METHOD FOR DETECTING DAMAGED AREAS ON AN IMAGE OF A ROOM USING A CONVOLUTIONAL NEURAL NETWORK

Abstract

The invention relates to a method for detecting at least one damaged area on an image of a part, comprising the following steps: a. acquiring an input image (1) of said part concerned; b. cutting at least a part of said input image (1) into thumbnails by cropping, without normalizing the size or the orientation; c. using a database of part images in which thumbnails are annotated to provide examples of transitions between healthy areas and damaged areas; d. processing the thumbnails by means of a convolutional neural network based on the FR-CCN ​​Resnet 150 architecture to detect anomalies in the texture of said input image, by exploiting local characteristics of the thumbnails, said neural network having been trained by means of said database; and e.identification, by said convolution neural network, of transitions (2) between damaged (3) and healthy (4) textures on said input image, the transitions (2) being determined and extracted automatically from the thumbnails by the learning process and the convolution layers of said convolution neural network. Figure for abstract: Fig 1.

Description

Title of the invention: METHOD FOR DETECTING DAMAGED AREAS ON AN IMAGE OF A ROOM BY MEANS OF A CONVOLUTIONAL NEURAL NETWORK Field of invention

[0001] The invention relates to a method for detecting at least one damaged area in an image of a room, for example a room in an apartment, a house or an industrial premises. To do this, the invention relates more specifically to a method for detecting damaged areas using a convolutional neural network, i.e. a specific artificial intelligence algorithm for identifying an image or a stream of images.

[0002] The aim of the invention is to limit the error rate in the detection of disaster areas.

[0003] The invention finds a particularly advantageous application for automatically detecting the surface area of ​​a claim and facilitating the assessment of claims for an expert. State of the art

[0004] Damage detection, especially in building rooms, is a major challenge for property owners, insurance companies, and repair professionals. Historically, this detection has been primarily performed through visual inspection by experts, who assess the damage and estimate repair costs. However, this method is often time-consuming, expensive, and may miss less visible damage.

[0005] With the advent of information technology, several automated solutions have been proposed. Among these, image- or video-based systems have been suggested to automatically detect damage in parts. These systems often rely on traditional image processing algorithms to identify damaged areas.

[0006] Recently, with the growing popularity of artificial intelligence (AI) and neural networks, solutions based on these technologies have been developed. These solutions rely on machine learning to identify and classify damage based on large image datasets.

[0007] However, despite these advances, many challenges remain. The variability in shapes, colors, and textures of disaster areas makes their accurate detection difficult. In addition, most existing solutions fail to detect subtle transitions between disaster and non-disaster areas, which can lead to inaccurate detections.

[0008] For example, document CN113177519A describes a method for assessing dirt and clutter in a kitchen based on density estimation. Although this method uses a neural network to assess the condition of the kitchen, it is specifically designed to detect dirt and clutter rather than damage such as water damage or cracks.

[0009] Thus, there is a real need to improve the accuracy and robustness of disaster detection systems, particularly in the context of building rooms, while further automating the process. Statement of the invention

[0010] The invention aims to overcome the limitations of current techniques for detecting damage in the rooms of a building by using an approach combining a specific neural network architecture and a strategy for dividing into images.

[0011] To this end, the invention relates to a method for detecting at least one damaged area on an image of a part, characterized in that it comprises the following steps: a. acquisition of an input image of said part concerned; b. cutting at least a portion of said input image into thumbnails by cropping, without standardizing the size or orientation; c. use of a database of part images in which thumbnail images are annotated to provide examples of transitions between healthy and damaged areas;

[0012] d. processing of the thumbnails by means of a convolutional neural network based on the FR-CCN ​​Resnet 150 architecture to detect anomalies in the texture of said input image, by exploiting local characteristics of the thumbnails, said neural network having been trained by means of said database; and e. identification, by said convolutional neural network, of transitions between damaged and healthy textures on said input image, the thresholds and transitions being determined and extracted automatically from the thumbnails by the learning process and the convolution layers of said convolutional neural network.

[0013] Thanks to these provisions, disaster areas can be detected reliably, without requiring a significant amount of calculations.

[0014] The thumbnails can have a definition of between 1 / 10th and 1 / 1000th of the definition of the input image, which is a simple and effective embodiment of the invention.

[0015] After the step of cutting the thumbnails, all the thumbnails can have the same definition, and form rows and columns of thumbnails in the input image, the overlap rate between two consecutive thumbnails on a row and / or a column being between 5 and 25%, which is a particularly reliable method and gives reproducible results.

[0016] Said identification of transitions may include a classification of the type of disaster among a predefined list, which makes it possible to process this disaster more efficiently.

[0017] Said method may also include a step of agglomerating several identified transitions to identify a disaster area, which makes it possible to have information on the size of the disaster in the input image.

[0018] The input image may include an object of known dimension, and the area of ​​the disaster may be estimated by comparing the dimensions of the disaster area and said object on the input image, which makes it possible to have information on the actual size of the disaster.

[0019] Said identification of transitions includes a classification of the importance or severity of the disasters on a predefined scale, which makes it possible to deal with this disaster more efficiently.

[0020] Said image database comprises more than 50,000 annotated part images, which provides good reliability to the invention. Brief description of the drawings

[0021] The present invention and its advantages will appear better in the following description of several embodiments given as non-limiting examples, with reference to the appended drawings, in which:

[0022] [Fig-1] [Fig.l] is a schematic view of an image on which has been detected a disaster area according to a detection method according to a preferred embodiment of the invention. Description of the embodiments

[0023] For the purposes of the invention, a “thumbnail” corresponds to a portion of the input image 1 of smaller dimension. For example, the thumbnails have a definition, that is to say a total number of pixels, of between 1 / 10th and 1 / 1000th of the definition of the input image 1. This ratio depends in particular on the definition of the input image 1.

[0024] Following extensive research and rigorous observations, it has been discovered that the combined use of these two elements - the specific neural network architecture and the image slicing - allows for significantly improved detection rates, surpassing the performance of traditional neural network-based systems. More specifically, these unexpected improvements in disaster detection are attributed to the ability of this combined approach to accurately capture and identify subtle transitions 2 between disaster-affected and non-disaster-affected areas.

[0025] The specific architecture of the neural network, based on the FR-CCN ​​Resnet 150 architecture, has proven particularly effective in processing thumbnails without normalization of size or orientation. This optimization allows the network to learn and identify relevant local features on the thumbnails, thus strengthening anomaly detection.

[0026] The thumbnail slicing strategy allows for the extraction of smaller, more manageable areas from the input image 1. These thumbnails are then processed individually by the neural network, allowing for a more detailed and targeted evaluation of each part of the image.

[0027] By combining these two essential elements, the invention capitalizes on the respective strengths of the specific neural network architecture and the thumbnail slicing strategy to provide a first-rate loss detection system capable of accurately and reliably detecting losses where other systems may fail.

[0028] Accordingly, this invention provides a novel and effective solution to the technical challenges associated with automatic loss detection, thereby addressing the unmet needs of property owners, insurance companies, and repair professionals.

[0029] In a preferred embodiment, before cutting at least a portion of the input image 1 into thumbnails, and if the definition of the input image 1 does not correspond to a predefined fixed definition, the definition of the input image 1 is modified in order to achieve the predefined fixed definition. The images used to train the algorithm then also undergo this processing. The predefined fixed definition may for example be 600x600 pixels, this example not being limiting, these values ​​being able to vary according to the applications and the type of input images 1 used.

[0030] Then, preprocessing can be carried out to detect one or more regions of interest in the input image 1. These regions of interest are classified, and their boundaries (coordinates of the “bounding boxes”) are determined by regression, according to a known method, for example of the RPN (“Region Proposal Network”) type.

[0031] During the step of dividing at least a portion of the input image 1 into thumbnails, the entire input image 1 may be divided, or only one or more regions of interest. Each thumbnail may be assigned the following attributes: an identifier, and position attributes allowing it to be located in the input image 1. The position attributes comprise, for example, the coordinates of four points corresponding to the four corners of the thumbnail, or the coordinates of two points corresponding to two opposite corners of the thumbnail, or the coordinates of a single point corresponding to a corner of the thumbnail, if the dimensions of the thumbnail are known.

[0032] The thumbnails are preferably of fixed size, that is to say that all the thumbnails have the same definition, for example 30x30 pixels, this example not being limiting, these values ​​being able to vary according to the applications and the type of input images 1 used.

[0033] The ratio between the definitions of the thumbnails and the definition of the input image 1 can be between 1 / 100 and 1 / 1000, and can be adjusted according to the type of input image 1 used.

[0034] During the step of dividing the input image 1 into thumbnails, the thumbnails can be defined in the form of a grid, and form rows and columns. The adjacent thumbnails comprise a common part of the input image 1, the overlap rate being for example between 5 and 25%. For example, if the thumbnails have a definition of 30x30 pixels, and the overlap rate is 10%, the three columns of pixels located to the left, respectively to the right, of a given thumbnail, are also part of the thumbnail located immediately to its right, respectively to its left, and the three lines of pixels located at the top, respectively at the bottom, of the given thumbnail, are also part of the thumbnail located immediately above, respectively below.

[0035] During the training phase of the neural network, the images are annotated, preferably by an expert who manually determines the transition zones 2 on the images in the training database. The objective is to annotate the transitions 2 between disaster zones and healthy zones, which represents several annotations for a single disaster.

[0036] The image database used for training the neural network may comprise several thousand images, preferably at least 50,000 annotated part images. This volume allows efficient training, and a low error rate in the processing by the neural network of new input images 1.

[0037] The annotation of the images in the database preferably includes information concerning the type of damage from a predefined list, for example: water damage, crack, mold. Thus, during its processing of the input image 1, the neural network is able to detect transitions 2 between damaged areas and healthy areas, and to additionally provide information on the type of damage.

[0038] After the step of identifying the transitions 2 between damaged textures 3 and healthy textures 4 on said input image 1, the method according to the invention preferably comprises a step of agglomeration of several identified transitions 2, in order to identify a damaged area. This agglomeration is done according to an algorithm in which the transitions that are sufficiently close are agglomerated, using for example a distance threshold in pixels. It can also be provided that the agglomeration is only authorized in the case of transitions 2 concerning damages of the same type.

[0039] Once a disaster has been identified, the detection process may include a step estimation of its surface area. In order to move from a pixel scale to a metric scale, it may be provided to include in the capture of the input image 1 the presence of a reference object. The detection method then comprises a step of searching on the input image 1 for the reference object, for example by known image processing techniques, and calculating a scale as a function of the number of pixels occupied by the reference object on the input image 1. The reference object may be, for example, a standard coin, the diameter of which is known. If this diameter is, for example, equal to 22 mm, and the size of the coin on the input image 1 is 20 pixels, it can be determined that the scale of the input image 1 is 1.1 mm per pixel.

[0040] The detection method may include a step of classifying the importance or severity of the losses, on a predefined scale. This classification step may be carried out within the neural network, or by a separate algorithm. This classification may be carried out by taking into account the dimensions of the loss, for example in number of pixels, or the actual surface area determined as described above. The type of loss may also be taken into account and related to its surface area, in order to determine the severity.

Claims

Claims

1. Method for detecting at least one damaged area on an image of a part, characterized in that W comprises the following steps: a. acquiring an input image (1) of said part concerned; b. cutting at least a part of said input image (1) into thumbnails by cropping, without normalizing the size or the orientation; c. using a database of part images in which thumbnails are annotated to provide examples of transitions between healthy areas and damaged areas; d. processing the thumbnails by means of a convolutional neural network based on the FR-CNN Resnet 150 architecture to detect anomalies in the texture of said input image, by exploiting local characteristics of the thumbnails, said neural network having been trained by means of said database; and e.identification, by said convolution neural network, of transitions (2) between damaged (3) and healthy (4) textures on said input image, the transitions (2) being determined and extracted automatically from the thumbnails by the learning process and the convolution layers of said convolution neural network.

2. Method for detecting damaged areas on an image of a room according to claim 1, in which the thumbnails have a definition of between 1 / 10th and 1 / 1000th of the definition of the input image.

3. Method for detecting damaged areas on an image of a room according to claim 1 or 2, in which after the step of cutting out the thumbnails, all the thumbnails have the same definition, and form rows and columns of thumbnails in the input image, the overlap rate between two consecutive thumbnails on a row and / or a column being between 5 and 25%.

4. Method for detecting damaged areas on an image of a room according to one of claims 1 to 3, in which said identification of the transitions (2) includes a classification of the type of damage from a predefined list.

5. Method for detecting damaged areas on an image of a room according to one of claims 1 to 4, in which said method also comprises a step of agglomerating several transitions (2) identified to identify a disaster area.

6. A method for detecting damaged areas on an image of a room according to claim 5, wherein the input image comprises an object of known dimension, and the area of ​​the damage is estimated by comparing the dimensions of the damaged area and said object on the input image.

7. A method of detecting damaged areas on an image of a room according to one of claims 1 to 6, wherein said identification of transitions includes a classification of the importance or severity of the damage on a predefined scale.

8. A method of detecting damaged areas on an image of a part according to one of claims 1 to 7, wherein said image database comprises more than 50,000 annotated part images.

Images