INTELLIGENT SYSTEM AND METHOD FOR THE RECOGNITION OF POTHOLES IN ROADS BASED ON AN ARTIFICIAL NEURAL NETWORK MODEL.

Abstract

The present invention relates to an intelligent system for pothole recognition on roads based on an artificial neural network model that utilizes a structure of inverted residual connections within a set of bottleneck connections. The ANN has been pre-trained using a function optimizer that accelerates convergence and reduces oscillations that may occur during the ANN training process.

Description

INTELLIGENT SYSTEM AND METHOD FOR RECOGNIZING POTHOLES ON ROADS BASED ON AN ARTIFICIAL NEURAL NETWORK MODEL FIELD OF INVENTION The present invention generally relates to an intelligent system and method for recognizing potholes on the surface of a road using a previously trained Artificial Neural Network (ANN) model, which has the ability to determine the existence of a pothole from an image, as well as its classification. BACKGROUND OF THE INVENTION Currently, roads, avenues, and streets constructed with asphalt or hydraulic pavement require precise and constant inspection to prevent accidents and damage to motor vehicles traveling on them. This is due to the formation of irregularities, commonly known as potholes, on the pavement surface. These potholes are caused by natural factors such as rain, earthquakes, or snow, but primarily by the constant high volume of traffic. Therefore, effective and efficient inspection and evaluation of the pavement's condition is essential for maintenance management, enabling the development of plans, methods, and budgets for pavement maintenance. There are manual methods for determining and classifying potholes on a roadway. These methods involve a visual inspection of the pavement surface by maintenance personnel or inspectors, either on foot or while driving a slow-moving vehicle. However, this visual inspection of the road surface is subject to the subjectivity of the personnel. Therefore, it requires a significant level of human intervention and time, considering the length of the road network. Furthermore, inspectors must be physically present on the roadway, exposing themselves to various risks. On the other hand, there are semi-automated methods for evaluating pavement condition where images of the roadways are automatically obtained by a high-speed vehicle. However, the process of recognizing and classifying potholes is subsequently carried out by inspectors in their offices. This approach improves inspector safety, but it relies on manual identification by a human expert, which is also a time-consuming task. For this, maintenance personnel first need to identify the existence of a pothole on a section of the road or highway using one or more images of a point of interest to assess whether a pothole exists, and then classify the detected pothole.Pothole classification allows for determining the degree of damage to the pavement surface, and based on this information, maintenance personnel can establish a maintenance program to repair the damage. However, as previously mentioned, manual pothole detection and classification has certain disadvantages, such as a heavy workload, low efficiency, high subjectivity, and lack of repeatability. Automated methods and systems for pavement condition assessment employ sophisticated autonomous vehicles that require a multitude of sensors (lasers, vibration sensors, accelerometers, cameras, sonar, etc.) to collect a vast amount of data. This data is not only related to detecting and classifying potholes on the pavement surface but also includes information on vehicle driving and all variables recorded during its autonomous operation. Consequently, the cost of the vehicle and the infrastructure required to process the sensor data is prohibitively high, even without considering future vehicle maintenance costs. Document IN 202041024592 A discloses an automated pothole detection system based on thermal imaging for the prevention of any accident, more particularly an artificial intelligence system that alerts the driver of a vehicle about potholes in real time, comprising 1.a an embedded computational unit with a wireless communication interface provided with a machine learning algorithm to process the input acquired with sensory components and provide an alert notification, a thermal image sensor, a digital image camera unit, a display unit, a memory unit, an alert notification unit that acts to inform of the pothole in the road and / or pavement before it is approached, an electrical power unit and a web and mobile Internet server and telephone infrastructure. While document US 10480939 B2 describes a mobile pavement surface scanning system for detecting pavement damage. The system comprises one or more light sources mounted on the mobile vehicle to illuminate a pavement, one or more stereoscopic imaging devices mounted on the vehicle to capture sequential images of an illuminated pavement surface, and a plurality of positioning sensors mounted on the mobile vehicle, the positioning sensors adapted to encode the movement of the mobile vehicle and provide a timing signal for the sequential images captured by one or more stereoscopic imaging devices.One or more computer processors are adapted to synchronize the intensity image pairs captured by each camera in the one or more stereoscopic image capture devices, perform a 3D reconstruction of the pavement from the intensity image pairs using stereoscopic principles, generate a depth image and an intensity image pair from the 3D reconstruction images, and process at least one of the depth image and the intensity image using one or more hazard detection modules to detect a type of damage to the pavement. However, in other automated processes for pothole identification, the processing of information collected by the autonomous vehicle's sensors is carried out using image processing and intelligent pattern recognition methods to identify the condition of the road surface. However, to achieve accuracy in recognizing pavement damage, it is necessary to post-process the outputs of these recognition methods, since images taken by moving vehicles are subject to occlusion, drastic changes in lighting, and / or alterations inherent to movement. For example, patent CN 111310558 A discloses an intelligent method for extracting pavement defects based on deep learning and image processing. The method comprises a deep learning (DL) component and an image processing technology component (1PT). The image processing method includes the steps of preprocessing an image of pavement cracks; performing image enhancement based on an avoidance algorithm; segmenting the image; performing image post-processing; and determining and evaluating the pothole type. The pothole length calculated by the method is the pixel length.Therefore, if you need to obtain the actual length of the pothole in the image, you must know the actual length of the pothole in one image and the calculated pixel length, and the lengths of the pothole in other images are calculated according to the ratio of the actual length of the pothole to the calculated pixel length, so that the calculated length is compared with the actual length and the accuracy of the calculation is determined. Therefore, there is a need for an intelligent system and method for recognizing potholes on roads based on an artificial intelligence model that uses a previously trained Artificial Neural Network (ANN), which does not require an autonomous vehicle or other type of vehicle with special sensor and camera adaptations to capture an image of a point on the pavement surface of the road of interest. The present invention possesses several features that make it faster and more accurate in determining and classifying potholes. As will be described later, the intelligent system and method for pothole recognition on roads is based on an Artificial Neural Network (ANN) which has been previously trained using a training method based on a function optimizer that accelerates convergence and reduces oscillations that may occur during the ANN's training. Therefore, those with experience in the technique know the need for a low-cost pothole detection system and method that contributes to the prompt maintenance of roads and presents the following advantages. OBJECTIVES OF THE INVENTION The main objective of the present invention is to provide an intelligent system and method for the recognition of potholes on the surface of a road based on a previously trained Artificial Neural Network (ANN) model. A second objective of the present invention is to provide a tool for capturing images of possible potholes on a mobile device that allows any user to generate a report based on an image captured of a possible pothole at a specific point on the surface of a road geolocated by the device. A third objective of the present invention is to provide an automated method for detecting and classifying a pothole on the surface of a roadway. A fourth objective of the present invention is to provide a method for training an ANN that is based on a function optimizer that accelerates convergence and reduces oscillations that may occur during the training of an ANN. A fifth objective of the present invention is to generate and send a message to the user, once the method and system of the invention has detected and classified a bump in the received image 5. A sixth objective of the present invention is to provide the user with a geolocation link where the coordinates of the pothole located by the system and method of the invention are established. A seventh objective of the present invention is to provide a database that stores the images processed by the system, so as to make it possible to retrain the ANN when the system requires it. BRIEF DESCRIPTION OF THE INVENTION The present invention relates to an intelligent system for pothole recognition on roads based on an artificial neural network model. This system comprises a mobile device with an image capture tool that identifies up to 20 potential potholes. The tool then generates a report indicating the possible presence of a pothole at a specific point on the road surface, recording an image of that point. The image capture tool also records the geolocation data of the point of interest where the image was taken.Subsequently, the tool sends the report with the image and geolocation data to a pothole recognition and reporting server that includes an artificial intelligence module (1A) which receives the image and geolocation data input from the report, where a previously trained Artificial Neural Network (ANN) has the ability to distinguish the existence of a pothole in the provided image. If the neural network (NN) detects a pothole in the provided image, it classifies the pothole into one of the types it has learned to distinguish: 1) no pothole; 2) crocodile or cracked pothole; 3) longitudinal pothole; or 4) pit-like pothole. When the NN identifies a non-pothole, it immediately ignores false reports of a possible pothole, thus expediting the processing of other reports. The pothole reporting server also has an alert module that generates an alert message to a pre-configurable address so that the damage located at the point on the road surface where the system has detected a pothole can be addressed. The alert message includes image data of the detected pothole, its geolocation coordinates represented on a map, and its classification. Furthermore, the system of the invention comprises a database that stores the images received from the report server, so that it is possible to retrain the ANN when the system requires it. Furthermore, the invention relates to a method for pothole recognition on roads based on an artificial neural network model comprising: generating a report of a possible pothole at a specific point on the road surface using a pothole image capture tool on a mobile device; recording, through the pothole image capture tool, an image of the road point of interest, where the tool also records the location data of the road point of interest; generating a pothole report from the image data; transmitting, via the mobile device through a wireless network, the pothole report to a pothole recognition and reporting server;Identify, using an artificial intelligence module comprising a previously trained artificial neural network (ANN), whether there is a pothole in the image, and classify it accordingly. If the AI ​​module identifies the existence of the pothole and its classification, it sends a pothole identification alert via message or email through a report generator module so that it can be addressed or identified on a geolocation map; and stores the image of the identified pothole in a database of the pothole recognition and reporting server for future retraining in case the accuracy of the ANN decreases in distinguishing and classifying a pothole. It is understood that this description is not limited to the specific configurations, processes, steps, and materials disclosed herein, as such configurations and materials may vary in certain details. It should also be understood that the terminology used herein is for the purpose of describing particular modalities and is not intended to be exhaustive. DESCRIPTION OF THE FIGURES AND DRAWINGS This and other objects, features, and advantages of the invention will become evident from the following detailed description of the invention considered in conjunction with the illustrated figures, in which: Figure 1 illustrates a schematic diagram of the intelligent pothole recognition system of the present invention. Figure 2 illustrates a flowchart of the intelligent method for pothole recognition at a point on a road surface. Figure 3 illustrates a block diagram of the ANN training model according to the invention. Figure 4 illustrates a block diagram of the ANN training method of the present invention. Figures 5a, 5b, 5c, and 5d illustrate our examples of each of the four classes of training pads for the ANN of the invention. DETAILED DESCRIPTION OF THE INVENTION First, it should be clearly understood that similar reference numbers are used to consistently identify the same elements in the various figures, since these elements can be explained or described in greater detail in the complete written specification, of which this detailed description is an integral part. Certain terms are used throughout the following description and claims to refer to particular features or components. As someone skilled in the art will appreciate, different people may refer to the same feature or component by different names. This document is intended to distinguish between components or features that differ in name but not in function. The modalities described herein comprise a combination of advantages and features intended to address various shortcomings associated with certain previous devices, systems, and methods. The foregoing has broadly outlined the technical characteristics and advantages of the disclosed modalities to facilitate a better understanding of the detailed description that follows. The various features and advantages described above, as well as others, will become evident to those skilled in the art after reading the detailed description and consulting the accompanying figures. It should be appreciated that the specific design and modalities disclosed can easily be used as a basis for modifying or designing other devices to accomplish the same purposes as the disclosed modalities. The present invention relates to an intelligent system and method for recognizing potholes on a road surface using a pre-trained Artificial Neural Network (ANN). Specifically, the ANN of the present invention has, as its first layers, bottleneck and inverted residual connections that alleviate vanishing and gradient explosion problems in deep networks. These layers are followed by additional layers of averaging pooling, a spatial filter mask layer (2,2), a dense layer of 128 neurons, a random neuron shutdown layer with a probability of 0.5, and finally, two dense layers of 32 and 4 neurons, respectively. This adaptive approach reduces poor training time performance, thus achieving a suitable balance between resources and accuracy in deep learning image analysis. Therefore, using the ANN architecture described earlier in the present invention, an optimization method for training an ANN is also proposed. The optimizer proposed by the invention uses a one-dimensional scalar filter that allows estimating the current state of the system from prior information to process each parameter of the objective function. The scalar filter is combined with an optimization methodology that uses the information from the first and second statistical moments of the gradient to calculate the maximum or minimum of the objective function. This allows training the neural network by reducing execution time and computational complexity, while achieving better accuracy. In addition, the proposed optimizer is suitable for handling problems involving large-scale databases, architectures with a large number of parameters, non-stationary objectives, and noisy and / or sparse gradients. According to Figure 1, the intelligent system for pothole recognition (10) at a point on the surface of a road comprises an image capture tool for possible potholes (11) provided in a user mobile device 20 (12) which displays through a screen of said device, a user interface (13) through which an image is recorded at a point of the road of interest by means of a camera of the user mobile device (12) and also records the location coordinates where the image of interest was captured and recorded 25, thus generating a report of a possible pothole (15). Subsequently, the pothole detection tool (11) on the user's mobile device (12) sends a pothole detection report (15) to a pothole recognition and reporting server (14). This report contains the image and location data of the point of interest on the roadway, recorded by the tool via a wireless signal (Z) connected to the user's mobile device (12). This signal can be a Wit'i, 3G, 4G, etc. The image and geolocation data are sent using a Python script that uses a POST request over the HTTP protocol. In addition, the pothole recognition and reporting server (14) comprises a report reception module (19) that receives the data input of the image and geolocation of the report generated (15) by the tool for capturing images of possible potholes (11); and an artificial intelligence (AI) module (17) that consists of a previously trained Artificial Neural Network (ANN) (16) which distinguishes whether or not there is a pothole in the received image based on its training. Thus, the ANN (16) receives and analyzes the image data of the report of a possible pothole (15) and identifies if there is a pothole and assigns it within one of the other three classes that the ANN has already learned to distinguish: 1) no pothole, 2) crocodile or cracked pothole, 3) longitudinal pothole or 4) pit pothole. Continuing with Figure 1, once the ANN (16) classifies the pothole detected in the image data of a possible pothole report (15), an alert module generates an alert message (18) containing data related to the processing results of the image data from the report (15) by the ANN (16). This alert message (18) includes the classification data of the detected pothole as well as its location data, so that it can be addressed by the responsible maintenance and repair personnel or this information can be provided to a geolocation map to alert drivers. If the AI ​​module (17) does not detect a pothole in the image, the server (14) discards the possible pothole report (15) and terminates the system execution without issuing a warning. Together, the image of the detected pothole is stored in a database (20) to perform future retraining of the ANN (16) in case its accuracy in distinguishing and classifying potholes decreases. The term server refers to a functionally related group of electrical components, such as a computer system, which may or may not be connected to a network and may include as many hardware and software components as are used to perform certain functions. The server may also be integrated with a database management system and one or more associated databases. In accordance with the professional practice of computer programming specialists, the invention is described below with reference to the operations performed by a computer system or similar electronic system. It will be noted that the operations symbolically represented include the manipulation, by means of a processor (e.g., a central processing unit), of electrical signals representing data bits and the storage of data bits in memory locations, such as system memory, as well as other signal processing. The memory locations where the data bits are stored are physical locations that have particular electrical, magnetic, optical, or organic properties corresponding to the data bits. Figure 2 shows a flowchart of the intelligent method for pothole recognition at a point on a road surface, which consists of: registering (21) using a tool for capturing images of possible potholes on a mobile device, an image of a point of interest on the road as well as its geolocation data; generating (22) using the tool for capturing images of possible potholes, a report of a possible pothole that includes the image data and location of the image of interest; and transmitting (23) using the mobile device in a wireless connection through a WiFi, 3G, 4G, etc. network, the report to a pothole recognition and reporting server for processing. Continuing with the method in Figure 2, the data contained in the report is received through an AI module of the pothole recognition and reporting server; an II 5 module of the pothole recognition and reporting server crops, scales, and analyzes (24) the received image, using the ANN that identifies whether a pothole appears in the image of the point of interest. If the ANN identifies (25) a pothole, it proceeds to classify (26) the identified pothole according to its 10 structural characteristics into one of the following types: 1) alligator or cracked pothole, 2) longitudinal pothole, or 3) pit pothole. Thus, an alert message is generated (27) by a server alert module. The message includes data (15) from the pothole classification result detected by the ANN, the image in question, and a link to a geolocation map with the coordinates of the detected pothole. Subsequently, the alert module transmits (28) the alert message to an email address or a previously established telephone number (20) so that the report of a possible pothole generated by the pothole image capture tool can be addressed by the personnel in charge or be useful for alerting drivers. In addition, the 25 pothole recognition and reporting server stores (29) the image of the detected pothole in a database to perform future re-training of the ANN if necessary. On the other hand, if the ANN does not distinguish a bump in the input image; it discards (30) the report of a possible bump, and ends (31) the execution of the system without generating and issuing an alert message. With reference to Figure 3, it shows a block diagram of the training model (40) of an ANN, in which a training module (41) trains the ANN to recognize objects in images using a first public image database (42) organized according to its hierarchy, in which each node of that hierarchy is represented by hundreds and thousands of images. The architecture of the ANN (16) consists of: A 2D convolution layer with a 2 offset, a 3x3 filter mask, and an output of 32 maps. Followed by a series of 17 bottleneck layers with a 3x3 filter mask and the following structure: Bottleneck layer with offset of 1 and expansion factor of 1, output of 16 maps; Bottleneck layers with offset of 2 and expansion factor of 6, output of 24 maps; Bottleneck layers with offset of 2 and expansion factor of 6, output of 32 maps; Bottleneck layers with offset of 2 and expansion factor of 6, output of 64 maps; Bottleneck layers with offset of 1 and expansion factor of 6, output of 96 maps; Bottleneck layers with offset of 2 and expansion factor of 6, output of 160 maps; Bottleneck layer with offset of 1 and expansion factor of 6, output of 320 maps This is followed by a 1x1 two-dimensional convolution layer with an offset of 1 and an output of 1280 A 7x7 average grouping layer A 1x1 two-dimensional convolutional layer Continuing with Figure 3, the training model (40) of the present invention uses a transfer learning module (43) which consists of taking the ANN to perform a task similar to the target task and cutting the architecture, i.e., keeping only the initial layers of the ANN. This results in a modified RNA (44) which comprises a total of 2,388,707 trainable parameters, with an architecture consisting of the following layers: Layer type Hyperparameters Number of parameters Average grouping Filter mask=(2,2) 0 Dense Neurons=128 Activation = Linear Rectifying Unit 163968 Random shutdown of neurons with fixed probability Probability=0.5 0 Dense Neurons--32 Activation =RELU 412 8 Dense NeuronsM Activation-Soft tmax 99 Continuing with Figure 3, an optimal training module (45) trains these trainable parameters of the modified ANN (44), using as input data images from a second public image database (46) with three classes for each pothole type, as exemplified in: Figure 5a - crocodile pothole; Figure 5b - longitudinal pothole; Figure 5c - pit pothole; and Figure 5d - no pothole. Likewise, an optimal training module (45) uses a function optimizer (47) that is responsible for finding the optimal parameters of the modified RNA (44) . Thus, the training module (45) uses a scalar filter as a function optimizer to estimate the gradient of the cost function in the cost rule. This attenuates updates to the cost function parameters when approaching the optimum point. The filter adds significant and relevant variations to the gradient to find better solutions in the loss function. In the optimization process, the objective parameters of the functions are updated iteratively with proportional steps in the negative direction of the gradient. This attenuates updates to the cost function parameters (weights and biases) when approaching the optimum point, where the gradient of the cost function is used as the zk measurement of a linear system under the following mathematical expressions: -Filter Parameters ak=qk=hk=í -Scalar filter Xk\kl=Xk-1\kl Pk\k-1—pk-l\kl + l _ P / c|k-lkPk\k-1 +rk Xk\k=Xk\k-1 +kk (Zk — Xk\k- 1) Pk\k=(l— Kk^p^kl Where ak, qk, hk are the dynamic model of the system to be estimated, the covariance of the system noise, and the measurement of the system respectively; and where: the a priori estimated state • p / c / <-i : is the error covariance associated with the a priori estimate • kk: is the Kalman gain • Xk\k: is the a posteriori estimated state! • pk k: is the error covariance associated with the posterior estimation With reference to Figure 4, which shows a block diagram of the ANN training method of the present invention, which is based on a function optimizer that uses the gradient of the cost function as the zk measurement of a linear system using the following mathematical expressions: gt^K^g^ τπ{—βιηΐι-ι + (_1—βι)^ι Pt=^2Ví-l + (l- / ?2)^t2 where the following expressions are all scalars: • gt is the scalar gradient at time t • K(^t) is a function that summarizes the process of all scalar filters, receives the gradient at time t and returns the gradient estimated by the respective scalar filters • mt are exponential moving averages over the previous gradients • Vt are exponential moving averages over the previous squared gradients • mt and vt are the exponential moving averages of the previous gradients and squared gradients, but with bias correction • 0t+i are the system parameters, in this case, the weights of the ANN and where: • η is the learning factor • βι, β? are exponential decay factors for the estimated statistical moments • γ is the exponential decay factor for the noise associated with the measurements • f(0) is the objective function to be optimized with parameters • θο is the initial vector of parameters • K(°) is the function that summarizes the process of all the scalar filters, receives the gradient gt • and returns the gradient estimated gt by the scalar filters Thus, the function optimizer uses exponential moving measures on the past gradient (mt) and the squared past gradient (mt] . According to the flowchart in Figure 4, which shows a flowchart of the modified ANN training method comprising the following steps: initialize (51) the first and second moment vectors mu and uo, the estimated state vector xo|o, the optimizer gain vector Ko, as well as the covariance vector po ioy and time t; if Θ does not converge (52), increment (53) the counter t=t+l ; increment (53) the time counter t=t+l; calculate (54) the cost function gradients with respect to 0Í-1, using the expression gt~ where the gradient (first order method) is useful to know the direction in which the network parameters should move, so that the synaptic weights and thresholds are constantly adjusted to perform an adequate classification; calculate (55) a priori states xk\ki=Xk-i\ki, where the a priori estimation of the states, in this case the gradients, 2b are calculated based on past information of their behaviors, that is, trying to estimate the next position of the states without using any extra information; Calculate (56) covariances estimated a priori under the expression p(t|ti)=p(ti | ti)+l, where the a priori estimate of the covariance refers to the transformations that the states undergo, these transformations occur due to their associated dynamics (behaviors) and external noise that may be present, in this case, it is assumed that the states have an autonomous behavior (evolve without depending on time) and present a constant noise; calculate (57) the measurement noise under the expression rt= , where the measurement noise induces a variable perturbation (using exponential decay) to the measurement of the states, in this way, the noise can help to reach new ideal values ​​for the gradients, which translates to finding new synoptic weights and desirable thresholds to improve the performance of the neural network; Calculate (58) the gains of the estimator filter according to the expression kt= _Lílíd—,where the calculated gains Pt ir-1+n are factors that consider the relationship between the covariances estimated a priori and the covariances estimated a posteriori of a previous time step (with their respective associated measurement noise), in other words, the gains indicate the relationship between the past transformations to the states and the transformations estimated for the states in the current time step; calculate (59) the posterior estimated states Xt|t=Xt|t—i +kt (zk — Xt|ti) , where the posterior estimates of the states can be considered as a correction to the prior estimates, to make the correction the innovation of the observation (difference between the prior estimated state and the observation to the real states) and the calculated gain of the estimator filter are used, in this way, the gain quantifies the innovation that is between the prior estimates and the real observations, which would make the posterior estimates of the states approximate the values ​​of the real states; calculate (60) the posterior estimated covariances pt|t= (1— Kt)pt|ti, where the estimated covariances also require a correction that considers the information of the current time step, then, the calculated posterior estimate of the covariances is obtained from the complement of the calculated gains for the estimator filter and the prior estimate of the covariances; obtain (61) the estimated gradients with the function K under the expression gt=K(gt)t where the function K summarizes steps 55 to 60, which are equivalent to applying the estimator filter process, which would result in the estimated gradients that were calculated in step 59; calculate (62) the estimate of the first moment 771(= / / 1771(-1 + (1- / / 1) gt.where the first moment estimate is an approximation of the average of the accumulated gradients up to the current time step; calculate (63) the estimate of the second moment V(= / / 2W(-i + (l— / / 2) gt2en where the estimate of the second moment is an approximation of the (non-centered) variance of the accumulated gradients up to the current time step; calculate (64) the first moment correction mt= that the first moment estimate is initialized at zero, it is necessary to make a correction to counteract this biased given; Calculate (65) the correction of the second moment vt= since the estimate of the second moment is initialized to zero, a correction is needed to counteract this bias; Update (66) the objective function parameters using Qt+y= Θ, — n , from the corrected estimates / Ot + e of the first and second moment the parameters of the neural network can be adjusted, where, steps are given that consider the average direction and these are extended according to the estimate of the standard deviation, that is, the smaller the estimated standard deviation, the larger the steps and, the larger the estimated standard deviation, the smaller the steps, if θ converges (52), the process is finalized (67). If 0 converges (52), the process is terminated (67). Thus, the method in Figure 4 uses exponential moving averages on the past gradient (mt) and the past squared gradient (Vt), with βA and β2 as coefficients for decreasing weighting, respectively. Therefore, a correction of these terms occurs because they are biased towards zero; consequently, the estimated corrected biases of the first moment (the mean) mt and the second moment (decentered deviation) vt are computed. Furthermore, the transfer learning scheme used allows for the reuse of models trained for other problems and their specialization for a specific problem. Transfer learning discards the last layers of a network, concatenates new layers, and retrains to a certain depth. Additionally, according to the descriptions in this document, the training method and the pothole recognition method, as used in this document and with respect to Figures 2-4, includes the hardware, firmware, or software running on a machine and / or combinations of each of these elements to perform a function or action, and / or to execute a function or action of another logical system, method, and / or system. The foregoing description is provided merely to illustrate the invention and is not intended to be limiting. Because those skilled in the art can infer modifications from the disclosed embodiments that incorporate the spirit and substance of the invention, it should be construed as including everything within the scope of the appended claims and their equivalents.

Claims

1. An intelligent system for pothole recognition on roads based on an artificial neural network model comprising: an image capture tool for possible potholes provided on a user mobile device that displays, through a screen of said device, a user interface through which a report of a possible pothole is generated, registering an image at a point on the road of interest, in addition to the location coordinates where the image of interest was captured;A pothole recognition and reporting server receives report data, and this server comprises: an artificial intelligence (AI) module comprising an artificial neural network (ANN) trained from a function optimizer that assumes each layer in the ANN has an associated true state vector of a linear dynamic system, where the estimated state vector has the same dimension as the sum of the stack dimensions, weights, and biases of its layer; where the ANN distinguishes whether there is a pothole in the provided image and assigns it to one of the classifications that the ANN learned to distinguish during its training; an alert module generates an alert message containing data related to the result of the image data processing from the ANN report; if the ANN does not distinguish a pothole in the report image, the server discards the report and terminates the system execution without issuing any alert.

2. The intelligent pothole recognition system according to claim 1, wherein the ANN has learned to distinguish three types of potholes consisting of: 1) crocodile or cracked pothole, 2) longitudinal pothole, or 3) pit pothole.

3. The intelligent pothole recognition system according to claim 1, wherein the alert message comprises the classification data of the detected pothole, as well as the location data.

4. The intelligent pothole recognition system according to claim 1, wherein the server comprises a database where the report image is stored when the module of 1A detects the presence of a pothole in it, in order to perform future re-training of the ANN in case its accuracy in distinguishing and classifying potholes decreases.

5. An intelligent method for pothole recognition on roads based on an artificial neural network model, characterized by comprising the steps of: generating a report of a possible pothole at a specific point on the surface of a road through a pothole image capture tool on a mobile device; registering an image of the road point of interest through the pothole image capture tool, where the pothole image capture tool also records the location data of the road point of interest; generating a report of a possible pothole from the image data; transmitting the report of a possible pothole to a pothole recognition and reporting server via a wireless network from the mobile device;Identify, using an artificial intelligence module comprising a pre-trained artificial intelligence network (ANN), whether a pothole exists in the report image; where the ANN is trained from a function optimizer that assumes that each layer in the ANN has an associated true state vector of a linear dynamic system, where the estimated state vector has the same dimension as the sum of the stack dimensions, weights, and biases of its layer; classify the detected pothole using the ANN into one of the classifications learned during its training; send a pothole identification alert through a report generator module of the pothole recognition and reporting server, so that it can be addressed or identified on a geolocation map;Store the image of the identified pothole in a database on the pothole recognition and reporting server for future retraining if the accuracy of the ANN decreases in distinguishing and classifying a pothole; if the ANN does not distinguish a pothole in the report image, discard the report by the pothole recognition and reporting server, where the server ends the execution of the system without issuing any alert.

6. The intelligent method for pothole recognition according to claim 5, wherein the pothole classification learned by the ANN is based on distinguishing three classes of potholes consisting of: 1) crocodile or cracked pothole, 2) longitudinal pothole, or 3) pit pothole.

7. The intelligent method for pothole recognition according to claim 5, wherein the alert message comprises the classification data of the detected pothole, as well as the location data.

8. The intelligent method for pothole recognition according to claim 5, further comprising storing in a database of the pothole recognition and reporting server, the image of the report when the AI ​​module detects the presence of a pothole in it.

9. An artificial neural network (ANN) trained to distinguish and classify a pothole in an image, characterized in that it comprises a function optimizer that assumes that each layer in the ANN has an associated true state vector of a linear dynamical system, where the estimated state vector has the same dimension as the sum of the stack dimensions, weights, and biases of its layer; wherein the ANN distinguishes whether there is a pothole in the provided image, and assigns it within one of its classifications that the ANN learned to distinguish in its training, wherein said classification consists of: 1) crocodile or cracked pothole, 2) longitudinal pothole, or 3) pit pothole.

Images