected data may be image data, or may be video data, text data, voice data, or other types of data, and the present invention does not impose any limitation on this.

[0062] The amount of collected data may be the sum of the amount of positive sample data and the amount of negative sample data. Positive samples may be samples with labels consistent with true sample labels, and negative samples may be samples with labels inconsistent with true sample labels.

[0063] Optionally, in step S100, the amount of collected data may be:

[0064] ;

[0065] ;

[0066] ;

[0067] .

[0068] Among them, N is the amount of collected data, N1 is the amount of positive sample data, and N2 is the amount of negative sample data; A is the number of target object categories, and A is the sum of subcategories of all target object categories; B is the number of camera points in the scene that are not repeated; C is the number of target object categories that appear simultaneously in each scene; K1 is the percentage of non-simulated data for each point; D1 is the number of non-simulated collections for each point; K2 is the percentage of simulated data for each point; D2 is the number of simulated collections for each point; K1+K2=1, and K1≥0, K2≥0; Q1 is the number of target object categories that appear simultaneously for each point, and Q2 is the number of collections for each target category that appears simultaneously; Q3 is the number of target object categories that can be simulated simultaneously; Q4 is the number of collections for each target category that can be simulated simultaneously.

[0069] For example, the collected data is image data, and the target object is an object in the image data that is desired to be detected and identified.

[0070] Exemplarily, taking the application scenario of relics detection in image data as an example, the target object is relics, the number of major target categories = 1 (relics), the number of subcategories of the target category = 5 (tool type), the number of non-repeating camera points B = 60, the number of target categories appearing simultaneously in each scene C = 2, the percentage of non-simulated data at each point K1 = 0, the number of non-simulated collections at each point D1 = 0, the percentage of simulated data at each point K2 = 1, the number of simulated collections at each point D2 = 5*10=50, the number of target categories appearing simultaneously at each point Q1 = 2, the number of target categories that can be simulated simultaneously Q3 = 5, the number of collections of each target category that can be simulated simultaneously = the number of viewing angle collections * the number of light collections * the number of distance collections * the number of orientation collections * the number of quality collections * the number of posture collections / (Q3 / C or Q1) = 2*2*2*3*1*1 / (5 / 2) = 9.6≈10, let the amount of negative sample data = the amount of positive sample data * 20%. Then the amount of collected data = 1*5*60 / 2*1*50 (1+0.2) = 9000. That is, in this application scenario, the amount of collected data required is 9000.

[0071] The present invention can determine the actual amount of data required for target detection in a specific image data through the calculation formula for the data amount of the above-mentioned collected data, thereby avoiding the problem that the actual cost is higher than the estimated cost due to the large amount of data and the actual amount of data, and the problem that the estimated cost is higher due to the small amount of data and the actual amount of data, thereby affecting the order conversion rate.

[0072] In step S200, the algorithm development system acquires the collected data based on the data volume collected according to the first preset collection requirement to obtain an algorithm development data set.

[0073] For example, after obtaining the required amount of collected data in step S100, the algorithm development system obtains the collected data of this amount (such as 9,000 images) collected by the user according to the first preset collection requirement, and generates an algorithm development data set based on the collected data of this amount.

[0074] Optionally, in step S200, the first preset collection requirement may include a general collection requirement, including:

[0075] For the collected data of simulated collection and non-simulated collection, data is collected for each target object category, and there is at least one target object in each target object category; and

[0076] ;

[0077] ;

[0078] Among them, F1 is the average number of each target category; F2 is the deviation ratio of the number of each target category, 0<F2<1.

[0079] For example, simulated data collection refers to the method of collecting data by simulating data through actual scenes and actual props, such as identifying scenes such as safety helmets and reflective clothing. Non-simulated data collection refers to the method of collecting data only through actual scenes, which cannot be simulated by actual scenes and actual props, such as identifying scenes such as peeling wall skin and missing excavator bucket teeth.

[0080] The present invention can prevent the situation where there is no data or too little data for some target object categories by setting the quantity deviation ratio of each target object category.

[0081] Optionally, in step S200, for the number of non-analog acquisitions at each point, the first preset acquisition requirement may further include a detailed acquisition requirement, including:

[0082] The number of collections of each simultaneously appearing target object category includes at least the number of perspective collections for collections at different perspectives, the number of light collections for collections at different light rays, the number of distance collections for collections at different distances, the number of orientation collections for collections at different orientations, the number of quality collections for collections at different qualities, and the number of posture collections for collections at different postures.

[0083] in:

[0084] ;

[0085] It is the number of viewing angle collections * the number of ray collections * the number of distance collections * the number of direction collections * the number of quality collections * the number of posture collections.

[0086] That is, the number of collections for each target category that appears simultaneously = the number of viewing angle collections * the number of light collections * the number of distance collections * the number of direction collections * the number of quality collections * the number of posture collections.

[0087] For example, the number of data collected for each target category that appears at the same time should include data collected from different perspectives, different light sources, different distances, different directions, different qualities, and different postures, and the amount of data of these different types should be as balanced as possible. With this setting, data simulation can be fully performed, and the algorithm model can perform better in various scenes with different perspectives, different light sources, different distances, different directions, different qualities, and different postures, thereby ensuring the generalization ability of the algorithm model.

[0088] Optionally, in step S200, for the number of simulated acquisitions at each point, the first preset acquisition requirement may further include a detailed acquisition requirement, including:

[0089] The number of target object categories that can be simulated simultaneously includes at least the number of perspective collections for different perspectives, the number of light collections for different light rays, the number of distance collections for different distances, the number of orientation collections for different orientations, the number of quality collections for different qualities, and the number of posture collections for different postures.

[0090] in:

[0091] ;or

[0092]

[0093] It is the number of viewing angle collections * the number of ray collections * the number of distance collections * the number of direction collections * the number of quality collections * the number of posture collections.

[0094] That is, the number of target object categories that can be simulated simultaneously = the number of viewing angle collections * the number of light collections * the number of distance collections * the number of orientation collections * the number of quality collections * the number of posture collections / (the number of target object categories that can be simulated simultaneously / the number of target object categories that appear simultaneously in each scene or the number of target object categories that appear simultaneously at each point).

[0095] For example, the number of data collected for each target category that can be simulated simultaneously should include data collected from different perspectives, different light sources, different distances, different directions, different qualities, and different postures, and the amount of data of these different types should be as balanced as possible. With this setting, data simulation can be fully performed, and the algorithm model can perform better in various scenes with different perspectives, different light sources, different distances, different directions, different qualities, and different postures, thereby ensuring the generalization ability of the algorithm model.

[0096] Optionally, in step S200, for the amount of negative sample data at each point, the first preset collection requirement may further include a detailed collection requirement, including:

[0097] The amount of negative sample data includes at least the number of ray collections of different ray collections.

[0098] In the application scenario of classification tasks, the ratio of the amount of positive sample data to the amount of negative sample data at each point is 1:1; in the application scenario of non-classification tasks, the ratio of the amount of positive sample data to the amount of negative sample data at each point is 8:2.

[0099] For example, the amount of negative sample data should include data collected under different light rays, and the amount of data collected under different light rays should be as balanced as possible. With this setting, negative sample data can be collected comprehensively, so that the algorithm model can perform better in various scenes under different light rays, thereby ensuring the generalization ability of the algorithm model. It can also avoid false positives and avoid the problem of no data or too little data for some negative sample types.

[0100] In step S300, the algorithm development system obtains an algorithm annotation data set based on the algorithm development data set.

[0101] In step S400, the algorithm development system develops the algorithm based on the algorithm annotation data set to obtain an initial development algorithm.

[0102] For example, the algorithm development system performs data preprocessing on the algorithm development data set, such as data frame extraction and data deduplication, to filter out valid data, and perform algorithm annotation on the valid data, so as to obtain the algorithm annotation data set. After that, the algorithm development system develops the initial version of the algorithm model and the initial version of the algorithm alarm logic based on the algorithm annotation data set to obtain the initial development algorithm.

[0103] In step S400, the algorithm development system determines whether the development accuracy of the initial development algorithm meets the preset accuracy requirement. If yes, it proceeds to step S500A (e.g. Figure 2 If not, proceed to step S500B (as shown in Figure 1 as shown).

[0104] In step S500A, the algorithm development system generates a target algorithm when the development accuracy of the initial development algorithm meets the preset accuracy requirement.

[0105] For example, the algorithm development system determines the current development accuracy of the initial development algorithm. The calculation formula for the current development accuracy can be:

[0106] Current development accuracy = (the current amount of collected data / the amount of collected data required for algorithm development) * target development accuracy.

[0107] It can be understood here that the current existing algorithm development accuracy evaluation method is based on the test data results and test data annotations of the algorithm, and the current development accuracy can be calculated using a specific calculation method. Compared with the prior art, the calculation method of the current development accuracy provided by the present invention does not rely on test data.

[0108] When the algorithm development system determines that the current development accuracy of the initial development algorithm meets the preset accuracy requirements, the algorithm is put online for testing to generate a target algorithm to complete the algorithm development.

[0109] In step S500B, the algorithm development system determines the iterative data volume of iterative data of the target algorithm when the development accuracy of the initial development algorithm does not meet the preset accuracy requirement.

[0110] For example, when the current development accuracy of the initial development algorithm does not meet the preset accuracy requirement, the algorithm development system performs algorithm iteration on the initial development algorithm. The algorithm development system responds to the user instruction and determines the iteration data amount of the iteration data for performing the algorithm iteration.

[0111] Optionally, in step S500B, in case of data misreporting, the iterative data volume of the iterative data is:

[0112] ;

[0113] Among them, M1 is the amount of iterative data in the case of data false alarms; P1 is the number of false alarm categories; P2 is the number of false alarm points; P3 is the number of false alarm categories that appear at each point at the same time; P4 is the number of samples collected at each point.

[0114] For example, taking the application scenario of non-motor vehicle detection in image data as an example, the number of false alarm types = 4 (such as vehicles, pedestrians, wheelchairs and carts), the number of false alarm points = 3, the number of false alarm types that appear at each point at the same time = 1, and the amount of data collected at each point = 20, then the required iterative data volume = 4*3 / 1*20=240, that is, the amount of data required for iterative data is 240.

[0115] Optionally, in step S500B, in the case of data underreporting, the iterative data amount of the iterative data is:

[0116] ;

[0117] ;

[0118] ;

[0119] Among them, M2 is the amount of iterative data in the case of data omission; R1 is the number of omission target categories, R1 is the sum of omission subcategories of all target categories; R2 is the number of omission points; R3 is the number of omission target categories that appear at each point at the same time; D3 is the number of non-simulated collections at each point in the case of data omission; Q5 is the number of omission target categories that appear at the same time for each scene or point; Q6 is the number of collections for each omission target category that appears at the same time; D4 is the number of simulated collections at each point in the case of data omission; Q7 is the number of omission target categories that can be simulated at the same time; Q8 is the number of collections for each omission target category that can be simulated at the same time.

[0120] For example, taking the application scenario of detecting disorderly stacking of materials in image data as an example, the number of missed target object categories = 3 (such as wooden boards, cement bags, and steel bars), the number of missed points = 5, the number of missed target object categories that appear simultaneously at each point = 1, k1 = 1, k2 = 0, the number of missed target object categories that appear simultaneously = the number of viewing angle collections * the number of light collections * the number of distance collections * the number of orientation collections * the number of mass collections * the number of posture collections = 2*2*2*3*2*1=48 pictures, the number of simulated collections at each point = 0, then the required amount of iterative data = 3*5 / 1*40=720 pictures, that is, the amount of data required for iterative data is 720 pictures.

[0121] The present invention can determine the amount of iterative data required for target detection in a specific image data through the calculation formula of the data amount of the above-mentioned iterative data, so as to accurately calculate the amount of data required for model iteration and effectively control the cost of model iteration.

[0122] In step S600, the algorithm development system acquires iterative data of an iterative data volume collected based on a second preset collection requirement to obtain an algorithm iteration data set.

[0123] For example, after the algorithm development system obtains the required iterative data volume in step S500B, it obtains the iterative data of this iterative data volume collected by the user according to the second preset collection requirement (such as the 240 or 720 image data mentioned above), and generates an algorithm iterative data set based on the iterative data of this iterative data volume.

[0124] Optionally, in step S600, in the case of data misreporting, the second preset collection requirement includes a general collection requirement, including:

[0125] For non-simulated data collection, data is collected for each false alarm target category, and there is at least one target in each false alarm target category; and

[0126] ;

[0127] ;

[0128] Among them, W1 is the average number of false alarms in each category; W2 is the deviation ratio of the number of false alarms in each category, 0<W2<1.

[0129] The present invention can prevent the situation where there is no data or too little data for some false alarm target categories by setting the deviation ratio of the number of each false alarm category.

[0130] In the case of data underreporting, the second preset collection requirements include general collection requirements, including:

[0131] For the collected data of simulated collection and non-simulated collection, data is collected for each missed target object category, and there is at least one target object in each missed target object category; and

[0132] ;

[0133] ;

[0134] Wherein, W3 is the average number of each category of missed target objects; W4 is the deviation ratio of the number of each category of missed target objects, 0<W4<1.

[0135] The present invention can prevent the situation where there is no data or too little data for some missed target object categories by setting the deviation ratio of the number of each missed target object category.

[0136] Optionally, in step S600, in the case of data omission, for the number of non-analog acquisitions at each point, the second preset acquisition requirement further includes a detailed acquisition requirement, including:

[0137] The number of acquisitions of each category of missed target objects at the same time at least includes the number of view acquisitions of acquisitions at different view angles, the number of light acquisitions of acquisitions at different light rays, the number of distance acquisitions of acquisitions at different distances, the number of orientation acquisitions of acquisitions at different orientations, the number of quality acquisitions of acquisitions at different qualities, and the number of posture acquisitions of acquisitions at different postures;

[0138] Among them, in the case of data underreporting:

[0139] ;

[0140] It is the number of viewing angle collections * the number of ray collections * the number of distance collections * the number of direction collections * the number of quality collections * the number of posture collections.

[0141] That is, the number of collections of each category of missed target objects at the same time = the number of viewpoint collections * the number of light collections * the number of distance collections * the number of orientation collections * the number of quality collections * the number of posture collections;

[0142] For example, the number of data collected for each category of missed target objects should include data collected from different perspectives, different light sources, different distances, different directions, different qualities, and different postures, and the amount of data of these different types should be as balanced as possible. With this setting, data can be collected comprehensively, and the algorithm model can perform better in various scenes with different perspectives, different light sources, different distances, different directions, different qualities, and different postures, avoiding missed reports, thereby ensuring the generalization ability of the algorithm model.

[0143] Optionally, in step S600, in the case of data omission, for the number of simulated acquisitions at each point, the second preset acquisition requirement further includes a detailed acquisition requirement, including:

[0144] The number of each missed target object category collection that can be simulated simultaneously includes at least the number of perspective collections collected at different perspectives, the number of light collections collected at different light rays, the number of distance collections collected at different distances, the number of orientation collections collected at different orientations, the number of quality collections collected at different qualities, and the number of posture collections collected at different postures;

[0145] in:

[0146] ;

[0147] It is the number of viewing angle collections * the number of ray collections * the number of distance collections * the number of direction collections * the number of quality collections * the number of posture collections.

[0148] That is, the number of missed target object categories that can be simulated simultaneously = the number of viewing angle collections * the number of light collections * the number of distance collections * the number of orientation collections * the number of quality collections * the number of posture collections / (the number of missed target object categories that can be simulated simultaneously / the number of missed target object categories that appear simultaneously in each scene or point).

[0149] For example, the number of data collected for each category of missed target objects that can be simulated at the same time should include data collected from different perspectives, different light sources, different distances, different directions, different qualities, and different postures, and the amount of data of these different types should be as balanced as possible. With this setting, data simulation can be fully performed, and the algorithm model can perform better in various scenes with different perspectives, different light sources, different distances, different directions, different qualities, and different postures, avoiding missed reports, thereby ensuring the generalization ability of the algorithm model.

[0150] In step S700, the algorithm development system performs algorithm iteration on the initial development algorithm according to the algorithm iteration data set to generate a target algorithm whose development accuracy meets the preset accuracy requirement.

[0151] For example, the algorithm development system iterates the initial development algorithm multiple times based on the iterative data set until the development accuracy of the generated algorithm reaches the preset accuracy requirement, and then completes the algorithm development, thereby obtaining a target algorithm that meets the user's customized needs.

[0152] Optionally, during the algorithm development process, the algorithm development system may also determine timeliness information such as the remaining time for data collection and the remaining time for algorithm delivery.

[0153] For example, the remaining time for data collection = ((the total amount of data required for algorithm customization - the amount of data currently collected) / the amount of data currently collected) * the collection time of the current amount of data collected.

[0154] The remaining delivery time of the algorithm = the remaining time for data collection + the remaining time for data labeling + the model iteration time + the algorithm logic optimization time.

[0155] Among them, the remaining data labeling time = (the total amount of data required for algorithm customization - the current amount of data collected) / the amount of labeled data per unit time; the model iteration time = the number of iterations * the time required for each iteration; the algorithm logic optimization time = the sum of the time for each optimization point. If the number of optimization points is 0, the algorithm logic optimization time is 0, so the remaining delivery time of the algorithm can be calculated.

[0156] Through the above embodiments, the present invention obtains the collected data of the amount of data collected based on the first preset collection requirement by determining the data volume of the collected data of the target algorithm to obtain the algorithm development data set. The present invention obtains the initial development algorithm through the algorithm development data set development, and when the development accuracy of the initial development algorithm does not meet the preset accuracy requirement, iterative data of the amount of iterative data collected based on the second preset collection requirement is obtained. The iterative data of the amount of iterative data can complete multiple iterations of the initial development algorithm to obtain the target algorithm that meets the preset accuracy requirement, thereby completing the algorithm development.

[0157] The present invention can accurately obtain an algorithm development data set that meets the algorithm development needs through the precise setting of the data volume and the first preset collection requirement, and can make the actual algorithm development cost closer to the estimated cost to meet user expectations; and the present invention can accurately obtain an algorithm iteration data set that meets the algorithm iteration through the precise setting of the iterative data volume and the second preset collection requirement when the development accuracy of the initial development algorithm does not meet the preset accuracy requirement, so that for the initial development algorithm that does not meet the preset accuracy requirement, the algorithm customization or iteration will be started only when the iterative data meets certain requirements, which can avoid the investment of advance workload or frequent algorithm iterations, and can save the algorithm development cost.

[0158] According to another aspect of the present invention, the present invention provides an algorithm development system for image detection and recognition, which is used to generate a specific algorithm model according to user-specific development requirements.

[0159] like Figure 3 As shown, the algorithm development system 1 includes a data processing module 10 and an algorithm development module 20 .

[0160] According to an example embodiment, the data processing module 10 determines the data volume of the collected data of the target algorithm.

[0161] For example, the target algorithm is determined based on the user's algorithm customization requirements. For example, the algorithm customization requirements include but are not limited to the algorithm name, requirement description (such as business alarm logic), sample data, accuracy requirements, data acquisition difficulty, camera type, camera angle, number of cameras, and number of non-repetitive scenes.

[0162] Exemplarily, the person in charge of algorithm development evaluates the algorithm customization requirements. If the evaluation result is passed, the person in charge of algorithm development outputs information such as algorithm boundary conditions, implementation plan, customization manpower, delivery time, data volume and collection requirements, acceptance plan and risks to determine the target algorithm. The present invention can avoid certain algorithm customization risks by evaluating the algorithm customization requirements.

[0163] After determining the target algorithm, the data processing module 10 responds to the user's instruction and determines the amount of data required to be collected for the target algorithm.

[0164] For example, the collected data may be image data, or may be video data, text data, voice data, or other types of data, and the present invention does not impose any limitation on this.

[0165] The amount of collected data may be the sum of the amount of positive sample data and the amount of negative sample data. Positive samples may be samples with labels consistent with true sample labels, and negative samples may be samples with labels inconsistent with true sample labels.

[0166] According to an example embodiment, the data processing module 10 acquires the collected data of the amount of data collected based on the first preset collection requirement to obtain an algorithm development data set.

[0167] For example, after obtaining the required amount of collected data, the data processing module 10 obtains the collected data of the amount of data collected by the user according to the first preset collection requirement, and generates an algorithm development data set based on the collected data of the amount of data.

[0168] The data processing module 10 obtains an algorithm annotation data set according to the algorithm development data set.

[0169] For example, the data processing module 10 performs data preprocessing on the algorithm development data set, such as data frame extraction and data deduplication processing, and then filters out valid data, and performs algorithm annotation on the valid data, thereby obtaining an algorithm annotation data set.

[0170] The algorithm development module 20 performs algorithm development based on the algorithm annotation data set to obtain an initial development algorithm.

[0171] For example, the algorithm development module 20 develops a preliminary algorithm model and performs preliminary algorithm alarm logic based on the algorithm annotation data set to obtain an initial development algorithm.

[0172] The algorithm development module 20 generates a target algorithm when the development accuracy of the initial development algorithm meets the preset accuracy requirement.

[0173] For example, the algorithm development module 20 determines the current development accuracy of the initial development algorithm. The calculation formula of the current development accuracy can be:

[0174] Current development accuracy = (the current amount of collected data / the amount of collected data required for algorithm development) * target development accuracy.

[0175] It can be understood here that the current existing algorithm development accuracy evaluation method is based on the test data results and test data annotations of the algorithm, and the current development accuracy can be calculated using a specific calculation method. Compared with the prior art, the calculation method of the current development accuracy provided by the present invention does not rely on test data.

[0176] When the algorithm development module 20 determines that the current development accuracy of the initial development algorithm meets the preset accuracy requirement, an algorithm online test is performed to generate a target algorithm to complete the algorithm development.

[0177] According to an exemplary embodiment, when the development accuracy of the initial development algorithm does not meet the preset accuracy requirement, the algorithm development module 20 determines the iterative data amount of the iterative data of the target algorithm.

[0178] For example, when the current development accuracy of the initial development algorithm does not meet the preset accuracy requirement, the algorithm development module 20 performs algorithm iteration on the initial development algorithm. The data processing module 10 responds to the user instruction and determines the iteration data amount of the iteration data for algorithm iteration.

[0179] According to an example embodiment, the data processing module 10 acquires iterative data of an iterative data amount collected based on a second preset collection requirement to obtain an algorithm iterative data set.

[0180] For example, after obtaining the required iterative data volume, the data processing module 10 obtains the iterative data of the iterative data volume collected by the user according to the second preset collection requirement, and generates an algorithm iterative data set according to the iterative data of the iterative data volume.

[0181] According to an example embodiment, the algorithm development module 20 performs algorithm iteration on the initial development algorithm according to the algorithm iteration data set to generate a target algorithm whose development accuracy meets a preset accuracy requirement.

[0182] For example, the algorithm development module 20 iterates the initial development algorithm multiple times based on the iterative data set until the development accuracy of the generated algorithm reaches the preset accuracy requirement, and then completes the algorithm development, thereby obtaining a target algorithm that meets the user's customized needs.

[0183] Optionally, during the algorithm development process, the algorithm development module 20 may also determine timeliness information such as the remaining time for data collection and the remaining time for algorithm delivery.

[0184] For example, the remaining time for data collection = ((the total amount of data required for algorithm customization - the amount of data currently collected) / the amount of data currently collected) * the collection time of the current amount of data collected.

[0185] The remaining delivery time of the algorithm = the remaining time for data collection + the remaining time for data labeling + the model iteration time + the algorithm logic optimization time.

[0186] Among them, the remaining data labeling time = (the total amount of data required for algorithm customization - the current amount of data collected) / the amount of labeled data per unit time; the model iteration time = the number of iterations * the time required for each iteration; the algorithm logic optimization time = the sum of the time for each optimization point. If the number of optimization points is 0, the algorithm logic optimization time is 0, so the remaining delivery time of the algorithm can be calculated.

[0187] According to an example embodiment, the algorithm development system provided by the present invention is used to execute the algorithm development method described above. The formulas for determining the data amount and the iterative data amount, the first preset collection requirements, the general collection requirements, and the fine collection requirements, as well as the second preset collection requirements, the general collection requirements, and the fine collection requirements have been described in detail above, so they will not be repeated here.

[0188] Through the above embodiments, the present invention obtains the collected data of the amount of data collected based on the first preset collection requirement by determining the data volume of the collected data of the target algorithm to obtain the algorithm development data set. The present invention obtains the initial development algorithm through the algorithm development data set development, and when the development accuracy of the initial development algorithm does not meet the preset accuracy requirement, iterative data of the amount of iterative data collected based on the second preset collection requirement is obtained. The iterative data of the amount of iterative data can complete multiple iterations of the initial development algorithm to obtain the target algorithm that meets the preset accuracy requirement, thereby completing the algorithm development.

[0189] The present invention can accurately obtain an algorithm development data set that meets the algorithm development needs through the precise setting of the data volume and the first preset collection requirement, and can make the actual algorithm development cost closer to the estimated cost to meet user expectations; and the present invention can accurately obtain an algorithm iteration data set that meets the algorithm iteration through the precise setting of the iterative data volume and the second preset collection requirement when the development accuracy of the initial development algorithm does not meet the preset accuracy requirement, so that for the initial development algorithm that does not meet the preset accuracy requirement, the algorithm customization or iteration will be started only when the iterative data meets certain requirements, which can avoid the investment of advance workload or frequent algorithm iterations, and can save the algorithm development cost.

[0190] According to another aspect of the present invention, an electronic device is provided, which includes: one or more processors; and a storage device for storing one or more programs, so that when the one or more programs are executed by the one or more processors, the one or more processors can implement the read and write processing method described above.

[0191] According to another aspect of the present invention, a non-volatile computer-readable storage medium is provided, wherein a computer program is stored on the storage medium, and when the computer program is executed by a processor, the read-write processing method described above can be implemented.

[0192] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention is described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions of the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for developing an algorithm for image detection and recognition, characterized in that: include: Determining the amount of collected data of a target algorithm, wherein the collected data at least includes image data; Acquire the amount of collected data collected based on the first preset collection requirement to obtain an algorithm development data set; Obtain an algorithm-labeled dataset based on the algorithm development dataset; Performing algorithm development based on the algorithm annotated data set to obtain an initial development algorithm; When the development accuracy of the initial development algorithm meets the preset accuracy requirement, generating the target algorithm; In the case where the development accuracy of the initial development algorithm does not meet the preset accuracy requirement, determining the iterative data amount of the iterative data of the target algorithm; Acquire iterative data of the iterative data amount collected based on the second preset collection requirement to obtain an algorithm iterative data set; Performing algorithm iteration on the initial development algorithm according to the algorithm iteration data set to generate a target algorithm whose development accuracy meets the preset accuracy requirement; The amount of data collected is: ； ； ； ； Wherein, N is the amount of data collected, N1 is the amount of positive sample data, and N2 is the amount of negative sample data; A is the number of target object categories, which is the sum of subcategories of all target object categories; B is the number of camera points in the scene that are not repeated; C is the number of target object categories that appear simultaneously in each scene; K1 is the percentage of non-simulated data for each point; D1 is the number of non-simulated collections for each point; K2 is the percentage of simulated data for each point; D2 is the number of simulated collections for each point; K1+K2=1, and K1≥0, K2≥0; Q1 is the number of target object categories that appear simultaneously for each point, and Q2 is the number of collections for each target category that appears simultaneously; Q3 is the number of target object categories that can be simulated simultaneously; Q4 is the number of collections for each target category that can be simulated simultaneously; For the number of non-analog acquisitions at each point, the first preset acquisition requirement also includes a detailed acquisition requirement, including: The number of each simultaneously appearing target object category collection includes at least the number of perspective collections collected at different perspectives, the number of light collections collected at different light rays, the number of distance collections collected at different distances, the number of orientation collections collected at different orientations, the number of quality collections collected at different qualities, and the number of posture collections collected at different postures; Regarding the number of simulated acquisitions at each point, the first preset acquisition requirement also includes a detailed acquisition requirement, including: The number of target object categories that can be simulated simultaneously includes at least the number of perspective collections for collections at different perspectives, the number of light collections for collections at different light rays, the number of distance collections for collections at different distances, the number of orientation collections for collections at different orientations, the number of quality collections for collections at different qualities, and the number of posture collections for collections at different postures.

2. The algorithm development method according to claim 1, characterized in that: The first preset collection requirement includes a general collection requirement, including: For the collected data of simulated collection and non-simulated collection, data is collected for each target object category, and each target object category has at least one target object; and ； ； Among them, F1 is the average number of each target category; F2 is the deviation ratio of the number of each target category, 0<F2<1.

3. The algorithm development method according to claim 1, characterized in that: ； It is the number of viewing angle collections × the number of light collections × the number of distance collections × the number of orientation collections × the number of quality collections × the number of posture collections; ;or 。 4. The algorithm development method according to claim 1, characterized in that: For the amount of negative sample data, the first preset collection requirement also includes a detailed collection requirement, including: The amount of negative sample data includes at least the number of light collections of different light collections; Among them, in the application scenario of classification tasks, the ratio of the amount of positive sample data to the amount of negative sample data at each point is 1:1; In the application scenario of non-classification tasks, the ratio of positive sample data to negative sample data at each point is 8:

2.

5. The algorithm development method according to claim 1, characterized in that: In the case of data misreporting, the iterative data volume of the iterative data is: ； Among them, M1 is the iterative data volume of the iterative data in the case of data false alarm; P1 is the number of false alarm categories; P2 is the number of false alarm points; P3 is the number of false alarm categories that appear at each point at the same time; P4 is the number of samples collected at each point; In the case of data underreporting, the iterative data volume of the iterative data is: ； ； ； Among them, M2 is the amount of iterative data of the iterative data in the case of data omission; R1 is the number of omission target object categories, which is the sum of the omission subcategories of all target object categories; R2 is the number of omission points; R3 is the number of omission target object categories that appear simultaneously at each point; D3 is the number of non-simulated collections at each point in the case of data omission; Q5 is the number of omission target object categories that appear simultaneously for each scene or point; Q6 is the number of collections for each omission target category that appears simultaneously; D4 is the number of simulated collections at each point in the case of data omission; Q7 is the number of omission target object categories that can be simulated simultaneously; Q8 is the number of collections for each omission target category that can be simulated simultaneously.

6. The algorithm development method according to claim 5, characterized in that: In the case of data misreporting, the second preset collection requirements include general collection requirements, including: For the collected data collected without simulation, data is collected for each false alarm target object category, and each false alarm target object category has at least one target object; and ； ； Where W1 is the average number of false positives in each category; W2 is the deviation ratio of the number of false positives in each category, 0<W2<1; In the case of data omission, the second preset collection requirement includes a general collection requirement, including: For the collected data of simulated collection and non-simulated collection, data is collected for each missed target object category, and each missed target object category has at least one target object; and ； ； Wherein, W3 is the average number of each category of missed target objects; W4 is the deviation ratio of the number of each category of missed target objects, 0<W4<1.

7. The algorithm development method according to claim 6, characterized in that: In the case of data omission, for the number of non-analog collections at each point, the second preset collection requirement also includes a detailed collection requirement, including: The number of acquisitions of each category of missed target objects at the same time includes at least the number of perspective acquisitions acquired at different perspectives, the number of light acquisitions acquired at different light rays, the number of distance acquisitions acquired at different distances, the number of orientation acquisitions acquired at different orientations, the number of quality acquisitions acquired at different qualities, and the number of posture acquisitions acquired at different postures; Among them, in the case of data underreporting: ； It is the number of viewing angle collections × the number of light collections × the number of distance collections × the number of orientation collections × the number of quality collections × the number of posture collections; In the case of data omission, for the number of simulated collections at each point, the second preset collection requirement also includes a detailed collection requirement, including: The number of each missed target object category collection that can be simulated simultaneously includes at least the number of perspective collections collected at different perspectives, the number of light collections collected at different light rays, the number of distance collections collected at different distances, the number of orientation collections collected at different orientations, the number of quality collections collected at different qualities, and the number of posture collections collected at different postures; in: ； It is the number of viewing angle collections × the number of light collections × the number of distance collections × the number of direction collections × the number of quality collections × the number of posture collections.

8. An algorithm development system for image detection and recognition, characterized in that: The algorithm development system is used to execute the algorithm development method according to any one of claims 1 to 7, and the algorithm development system comprises: A data processing module determines the amount of collected data of a target algorithm; obtains the collected data of the amount collected based on the first preset collection requirement to obtain an algorithm development data set; and obtains an algorithm annotation data set based on the algorithm development data set; An algorithm development module performs algorithm development based on the algorithm annotation data set to obtain an initial development algorithm; and generates the target algorithm when the development accuracy of the initial development algorithm meets the preset accuracy requirement; The data processing module determines the iterative data volume of the collected data of the target algorithm when the development accuracy of the initial development algorithm does not meet the preset accuracy requirement; and obtains the iterative data of the iterative data volume collected based on the second preset collection requirement to obtain an algorithm iteration data set; The algorithm development module performs algorithm iteration on the initial development algorithm according to the algorithm iteration data set to generate a target algorithm whose development accuracy meets the preset accuracy requirement.

9. An electronic device, characterized in that: include: one or more processors; A storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the algorithm development method according to any one of claims 1 to 7.

Images