Classifiers

The attributes of vehicles or license plates (LPs) are determined by classifiers. This section describes the available classifiers in the CARS API and their fields.

Vehicle classifiers

This section describes the available vehicle classifiers in the CARS API and their fields.

«car_brand_model_v2»

This classifier recognizes the vehicle brand and model. The response includes the brand and model names in Latin characters, along with the recognition accuracy score (Table 4).

The «car_brand_model_v2» classifier supports the recognition of over 160 brands and more than 1700 models.

The complete list of supported brands and models is provided in Appendix 4, Table 1.

Vehicle brand and model recognition in the system is based on the vehicle's body appearance. When using the «car_brand_model_v2» classifier, vehicles of different brands and models but with similar body appearances are classified into one brand and model (group), which is the most frequently encountered in Russia. For example, ZAZ Sens, ZAZ Lanos, ZAZ Chance, Daewoo Sens, Daewoo Lanos, Chevrolet Lanos, and ZAZ Lanos Furgon have similar appearances and are grouped together, and CARS API identifies these models as Chevrolet Lanos. Therefore, when an image of any of these vehicles is sent, the «Brand» and «Model» fields will display «Chevrolet» and «Lanos» respectively.

The complete list of related brand and model groupings is provided in Appendix 4, Table 2.

An example of an input image is shown in Figure 3.

Example input image for «car_brand_model_v2» classifier

Response example:

{
    "brand": "Skoda",
    "classifier": "car_brand_model_v2",
    "execution_time": 84,
    "model": "Kodiaq",
    "score": 0.9844
}

Table 4. Description of the fields in the «car_brand_model_v2» classifier

Field name	Type	Description	Possible values
brand	string	Vehicle brand	Skoda
classifier	string	Name of classifier	car_brand_model_v2
execution_time	int	Execution time in milliseconds	0…1000
model	string	Vehicle model	Kodiaq
score	float	Assessment of accuracy recognition of vehicle brand and model	0.0000…1.0000

«detailed_vehicle_color»

The classifier recognizes the color type and color of the vehicle. The response includes the color type, the vehicle color name in English, and the confidence score for recognition (Table 5).

The classifier supports recognition of 12 colors and 4 types of color schemes. The model classifies the vehicle's color scheme into 4 categories: monochrome, multicolor, with stickers/decals/paintings, or with color graphics.

An example of an input image is shown in Figure 4.

Example input image for «detailed_vehicle_color» classifier

Response example:

{
    "classifier": "detailed_vehicle_color",
    "execution_time": 20,
    "color_type": "pics",
    "color_type_score": 0.69997,
    "colors": [
        {
            "vehicle_color": "white",
            "vehicle_color_score": 0.9986
        }
    ],
    "execution_time": 87
}

Table 5. Description of the fields in the «detailed_vehicle_color» classifier

Field name	Type	Description	Possible values
classifier	string	Name of classifier	detailed_vehicle_color
execution_time	int	Execution time in milliseconds	0…1000
color_type	string	Vehicle color scheme type	mono - monochrome;
			multi - multicolor;
			colorgraph - with stickers/decals/paintings;
			pics - with color graphics
color_type_score	float	Confidence score for the color scheme recognition	0.0000…1.0000
colors		Array containing information about the vehicle color and the color recognition score	vehicle_color, vehicle_color_score
vehicle_color	string	Vehicle color	red, orange, yellow, green, light_blue, blue, purple_or_pink, black, white, brown, grey_or_silver, beige
vehicle_color_score	float	Confidence score for the vehicle color recognition	0.0000…1.0000
execution_time	int	Execution time in milliseconds	0…1000

«vehicle_type»

This classifier recognizes the type of vehicle. It returns the vehicle type and recognition accuracy score in the response (Table 7).

Vehicle classification is carried out according to GOST R 52051-2003 «Motor vehicles and trailers. Classification and definitions.» The description of each vehicle type is presented in Table 6.

Table 6. Supported vehicle types

Vehicle type	Description	Standard type	Example
A_light	Vehicles that do not require a special driving license	-	Bicycles, mopeds, gyroscooters, scooters, electric scooters, etc.
А_heavy	Motorized vehicles that require a special driving license	L1 – L7	Motorcycles, tricycles, quad bikes, etc.
B_light	Passenger vehicles used for transporting passengers, with no more than eight seats (excluding the driver)	M1	Sedans, SUVs, minivans, microbuses, etc., including pickups with enclosed cargo beds.
B_heavy	Passenger vehicles used for transporting goods	N1	Pickups without an enclosed cargo bed, trucks weighing less than 3.5 tons, motorhomes.
C_light	Trucks with a maximum weight of 3.5 to 12 tons	N2	Trucks except semi-trailers trucks
C_heavy	Trucks with a maximum weight of 3.5 to 12 tons	N3	Trucks except semi-trailers trucks
D_light	Passenger vehicles that have more than eight seats (excluding the driver) with a maximum weight of up to 5 tons	M2	Buses
D_heavy	Passenger vehicles that have more than eight seats (excluding the driver) with a maximum weight of more than 5 tons	M3	Buses and trolleybuses, excluding articulated ones
D_long	D_heavy with one or more articulated sections	M3	Articulated buses and trolleybuses
E_light	Trucks with a maximum weight of more than 3.5 tons	N2, N3	Trucks without trailers
E_heavy	Trucks with a maximum weight of more than 3.5 tons and a trailer weight of more than 0.75 tons	Vehicle: N2, N3	Articulated trucks with semi-trailers and road trains
		Trailer: O2 – O4
P_light	Trailers with a maximum weight of no more than 0.75 tons	O1, O2	Trailers used by passenger vehicles
P_heavy	Trailers with a maximum weight of more than 0.75 tons	O2 – O4	Trailers used by cargo vehicles
Other	Other vehicles	T, G, and without category	Agricultural machinery, off-road vehicles, trams, etc.

An example of an input image is shown in Figure 5.

Example input image for «vehicle_type» classifier

Response example:

{
    "classifier": "vehicle_type",
    "execution_time": 20,
    "ts_type_ai": "E_light",
    "ts_type_ai_score": 0.9997
}

Table 7. Description of the fields in the «vehicle_type» classifier

Field name	Type	Description	Possible values
classifier	string	Name of classifier	vehicle_type
execution_time	int	Execution time in milliseconds	0…1000
ts_type_ai	string	Vehicle type	А_light, А_heavy, В_light, B_heavy, С_light, С_heavy, D_light, D_heavy, D_long, E_light, E_heavy, P_light, P_heavy, Other
ts_type_ai_score	float	Assessment of the accuracy of determining the vehicle type	0.0000…1.0000

«vehicle_position_v1»

The classifier recognizes the position of a vehicle in the image by returning the coordinates of key points (leftFront: front left corner, leftRear: rear left corner, rightFront: front right corner, rightRear: rear right corner).

An example of an input image is shown in Figure 6.

Example input image for «vehicle_position_v1» classifier

Example response:

{
  "classifier": "vehicle_position_v1",
  "execution_time": 20,
  "leftFront": {
    "x": 13.9383,
    "y": 291.3493
  },
  "leftRear": {
    "x": 344.9959,
    "y": 472.0356
  },
  "rightFront": {
    "x": 174.0835,
    "y": 260.4908
  },
  "rightRear": {
    "x": 565.8995,
    "y": 442.9853
  }
}

Table 8. Description of the «vehicle_position_v1» classifier fields

Field	Type	Description	Possible values
classifier	string	Name of classifier	vehicle_position_v1
execution_time	int	Execution time in milliseconds	0…1000
leftFront	object	Coordinates of the front left corner	{x: float, y: float}
leftRear	object	Coordinates of the rear left corner	{x: float, y: float}
rightFront	object	Coordinates of the front right corner	{x: float, y: float}
rightRear	object	Coordinates of the rear right corner	{x: float, y: float}

«vehicle_orientation_v1»

The classifier recognizes the orientation of a vehicle in an image, returning 8 possible values that represent the vehicle's orientation relative to the camera. These orientations cover all major angles: front, back, left, right, as well as their combinations.

An example of an input image is shown in Figure 7.

Example input image for «vehicle_orientation_v1» classifier

Example Response:

{
    "classifier": "vehicle_orientation_v1",
    "execution_time": 20,
    "orientation": "front_right",
    "orientation_score":  0.8609
}

Table 9. Description of the fields in the «vehicle_orientation_v1» classifier

Field	Type	Description	Possible values
classifier	string	Name of classifier	vehicle_orientation_v1
execution_time	int	Execution time in milliseconds	0…1000
orientation	string	Vehicle orientation	front, back, left, right, front_right, front_left, back_left, back_right
orientation_score	float	Assessment of orientation recognition accuracy	0.0000…1.0000

«detailed_vehicle_emergency» и «detailed_vehicle_emergency_v2»

The classifier is used to determine the affiliation of a vehicle to emergency services and the presence of flashing lights. The classifier analyzes vehicle images, classifies the type of emergency service, and detects the presence of a flashing light.

An example of an input image is shown in Figure 8.

Example input image for «detailed_vehicle_emergency_v2» classifier

Response example:

{
    "classifier": "detailed_vehicle_emergency_v2",
    "execution_time": 20, 
    "emergency_type": "01", 
    "emergency_type_score": 0.6714, 
    "flashing_light": "without_flashing_light", 
    "flashing_light_score": 0.0896
}

Table 10. Description of the fields in the «detailed_vehicle_emergency_v2» classifier

Field	Type	Description	Possible values
classifier	string	Name of classifier	detailed_vehicle_emergency_v2
execution_time	int	Execution time in milliseconds	0…1000
emergency_type	string	Type of emergency service	01 – Fire department
			02 – Police
			03 – Ambulance
			04 — Emergency services
			05 — EMERCOM
			112 – Emergency
			not_emergency – Vehicles not belonging to emergency services (passenger, military vehicle etc.)
emergency_type_score	float	Assessment of the emergency type recognition accuracy	0.0000…1.0000
flashing_light	string	Presence of flashing lights	with_flashing_light / without_flashing_light
flashing_light_score	float	Assessment of flashing light detection accuracy	0.0000…1.0000

«public_transport_type»

The classifier recognizes the type of public transport by returning the type of public transport in the response and an estimate of the recognition accuracy.

Taxis are considered public transport if their color schemes comply with GOST R 58287-2018 «Distinctive Signs and Information Support for Passenger Ground Transport Vehicles, Stop Points, and Bus Stations. General Technical Requirements». Carsharing is considered public transport if their color schemes comply with the order of the Moscow Department of Transport dated September 2, 2015, No. 61-02-283/5 «On Approval of the Requirements for the Color Scheme of Vehicles Used for Car Sharing Services».

An example of an input image is shown in Figure 9.

Example input image for «public_transport_type» classifier

Response example:

    {
      "classifier": "public_transport_type",
      "execution_time": 20,
      "public_type": "taxi",
      "public_type_score": 0.9999
    }

Table 11. Description of the fields in the «public_transport_type» classifier

Field name	Type	Description	Possible values
classifier	string	Name of classifier	public_transport_type
execution_time	int	Execution time in milliseconds	0…1000
public_type	string	Public transport category	- route_transport,
			- taxi,
			- carshering,
			- other
public_type_score	float	Assessment of accuracy recognition	0.0000…1.0000

«special_transport_type»

The classifier recognizes the type of special transport by returning the type of special transport in the response and an estimate of the recognition accuracy.

An example of an input image is shown in Figure 10.

Example input image for «special_transport_type» classifier

Response example:

    {
      "classifier": "special_transport_type",
      "execution_time": 20,
      "special_type": "garbage_truck",
      "special_type_score": 0.9999
  }

Table 12. Description of the fields in the «special_transport_type» classifier

Field name	Type	Description	Possible values
classifier	string	Name of classifier	special_transport_type
execution_time	int	Execution time in milliseconds	0…1000
special_type	string	Special transport category	- forklift,
			- paver,
			- bulldozer,
			- grader,
			- roller,
			- truck_crane,
			- concrete_mixer,
			- tractor,
			- excavator,
			- sweeper,
			- garbage_truck,
			- sprinkler_truck,
			- dump_truck,
			- truck,
			- other_special,
			- other_non_special
special_type_score	float	Assessment of accuracy recognition	0.0000…1.0000

«vehicle_axles»

The classifier detects the vehicle's axles: their coordinates and count. For each axle, the detection accuracy is determined, and the overall detection accuracy for all axles is provided.

An example of an input image is shown in Figure 11.

Example input image for «vehicle_axles» classifier

Response example:

{
    "classifier": "vehicle_axles",
    "execution_time": 20,
    "aggregatedScore": 0.9006090760231018,
    "detections": [
        {
            "height": 119.57168579101563,
            "score": 0.917134165763855,
            "width": 109.78546142578125,
            "x": 180.5765380859375,
            "y": 385.35791015625
        },
        {
            "height": 93.49630737304688,
            "score": 0.8918288946151733,
            "width": 66.70538330078125,
            "x": 565.5817260742188,
            "y": 359.1695861816406
        },
        {
            "height": 91.6329345703125,
            "score": 0.8791476488113403,
            "width": 84.33480834960938,
            "x": 477.0045471191406,
            "y": 372.5871276855469
        }
    ],
    "numberOfWheelAxles": 3
}

Table 13. Description of the fields in the «vehicle_axles» classifier

Field name	Type	Description	Possible values
classifier	string	Name of classifier	vehicle_axles
execution_time	int	Execution time in milliseconds	0…1000
aggregatedScore	float	Overall detection accuracy for all vehicle axles	0…1
detections	array	Array containing coordinates and size of each axle detection on the image, as well as the detection accuracy for each axle	-
height	int	Height of the axle's BBox (in pixels)	0…1080
score	float	Detection accuracy for the axle	0.0000…1.0000
width	int	Width of the axle's BBox (in pixels)	0…1920
x	int	Horizontal coordinate of the top-left corner of the axle's BBox	0…1920
y	int	Vertical coordinate of the top-left corner of the axle's BBox	0…1080
numberOfWheelAxles	int	Number of detected axles	1…n

«vehicle_descriptor_v2»

The classifier extracts the vehicle descriptor.

This descriptor is used for searching vehicles by image in the CARS Analytics interface. Image is encoded in Base64 format. Image-based search is especially useful when it is necessary to find records in the archive related to a specific vehicle.

More details on the process of vehicle search by image can be found in the «CARS Analytics. User Guide». The vehicle descriptor can be up to several thousand characters long.

An example of an input image is shown in Figure 12.

Example input image for «vehicle_descriptor_v2» classifier

Response example:

        {
            "classifier": "vehicle_descriptor_v2",
            "execution_time": 23,
            "vehicle_descriptor": "AACAvQAAALwAAAA9..."
        }

Table 14. Description of the fields in the «vehicle_descriptor_v2» classifier

Field name	Type	Description	Possible values
classifier	string	Name of classifier	vehicle_descriptor_v2
execution_time	int	Execution time in milliseconds	0…1000
vehicle_descriptor	string	Vehicle descriptor	AACAvQ...

«vehicle_tracks_reid_descriptor»

The classifier is used to detect track breaks of the vehicle and allows continuing the existing track instead of creating a new one.

It extracts the vehicle descriptor, which is necessary for searching the vehicle by image in the CARS Analytics interface. Image is encoded in Base64 format. Image-based search helps find records in the archive related to a specific vehicle.

More information about the process of searching for a vehicle by image can be found in the «CARS Analytics. User Guide». The vehicle descriptor can be several thousand characters long.

The «vehicle_tracks_reid_descriptor» classifier works faster than «vehicle_descriptor_v2», but its accuracy is slightly lower.

An example of an input image is shown in Figure 13.

Example input image for «vehicle_tracks_reid_descriptor» classifier

Response example:

        {
            "classifier": "vehicle_tracks_reid_descriptor",
            "execution_time": 20,
            "vehicle_descriptor": "AACAvQAAALwAAAA9..."
        }

Table 15. Description of the fields in the «vehicle_tracks_reid_descriptor» classifierr

Field name	Type	Description	Possible values
classifier	string	Name of classifier	vehicle_tracks_reid_descriptor
execution_time	int	Execution time in milliseconds	0…1000
vehicle_descriptor	string	Vehicle descriptor	AACAvQ...

License plate classifiers

This section describes the available LP classifiers in the CARS API and their fields.

Operating logic of LP recognition networks

The logic scheme of LP recognition networks is shown in Figure 14.

LP networks operation logic

The description of the scheme is provided in Table 16.

Table 16. Description of the operating logic of LP recognition networks

Step	Description
0	The CARS API receives a frame containing an LP number.
1	The LP country recognition classifier «grz_all_countries» is launched.
2	The country of affiliation of the LP is determined.
3	If the country of affiliation is NOT defined, the «eu_cis_recognition_v4» classifier is launched.
4	If the country of affiliation is determined, the corresponding classifier from Table 17 is launched.

Table 17. Correspondence of classifiers and countries of LP affiliation:

Classifier	Country of LP affiliation
rus_plate_recognition_v3	Russian Federation, Abkhazia, South Ossetia, DPR and LPR.
grz_country_recognition_v4, grz_all_countries	Russia, Abkhazia, South Ossetia, DPR, LPR, Belarus, Kazakhstan, Ukraine, Azerbaijan, Armenia, Georgia, Kyrgyzstan, Moldova, Tajikistan, Turkmenistan, Uzbekistan, Turkey, Austria, Andorra, Albania, Belgium, Bulgaria, Bosnia and Herzegovina, Vatican, Hungary, Germany, Greece, Denmark, Ireland, Spain, Italy, Cyprus, Kosovo, Latvia, Lithuania, Liechtenstein, Luxembourg, Malta, Monaco, Netherlands, Norway, Poland, Portugal, Transnistria, Romania, San Marino, North Macedonia, Serbia, Slovakia, Slovenia, Finland, France, Croatia, Montenegro, Czech Republic, Switzerland, Sweden, Estonia, Israel, Iran, Palestine, Pakistan, Qatar, Oman, Kuwait, Bahrain, Saudi Arabia, United Kingdom, Singapore, Hong Kong, Macau, Indonesia, Vietnam, Malaysia, Sri Lanka, Taiwan, Philippines, Thailand, UAE, India, China, Mongolia, Japan, South Korea, Brazil, Bangladesh, Canada, USA, Mexico
uae_plate_recognition_v2	UAE
eu_cis_recognition_v3, eu_cis_recognition_v4	CIS and EU countries. Ukraine, Kazakhstan, Belarus, Georgia, Moldova, Armenia, Azerbaijan, Tajikistan, Turkmenistan, Kyrgyzstan, Uzbekistan, Andorra, Albania, Austria, Bosnia and Herzegovina, Belgium, Bulgaria, Switzerland, Cyprus, Czech Republic, Germany, Denmark, Estonia, Spain, Finland, France, Greece, Croatia, Hungary, Ireland, Israel, Palestine, Guernsey, Iran, Italy, Liechtenstein, Lithuania, Luxembourg, Latvia, Monaco, Montenegro, North Macedonia, Malta, Netherlands, Norway, Poland, Transnistria, Portugal, Romania, Serbia, Sweden, Slovenia, Slovakia, San Marino, Turkey, Vatican, Kosovo
emirate_recognition_v1	UAE
grz_uk_asia_recognition_v2	India, Pakistan, Vietnam, Indonesia, Malaysia, Taiwan, Philippines, Sri Lanka, Macau, Singapore, Hong Kong, the United Kingdom
chn_plate_recognition_v2	China
mng_recognition_v1	Mongolia
america_plate_recognition_v1	USA, Canada, Mexico, Brazil
persian_plate_recognition_v1	Bahrain, Kuwait, Oman, Qatar, Saudi Arabia
thailand_plate_recognition_v1	Thailand

«license_plate_ags_v1»

The classifier evaluates the quality of the LP image. It is used to optimize selection of the best frame.

To filter LPs images by quality, you need to set upper and lower score threshold values. More information about the process of filtering LPs images by quality can be found in the «CARS Analytics. User Guide».

An example of an input image is shown in Figure 15.

Example input image for «license_plate_ags_v1» classifier

Response example:

        {
            "classifier": "license_plate_ags_v1",
            "execution_time": 20,
            "score": 0.9963796734809875,
        }

Table 17. Description of the fields in the ««license_plate_ags_v1» classifier

Field name	Type	Description	Possible values
classifier	string	Name of classifier	license_plate_ags_v1
execution_time	int	Execution time in milliseconds	0…1000
score	float	Estimation of LPs image quality	0.0000…1.0000

«rus_plate_recognition_v3»

The classifier recognizes LP symbols and their features, evaluates the accuracy of their recognition, and provides an score accuracy of LP recognition.

Classifier «rus_plate_recognition_v3» is available for LP images from Russia, Abkhazia, South Ossetia, the DPR, and LPR. A detailed list of supported LPs is provided in Appendix 3.

An example of an input image is shown in Figure 16.

Example input image for «rus_plate_recognition_v3» classifier

Response example:

        {
            "classifier": "rus_plate_recognition_v3",
            "execution_time": 20,
            "features": [
                {
                    "score": 0.9998,
                    "type": "rus_spec_type",
                    "value": "regular"
                },
                {
                    "score": 0.9987,
                    "type": "rus_region_number",
                    "value": "94"
                }
            ],
            "regno_ai_score": 0.9959,
            "symbol_scores": [
                0.9998,
                0.9996,
                0.9994,
                0.9991,
                0.9996, 
                0.9996,
                0.9994,
                0.9994
            ],
            "symbols": [
                "K", 
                "2", 
                "1", 
                "2", 
                "K", 
                "K", 
                "9", 
                "4"
            ]
        }

Table 18. Description of the fields in the ««rus_plate_recognition_v3» classifier

Field name	Type	Description	Possible values
classifier	string	Name of classifier	rus_plate_recognition_v3
execution_time	int	Execution time in milliseconds	0…1000
score	float	Accuracy score for recognizing LP features	0.0000…1.0000
type	string	Type of LP feature	- rus_spec_type – LP format
value	string	LP feature	- regular
			- police
			- diplomatic
			- military
			- machines&moto
			- trailer
			- transit
			- taxi
score	float	Accuracy score for recognizing the LP region	0.0000…1.0000
type	string	LP region	- rus_region_number – vehicle registration region
value	string	Array of recognized LP symbols	0…9
regno_ai_score	float	Accuracy score for LP recognition	0.0000…1.0000
symbol_scores	float	Accuracy score for recognizing each LP symbol	0.0000…1.0000
symbols	string	Array of recognized LP symbols	0…9; A, B, C, E…

«grz_country_recognition_v4»

The classifier recognizes the country of the LP and evaluates the accuracy of recognition.

The classifier separates boxes into «relevant» and «irrelevant» to keep only relevant plates. It determines whether the box belongs to the main car or a neighboring one. All plates on a single car are considered «relevant». A detailed list of possible values for the country parameter is provided in Appendix 3.

An example of an input image is shown in Figure 17.

Example input image for «grz_country_recognition_v4» classifier

Response example:

{
   classifier:  grz_country_recognition,
   "execution_time": 20,
   country: RUS,
   country_score : 0.9718532562255859
}

Table 19. Description of the fields in the «grz_country_recognition_v4» classifier

Field name	Type	Description	Possible values
classifier	string	Name of classifier	«grz_country_recognition_v4»
execution_time	int	Execution time in milliseconds	0…1000
country	string	Country of the LP	RUS – Russian Federation (RF)
country_score	float	Accuracy score for country recognition	0.0000…1.0000

«grz_all_countries»

The classifier recognizes the country of registration of the LP and their features, evaluates the accuracy of their recognition, and provides an score accuracy  of LP recognition.

A detailed list of supported LPs is provided in Appendix 3.

An example of an input image is shown in Figure 18.

Example input image for «grz_all_countries» classifier

Response example:

        {
            "classifier": "grz_all_countries",
            "execution_time": 20,
            "features": [
                {
                    "score": 0.9998,
                    "type": "rus_spec_type",
                    "value": "regular"
                },
                {
                    "score": 0.9987,
                    "type": "rus_region_number",
                    "value": "94"
                }
            ],
            "regno_ai_score": 0.9959,
            "symbol_scores": [
                0.9998,
                0.9996,
                0.9994,
                0.9991,
                0.9996, 
                0.9996,
                0.9994,
                0.9994
            ],
            "symbols": [
                "K", 
                "2", 
                "1", 
                "2", 
                "K", 
                "K", 
                "9", 
                "4"
            ]
        }

Table 20. Description of the fields in the «grz_all_countries» classifier

Field name	Type	Description	Possible values
classifier	string	Name of classifier	grz_all_countries
execution_time	int	Execution time in milliseconds	0…1000
score	float	Accuracy score for recognizing LP features	0.0000…1.0000
type	string	Type of LP feature	- rus_spec_type – LP format
value	string	LP feature	- regular
			- police
			- diplomatic
			- military
			- machines&moto
			- trailer
			- transit
			- taxi
score	float	Accuracy score for recognizing the LP region	0.0000…1.0000
type	string	LP region	- rus_region_number – vehicle registration region
value	string	Array of recognized LP symbols	0…9
regno_ai_score	float	Accuracy score for LP recognition	0.0000…1.0000
symbol_scores	float	Accuracy score for recognizing each LP symbol	0.0000…1.0000
symbols	string	Array of recognized LP symbols	0…9; A, B, C, E…

«uae_plate_recognition_v2»

The classifier recognizes the emirate of registration of the LP, the symbols of the LP and the evaluation of the accuracy of their recognition.

A detailed list of supported LPs is provided in Appendix 3.

An example of an input image is shown in Figure 19.

Example input image for «uae_plate_recognition_v2» classifier

Response example:

{
    "classifier": "uae_plate_recognition_v2",
    "execution_time": 43,
    "features": [
        {
            "score": 0.848,
            "type": "emirate_name",
            "value": "SHARJAH"
        }
    ],
    "regno_ai_score": 0.5661,
    "symbol_scores": [
        0.9044,
        0.9116,
        0.9182,
        0.9042,
        0.9087,
        0.9101
    ],
    "symbols": [
        "2",
        "1",
        "4",
        "5",
        "6",
        "7"
    ]
}

Table 21. Description of the fields in the «uae_plate_recognition_v2» classifier

Field name	Type	Description	Possible values
classifier	string	Name of classifier	uae_plate_recognition_v2
execution_time	int	Execution time in milliseconds	0…1000
features	array	List of LP features	emirate_name
score	float	Assessment of the accuracy of determining the emirate	0.0000…1.0000
type	string	Feature type	emirate_name
value	string	Recognition result	DUBAI, ABU_DHABI, SHARJAH, KHAIMAN, QUWAIN, AJMAN, FUJAIRAH
regno_ai_score	float	General assessment of the accuracy of LP recognition	0.0000…1.0000
symbol_scores	float	Accuracy score for recognizing each LP symbol	0.0000…1.0000
symbols	string	Array of recognized LP symbols. Symbols for recognition include:	0…9; A, B, C, E…
		- series;
		- LP symbols;
		- region of registration code/name

«eu_cis_recognition_v3» и «eu_cis_recognition_v4»

The classifiers recognize LP symbols from CIS countries and the European Union, as well as individual country plates within the EU, and evaluate the accuracy of their recognition.

The «eu_cis_recognition_v4» classifier is capable of recognizing more types of LP.

A detailed list of supported LPs is provided in Appendix 3.

An example of an input image is shown in Figure 20.

Example input image for «eu_cis_recognition_v4» classifier

Response example:

{
  "classifier": "eu_cis_recognition_v4",
  "execution_time": 20,
  "regno_ai": {
    "length_scores": [
      0.0, 
      0.0, 
      0.0, 
      0.0, 
      0.0, 
      1.0, 
      0.0, 
      0.0, 
      0.0, 
      0.0, 
      0.0
    ],
    "scores": [
      0.9137,
      0.9112,
      0.9082,
      0.9054,
      0.9043
    ],
    "symbols": [
      "K",
      "U",
      "H",
      "L",
      "1"
    ]
  },
  "regno_ai_score": 0.6191
}

Table 22. Description of the fields in the «eu_cis_recognition_v4» classifier

Field name	Type	Description	Possible values
classifier	string	Name of classifier	eu_cis_recognition_v4
length_scores	float	An array of 11 elements showing the probability of the number of characters in the LP: - The 1st element corresponds to 0 characters, the 11th to 10 characters. The sum of all elements equals 1. The highest value indicates the number of characters in the LP.	0.0000…1.0000
scores	float	An array displaying the assessment of the recognition accuracy of each LP symbol	0.0000…1.0000
symbols	string	Array of recognized LP symbols	0…9; A, B, C, E…
regno_ai_score	float	Assessment of the recognition accuracy of the LP	0.0000…1.0000

«grz_uk_asia_recognition_v2»

The classifier recognizes LP symbols of Asian and British countries, individual numbers of Asian countries, and also evaluates the accuracy of their recognition.

A detailed list of supported LPs is provided in Appendix 3.

An example of an input image is shown in Figure 21.

Example input image for «grz_uk_asia_recognition_v2» classifier

Response example:

{
    "classifier": "grz_uk_asia_recognition_v2",
    "execution_time": 55,
    "regno_ai": {
        "length_scores": [
            0.0,
            0.0,
            0.0,
            0.0,
            0.0,
            0.0,
            0.0,
            0.0,
            0.0,
            0.0,
            1.0
        ],
        "scores": [
            1.0,
            1.0,
            1.0,
            1.0,
            1.0,
            1.0,
            1.0,
            1.0
            1.0
            1.0
        ],
        "symbols": [
            "G",
            "J",
            "0",
            "8",
            "C",
            "G",
            "2",
            "5",
            "3",
            "1",
        ]
    },
    "regno_ai_score": 1.0
}

Table 23. Description of the fields in the «grz_uk_asia_recognition_v2» classifier

Field name	Type	Description	Possible values
classifier	string	Name of classifier	grz_uk_asia_recognition_v2
execution_time	int	Execution time in milliseconds	0…1000
length_scores	float	An array of 11 elements showing the probability of the number of characters in the LP: - The 1st element corresponds to 0 characters, the 11th to 10 characters. The sum of all elements equals 1. The highest value indicates the number of characters in the LP.	0.0000…1.0000
scores	float	An array displaying the assessment of the recognition accuracy of each LP symbol	0.0000…1.0000
symbols	string	Array of recognized LP symbols	0…9; A, B, C, E…
regno_ai_score	float	Assessment of the recognition accuracy of the LP	0.0000…1.0000

«chn_plate_recognition_v2»

The classifier recognizes LP symbols of the China, evaluates the accuracy of recognition of each symbol and the entire LP.

A detailed list of supported LPs is provided in Appendix 3.

An example of an input image is shown in Figure 22.

Example input image for «chn_plate_recognition_v2» classifier

Response example:

{
"classifier": "chn_plate_recognition_v2",
    "execution_time": 24,
    "features": [
        {
            "score": 0.8903,
            "type": "spec_type",
            "value": "regular"
        }
    ],
    "regno_ai_score": 0.4932,
    "symbol_scores": [
        0.8961,
        0.9147,
        0.9064,
        0.9024,
        0.9049,
        0.901,
        0.9023
    ],
    "symbols": [
        "吉",
        "H",
        "A",
        "Y",
        "4",
        "5",
        "7"
    ]
}

Table 24. Description of the fields in the «chn_plate_recognition_v2» classifier

Field name	Type	Description	Possible values
classifier	string	Name of classifier	chn_plate_recognition_v2
execution_time	int	Execution time in milliseconds	0…1000
score	float	Assessment of the recognition accuracy of each feature	0.0000…1.0000
type	string	Type of feature	- spec_type – LP form
value	string	The value of LP feature	- regular
			- large&moto – motorcycles, trucks, buses;
			- electric – electric cars;
			- cross-boarder – numbers for crossing the border with Macau and Hong Kong;
			- trailer;
			- temporary – transit;
			- police;
			- training.
regno_ai_score	float	Assessment of the recognition accuracy of the LP	0.0000…1.0000
symbol_scores	float	Accuracy score for recognizing each LP symbol	0.0000…1.0000
symbols	string	Array of recognized LP symbols	0…9; A, B, C, E…; O, 皖, 沪, 津…

«mng_recognition_v1»

The classifier recognizes LP symbols of the Mongolia, evaluates the accuracy of recognition of each symbol and the entire LP.

A detailed list of supported LPs is provided in Appendix 3.

An example of an input image is shown in Figure 23.

Example input image for ««mng_recognition_v1» classifier

Response example:

{
    "classifier": "mng_recognition_v1",
    "execution_time": 206,
    "regno_ai": {
        "length_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            1,
            0,
            0,
            0,
            0
        ],
        "scores": [
            0.9869,
            0.9909,
            0.9974,
            0.9981,
            0.9529,
            0.9952
        ],
        "symbols": [
            "1",
            "2",
            "3",
            "4",
            "Д",
            "К"
        ]
    },
    "regno_ai_score": 0.9231
}

Table 25. Description of the fields in the ««mng_recognition_v1» classifier

Field name	Type	Description	Possible values
classifier	string	Name of classifier	mng_recognition_v1
execution_time	int	Execution time in milliseconds	0…1000
length_scores	float	An array of 11 elements showing the probability of the number of characters in the LP: - The 1st element corresponds to 0 characters, the 11th to 10 characters. The sum of all elements equals 1. The highest value indicates the number of characters in the LP.	0.0000…1.0000
scores	float	An array displaying the assessment of the recognition accuracy of each LP symbol	0.0000…1.0000
symbols	string	Array of recognized LP symbols	0…9; A, B, C, E…
regno_ai_score	float	Assessment of the recognition accuracy of the LP	0.0000…1.0000

«america_plate_recognition_v1»

The classifier recognizes LP symbols of the USA, Canada, Mexico and Brazil, evaluates the accuracy of recognition of each symbol and the entire LP.

A detailed list of supported LPs is provided in Appendix 3.

An example of an input image is shown in Figure 24.

Example input image for «america_plate_recognition_v1» classifier

Response example:

{
    "classifier": "america_plate_recognition_v1",
    "execution_time": 190,
    "regno_ai": {
        "length_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            1,
            0,
            0,
            0
        ],
        "scores": [
            0.9971,
            0.9943,
            0.9947,
            0.9997,
            0.999,
            0.9989,
            0.9992
        ],
        "symbols": [
            "К",
            "N",
            "М",
            "1",
            "9",
            "5",
            "4"
        ]
    },
    "regno_ai_score": 0.983
}

Table 26. Description of the fields in the «america_plate_recognition_v1» classifier

Field name	Type	Description	Possible values
classifier	string	Name of classifier	america_plate_recognition_v1
execution_time	int	Execution time in milliseconds	0…1000
length_scores	float	An array of 11 elements showing the probability of the number of characters in the LP: - The 1st element corresponds to 0 characters, the 11th to 10 characters. The sum of all elements equals 1. The highest value indicates the number of characters in the LP.	0.0000…1.0000
scores	float	An array displaying the assessment of the recognition accuracy of each LP symbol	0.0000…1.0000
symbols	string	Array of recognized LP symbols	0…9; A, B, C, E…
regno_ai_score	float	Assessment of the recognition accuracy of the LP	0.0000…1.0000

«persian_plate_recognition_v1»

The classifier recognizes LP symbols of of Bahrain, Kuwait, Oman, Oman, Qatar, Saudi Arabia, evaluates the accuracy of recognition of each symbol and the entire LP.

A detailed list of supported LPs is provided in Appendix 3.

An example of an input image is shown in Figure 25.

Example input image for «persian_plate_recognition_v1» classifier

Response example:

{
    "classifier": "persian_plate_recognition_v1",
    "execution_time": 49,
    "regno_ai": {
        "length_scores": [
            0,
            0,
            0,
            0,
            0,
            1,
            0,
            0,
            0,
            0,
            0
        ],
        "scores": [
            0.9047,
            0.9088,
            0.9063,
            0.9099,
            0.9062
        ],
        "symbols": [
            "1",
            "2",
            "3",
            "4",
            "5"
        ]
    },
    "regno_ai_score": 0.6144
}

Table 27. Description of the fields in the «persian_plate_recognition_v1» classifier

Field name	Type	Description	Possible values
classifier	string	Name of classifier	persian_plate_recognition_v1
execution_time	int	Execution time in milliseconds	0…1000
length_scores	float	An array of 11 elements showing the probability of the number of characters in the LP: - The 1st element corresponds to 0 characters, the 11th to 10 characters. The sum of all elements equals 1. The highest value indicates the number of characters in the LP.	0.0000…1.0000
scores	float	An array displaying the assessment of the recognition accuracy of each LP symbol	0.0000…1.0000
symbols	string	Array of recognized LP symbols	0…9; A, B, C, E…
regno_ai_score	float	Assessment of the recognition accuracy of the LP	0.0000…1.0000

«thailand_plate_recognition_v1»

The classifier recognizes the region of registration of the LP, the symbols of the LP, individual numbers of Thailand, evaluates the accuracy of recognition of each symbol and the entire LP.

A detailed list of supported LPs is provided in Appendix 3.

An example of an input image is shown in Figure 26.

Example input image for «thailand_plate_recognition_v1» classifier

Response example:

{
  "classifier": "thailand_plate_recognition_v1",
  "execution_time": 20,
  "features": [
    {
      "score": 0.9891,
      "type": "region",
      "value": "Ratchaburi"
    },
    {
      "score": 0.9899,
      "type": "spec_type",
      "value": "private"
    }
  ],
  "regno_ai_score": 0.991,
  "symbol_scores": [
    0.9992,
    0.9937,
    0.9995,
    0.9997,
    0.9993,
    0.9996
  ],
  "symbols": [
    "ก",
    "ก",
    "7",
    "7",
    "4",
    "2"
  ]
}

Table 28. Description of the fields in the «thailand_plate_recognition_v1» classifier

Field name	Type	Description	Possible values
classifier	string	Name of classifier	thailand_plate_recognition_v1
execution_time	int	Execution time in milliseconds	0…1000
score	float	Assessment of the accuracy of determining the region	0.0000…1.0000
type	string	Feature type	region
value	string	Recognition result	Bangkok, Chai Nat, Sing Buri, Ratchaburi…etc.
score	float	Assessment of the recognition accuracy of each feature	0.0000…1.0000
type	string	Type of feature	- spec_type – LP form
value	string	The value of LP feature	- regular
			- bus&truck – trucks and buses;
			- private – private individuals;
			- business – commercial vehicles;
			- taxi
			- temporary – dealer vehicles;
			- non-regular – individual;
			- moto – motorcycles
regno_ai_score	float	Assessment of the recognition accuracy of the LP	0.0000…1.0000
symbol_scores	float	Accuracy score for recognizing each LP symbol	0.0000…1.0000
symbols	string	Array of recognized LP symbols	0…9; ก, ข, ค, ฆ…

Classifier accuracy

The accuracy of the classifiers was measured for several datasets. All data were obtained from various sources obtained under different conditions.

The accuracy values for recognition of vehicle and LP attributes by classifiers are given in Table 29.

Table 29. Data on the accuracy of classifiers

Classifier	Number of images	Accuracy (%)
Car_brand_model_v2 (brand)	213 000	98.65
Car_brand_model_v2 (model)	213 000	96.77
Vehicle_color	13 323	96.39
Vehicle_type	10 413	99.34
Vehicle_emergency_type	18 016	82.00
Public_transport_type	12 683	98.40
Special_transport_type	24 844	97.95
LP attributes classifiers	10 000	99.70