Method for analyzing the degree of liberation of iron ore using computer vision

Description

1 / 33 METHOD FOR ANALYZING THE DEGREE OF LIBERATION OF IRON ORE USING COMPUTER VISION Field of Invention

[001] The present invention is in the field of image processing and classification, more specifically, in the field of image processing and classification using deep learning for analysis of the degree of liberation of iron ore. Background of the Invention

[002] Many researchers have sought methods for quantifying the degree of mineral liberation, aiming to predict the behavior of particles during separation processes. Gaudin (1939) pioneered the development of a binary calculation method for the visual estimation of liberation by size range in an optical microscope, through the quantification of the liberation index.

[003] Gaudin's method for determining the degree of ore liberation is a procedure used in mining and mineral processing to assess the degree to which minerals of interest have been liberated from gangue particles (non-valuable minerals) after comminution processes, such as grinding. It is one of the simplest methods for measuring this quantity.

[004] Gaudin's method involves the microscopic analysis of particles of different granulometric fractions, classifying them as liberated particles (composed only of the mineral of interest or only of gangue) or as intergrown particles (with a mixture of both). The degree of liberation is calculated as the proportion of liberated particles in relation to the total, usually expressed as a percentage. The following mathematical formula is used for the liberation percentages. i index of free particle release \ Degree of ZZberation = —--—------------—·— -------—— * 100 \index of release of free + mixed particles / Petition 870250012538, dated 02 / 14 / 2025, page 10 / 61 2 / 33

[005] Mineral-ore particles completely separated from the gangue receive the maximum liberation index, equivalent to 10. For mixed particles, composed of mineral-ore and gangue, the indices vary comparatively, depending on each particle and the product analyzed.

[006] Although widely disseminated, Gaudin's method is limited due to the imprecision of the results, as it is a subjective process, dependent on the opinion and experience of the executing technician, there is variation in the size of the particles within the same fraction[1].

[007] Advances in the assessment of the degree of liberation and image processing technologies have enabled the development of more precise and rapid techniques for quantifying liberation. Two of the most widely used techniques by professionals in the mineral industry today are MLA and QEMSCAN. In addition to these, TIMA and μ^ are being applied for this purpose.

[008] MLA[2] is a characterization system that uses a SEM microscope equipped with an energy-dispersive X-ray spectrometer (EDS) for mineral identification and automated mineral liberation characterization. The method allows the collection of various data from polished rock surfaces, sediments, or particulate samples, including modal mineralogy, porosity, particle size and shape, mineral associations, and digital textural maps. Mineral liberation measurements are performed by identifying phases through the analysis of backscattered electron (BSE) images, which allows determining particle boundaries for acquisition and classification of characteristic X-ray spectra of each mineral species in comparison with a library of reference spectra.

[009] However, MLA has some disadvantages such as a high investment cost, difficulty in distinguishing minerals with very similar compositions and inability to distinguish polymorphs[3].

[010] Like MLA, QEMSCAN is a technology that allows the quantitative determination of the degree of liberation through the use of coupled SEM. Petition 870250012538, dated 02 / 14 / 2025, page 11 / 61 3 / 33 with EDS and combined with image analysis and data processing software. In the QEMSCAN system, X-ray spectra are mapped for mineral classification and particles are subsequently distinguished by BSE. In contrast, in the MLA system, particles are first defined by BSE and subsequently classified by an X-ray spectrum per particle[4].

[011] TIMA consists of a SEM system coupled with up to four EDS detectors. Thus, the high throughput of the technology is due to the use of multiple EDS detectors that act as a speed multiplier and reduce the impact of shadowing on the measurement, these being the main advantages of TIMA in relation to MLA and QEMSCAN[5].

[012] X-ray microtomography (X-ray CT) is an advanced technique that enables high-resolution three-dimensional (3D) analysis of objects, allowing the estimation of characteristics such as size, shape, particle distribution, as well as identifying porosities, internal defects, and structural heterogeneities. The method is based on exposing the sample to an incident X-ray beam while it is rotated, capturing a series of two-dimensional (2D) projections at different angles. These projections are subsequently reconstructed using computational algorithms, generating a detailed 3D model. CT is widely used in various fields, such as materials science, geology, biology, and engineering, due to its ability to provide non-destructive and detailed information about the internal microstructure of samples without the need for physical cutting.

[013] Although the techniques mentioned are fast and their results are more representative, accurate and reliable, they are high-cost investment technologies compared to optical microscopy analysis methods, limiting their use.

[014] In the state of the art, some prior techniques have already attempted to provide solutions to similar problems. Examples of the state of the art include: CA3006491C, US5519793 and CN111507379. Petition 870250012538, dated 02 / 14 / 2025, page 12 / 61 4 / 33

[015] Document CA3006491C discloses a method and system for monitoring and controlling a comminution process, specifically a grinding circuit for iron ore, using 3D imaging and particle size analysis. In CA3006491C, the method begins with measuring 3D reconstruction data for incoming ore and particle size data for outgoing ore, allowing a particle distribution profile at each stage of processing. Using these measurements, the system calculates the ore characteristics, including hardness, to regulate the grinding circuit. The imaging system, placed near a conveyor, captures 3D reconstruction data of the ore, while the particle size analysis equipment measures the outgoing ore to determine a characteristic value. This data is then processed to optimize the grinding circuit by adjusting parameters such as mill speed, water and ball addition, and load.

[016] CA3006491C, however, does not mention a computer vision monitoring method using images and deep learning for analyzing the degree of liberation of iron ore.

[017] Document US5519793 discloses a computer vision system and an image processing method for accurately determining the size and shape of coarsely textured, irregular, layered or similar pieces of ore. The system comprises a camera to produce an image of the ore, a digitizer to digitize the camera image, and a video processor to process the digitized image. An optional digital computer may be used to control the system as a whole and to determine the dimensions, mass, and surface area of ​​each piece of ore based on the processed image.The digitizer divides the camera image into a plurality of pixels, and a set of line rotation masks, each including a line of negative unit values ​​or center-weighted values ​​and a central positive value equal to the absolute value of the negative values, with the line arranged in a different orientation, are multiplied by the digital values ​​of each of the pixels to determine new values ​​for the pixels. A composite pixel field. Petition 870250012538, dated 02 / 14 / 2025, p. 13 / 61 5 / 33, using the lowest pixel values, is used to determine the sizes and shapes of the particles.

[018] US5519793, however, does not propose a solution for analyzing the degree of liberation of iron ore using computer vision. Furthermore, US5519793 does not use deep learning nor does it evaluate the colors of the ore for ore classification, so that US5519793 only evaluates size and shape, but not color.

[019] Document CN111507379 discloses a deep learning-based automatic ore identification and coarse sorting system, and the system comprises: a model building module, which is used to build a deep learning model for automatic ore identification and coarse sorting; a model training module, which is used to form a training set and a test set from images of the selected ore slurry block mixture, to train a deep learning model for automatic ore identification and coarse sorting; a recognition module, which is used to input images of the ore mixture and slurry blocks captured on the conveyor belt into a deep learning model trained for automatic ore recognition and coarse sorting, in order to obtain a real-time recognition result;and a screening module, which is used to control the conveyor belt to feed the batch of ore to the next procedure, according to the recognition result of the recognition module. If the mud content after recognition is less than a defined limit value, otherwise, it controls the high-pressure water gun to perform a spot wash on the recognized mud blocks and then feed the mud blocks to the next procedure. According to CN111507379, deep learning is creatively introduced into the ore screening process, achieving full automation of coarse ore screening.

[020] CN111507379, however, reveals a solution for the coarse screening of iron ore and does not mention the possibility of estimating the degree of liberation of the iron ore. Objectives of the Invention Petition 870250012538, dated 02 / 14 / 2025, page 14 / 61 6 / 33

[021] The present invention aims to present a method for calculating the degree of liberation of iron ore, using computer vision techniques, with the purpose of making this process cheaper, simpler, faster and more efficient. Brief Description of the Invention

[022] The present invention discloses a method for analyzing the degree of liberation of iron ore using computer vision, comprising at least the following steps: a) capturing images of particle samples using a microscope; b) segment the images to identify and isolate each particle, separating it from the background and neighboring particles, creating the dataset for model validation; c) create and validate the model used for segmentation; d) calculate the percentage of iron present in the segmented particles using cluster-based analysis algorithms; and e) Calculate the degree of liberation for each image taken from a microscope composed of a set of segmented particles using analysis algorithms based on image recognition and clustering to determine the percentage of iron present, calculating the degree of liberation using the Gaudin method.

[023] In the proposed method, mixed particles correspond to particles containing iron and quartz, quartz particles comprising less than a lower limit of iron and iron particles comprising more than an upper limit of iron, obtained in item e).

[024] In step e), particles that are quartz particles are excluded for the liberation calculation.

[025] The degree of liberation is calculated as the ratio between the liberation index of iron particles and the sum of the indices of iron particles and mixed particles. Petition 870250012538, dated 02 / 14 / 2025, page 15 / 61 7 / 33

[026] Step a) comprises the sub-step of preparing the iron ore samples, including separation by shaking to avoid overlapping.

[027] Furthermore, preferably, step a) comprises the sub-steps of: setting up the microscope magnification; setting up the microscope illumination; and adjusting the sample position, wherein microscope magnification is obtained using at least one eyepiece and at least one objective lens and images with inadequate illumination, noise or tilt are discarded.

[028] In step b), Convolutional Neural Networks (Mask R-CNN) are used for precise segmentation of mineral particles in the image.

[029] In step c), the Mask R-CNN architecture with transfer learning is used on the MS COCO dataset, to create the dataset for model validation.

[030] In step c), Mask R-CNN uses pre-trained backbones to create and validate the model used for particle segmentation.

[031] In step d), the SLIC clustering algorithm is used to identify particles (in shades of white and gray) and calculate the amount of iron in the mineral particles.

[032] In step d), certain upper and lower thresholds are applied during segmentation to correct for reflection or noise effects and to calculate the percentage of iron ore in each particle.

[033] In step d), a threshold of 55% (defined as the percentage of white tones) is defined to correct for reflection or noise effects in determining the iron content calculation of a segmented particle.

[034] In step e), aiming at the identification of mixed particles, compared to purely iron and quartz particles, the lower limit for the identification of a mixed particle is 0.1% of the percentage of iron identified in step d) and the upper limit of the particles is 70%. Petition 870250012538, dated 02 / 14 / 2025, page 16 / 61 8 / 33

[035] The present invention further discloses a method for analyzing the degree of liberation of iron ore using computer vision, to be applied to an ore sample collected on a conveyor belt. Brief Description of the Drawings

[036] The present invention will be described in more detail after the presentation of the figures, which contain a preferred configuration. The figures show:

[037] Figure 1 illustrates a flowchart with a configuration of the application of the method for analyzing the degree of liberation of iron ore using computer vision.

[038] Figure 2 illustrates the steps of a preferred configuration for applying the method.

[039] Figure 3 illustrates a configuration of the ore sample container being analyzed by an optical microscope.

[040] Figure 4 illustrates an image configuration of the ore with 50x zoom.

[041] Figure 5 illustrates a flowchart of the preferred configuration for applying steps a1) and a2) of the method.

[042] Figure 6 illustrates the classification stage of the method, where the outline of each particle is made, generating polygons in the same shape as the particles and defining their classes.

[043] Figure 7 illustrates an example of the feature extraction process, using Mask R-CNN from the state of the art[6].

[044] Figure 8 illustrates a flowchart for training and testing the segmentation model / backbone.

[045] Figure 9 illustrates the process of training and testing the images to choose the best zoom. Petition 870250012538, dated 02 / 14 / 2025, page 17 / 61 9 / 33

[046] Figure 10 illustrates three examples of particles with a grid.

[047] Figure 11 illustrates an example of signaling the exclusion of particles touching the edges (red), signaling the exclusion of particles without iron (yellow) and signaling of particles with iron (blue), following the Gaudin method.

[048] Figure 12 illustrates all the particles identified with the area containing iron highlighted and the percentage value of iron found, using the method for analyzing the degree of liberation of iron ore using computer vision.

[049] Figure 13 illustrates a first example of comparison of images obtained manually (a) by the Gaudin method and computationally (b) through the method for analyzing the degree of liberation of iron ore using computer vision.

[050] Figure 14 illustrates an example of a mixed particle image with a grid.

[051] Figure 15 illustrates a particle represented by numbers 1s and 0s.

[052] Figure 16 illustrates an example of a particle with iron calculation with varying segments.

[053] Figure 17 illustrates an iron threshold selection experiment with the SLIC algorithm.

[054] Figure 18 illustrates a graph with the result of choosing a threshold.

[055] Figure 19 illustrates an example of an image with particles reflecting light.

[056] Figure 20 illustrates the result of the calculation of the computational mineral liberation degree with upper and lower cut.

[057] Figure 21 illustrates a second example of comparing images obtained manually (a) by the Gaudin method and computationally (b) using the method for analyzing the degree of liberation of iron ore using computer vision. Petition 870250012538, dated 02 / 14 / 2025, page 18 / 61 10 / 33

[058] Figure 22 illustrates a comparison graph between the manual and computational methods.

[059] The reference numbers used in this application are: 1 - particles; 2 - microscope; 3 - container; 4 - iron particles; 5 - quartz particles; and 6 mixed particles. Detailed Description of the Invention

[060] Before the invention is described in detail, it should be understood that it is not limited to the specific steps of the described method, as such steps may vary. It should also be understood that the terminology used herein is only for the purpose of describing particular embodiments and is not intended to be limiting. It should be noted that, as used in the descriptive report and appended claims, the singular forms “a,” “an,” and “the” include singular and / or plural referents, unless the context clearly indicates otherwise. Furthermore, it should be understood that, in the case of parameter ranges delimited by numerical values ​​being provided, the ranges are considered to include these limiting values.

[061] It should also be understood that the modalities disclosed herein should not be understood as individual modalities that are unrelated to each other. The characteristics discussed with one modality should also be disclosed in connection with other modalities shown herein. If, in one case, a specific characteristic is not disclosed with one modality but with another, the person skilled in the art would understand that this does not necessarily mean that the characteristic is not intended to be disclosed with the other modality. The person skilled in the art would understand that the essence of this request is to disclose the characteristic also for the other modality, but that only for the sake of clarity and to keep this descriptive report to a manageable volume, this has not been done.

[062] The present invention relates to a method for analyzing the degree of liberation of iron ore using computer vision. To achieve this objective, an approach based on Convolutional Neural Networks is employed. Petition 870250012538, dated 02 / 14 / 2025, page 19 / 61 11 / 33 Region-based CNNs (R-CNN), a deep learning architecture specifically designed for object detection and segmentation in images.

[063] To achieve the proposed objective, the present method used sample preparation, image acquisition, dataset creation, machine training, validation, analysis and comparison with other methods.

[064] Sample preparation and image acquisition involves separating samples and obtaining images, checking the illumination and microscope settings at different magnifications, in order to determine the combination with the best result. This describes the preparation of iron ore samples for image acquisition using an optical microscope at magnifications of 16x, 32x, and 50x to determine the ideal magnification for creating the images that make up the dataset. The quality and quantity of the images are important for the following steps (especially for reproducibility), therefore mentioning the microscope adjustment and illumination is part of this step.

[065] The microscope model 2 used in this configuration does not have its own light source, so external light sources were used, which were defined during the execution of the method. The position of incidence of the light also influences the result of the experiment, therefore light emission positions that improve the model's result, reducing shadowing and contributing to the uniformity of the illumination, were analyzed.

[066] The creation of a dataset with segmented images using cooperative learning and transfer learning techniques after an experiment to define the best image magnification adjustment was carried out to choose the ideal magnification.

[067] A deep learning model is an advanced artificial intelligence tool designed to identify and learn complex patterns in data. It works in several stages organized in layers, where each layer processes the information received from the previous one, gradually refining the understanding of the original data. Petition 870250012538, dated 02 / 14 / 2025, page 20 / 61 12 / 33

[068] A framework, in turn, is software that brings together a set of tools and resources that facilitate the development and execution of deep learning models. It provides support for both the training of models and the practical application of the results obtained.

[069] In a preferred embodiment of the present invention, a framework is used that already includes pre-trained models on a widely used dataset. These models serve as a basis for accelerating development by applying a technique called transfer learning. This approach leverages the learning already acquired by a model for a similar task, reducing the need to collect large volumes of new data.

[070] To acquire images at different magnifications, microscope 2 was configured and adjusted by the operator by changing its eyepiece and objective lenses following this configuration: • Images with 16x magnification: 4x eyepiece; 4x objective lens; • Images with 32x magnification: 4x eyepiece; 8x objective lens; • Images with 50x magnification: 5x eyepiece; 10x objective lens.

[071] Based on the choice of the ideal magnification, four hundred images were created which were used in the next steps.

[072] Of these four hundred images, two hundred were segmented using the CML (Cooperative Machine Learning) technique. In this stage, nineteen images of each zoom were sampled, produced with lighting and particle arrangement of the samples performed by the laboratory technician specialized in this type of acquisition. With the nineteen samples of each zoom, the first training of the first neural network was carried out with the objective of choosing the images with the ideal zoom.

[073] Machine training involves choosing the best segmentation model applied to test images. Petition 870250012538, dated 02 / 14 / 2025, page 21 / 61 13 / 33

[074] For this, the nineteen images generated from each magnification (16x, 32x, 50x) were used with a segmentation algorithm provided by the Detectron2 framework.

[075] This framework provides several backbones with the Mask R-CNN algorithm. One was chosen that already had training data in some public database for use with the transfer learning technique.

[076] Once the use of frameworks such as Detectron2 to facilitate the training and implementation of deep learning models has been established, it is important to highlight the architecture used in this context. Region-based Convolutional Neural Networks (R-CNNs) are one of the main approaches employed for object detection in images, due to their ability to efficiently identify and process regions of interest.

[077] Region-based Convolutional Neural Networks (R-CNNs) are a deep learning architecture developed for object detection in images. They work by extracting regions of interest (ROIs) from an image, generating candidates from algorithms such as Selective Search. These regions are then passed through a CNN to extract features, which feed a classifier to identify the object present and predict the coordinates of the bounding box.

[078] Fast R-CNN is an evolution of the R-CNN model that improves efficiency and reduces the computational cost of the object detection process. Unlike R-CNN, where regions of interest (ROIs) are processed individually by a convolutional neural network, Fast R-CNN processes the entire image by a CNN only once, extracting a feature map. The ROIs are then projected onto this feature map using an operation called ROI pooling, which reduces the dimensions of each ROI to a fixed size. The features of the ROIs are used to classify objects and predict bounding boxes, making the process significantly faster and more integrated compared to the original R-CNN. Petition 870250012538, dated 02 / 14 / 2025, page 22 / 61 14 / 33

[079] Mask R-CNN, on the other hand, is an extension of Faster R-CNN, which not only detects objects and predicts bounding boxes, but also performs instance segmentation. It adds a new branch to the architecture, capable of generating binary masks for each detected object, allowing the exact shape of each object in the image to be identified. The main difference lies in Mask R-CNN's focus on providing detailed segmentation of individual instances, in addition to detection, while R-CNN focuses only on locating and classifying objects without precise segmentation.

[080] With the advanced capability of Mask R-CNN to perform detailed and precise segmentation of individual instances, the present invention applies this computer vision technology to solve specific challenges in the analysis of mineral particles. Using deep learning techniques, the images required for this analysis are preferably captured by means of an optical microscope 2, ensuring the quality required for the segmentation and classification of the particles.

[081] The 'R-101-FPN-3x' backbone was chosen, and its training was performed on the 'MS COCO' database. To use this backbone with the trained data, it is necessary to define two configuration parameters before running the model training and testing. The two items required for configuration are the backbone and the weights of the already trained model, both configured with the value: • “COCO-InstanceSegmentation / mask_rcnn_R_101_FPN_3x.yaml”.

[082] The features of an image are essential information, such as shapes, textures, or patterns, that help deep learning algorithms recognize and detect objects. In MASK R-CNN, these features are captured through a process called feature map generation, which creates a detailed representation of the important parts of the image.

[083] This feature extraction process is configured through an architectural element called the backbone. The backbone functions as the base Petition 870250012538, dated 02 / 14 / 2025, page 23 / 61 15 / 33 is responsible for identifying and organizing image information. It can be adjusted to choose the most efficient configuration, depending on the images and the problem's objective, ensuring more accurate analysis results. For training and testing, all generated images were classified and segmented using the MAKESENSE.AI tool.

[084] After the necessary configurations were made, training and testing of each set of images was carried out, with nineteen images being selected, fifteen for training and four for testing. Few images were used because they were used to choose the best magnification and at the beginning of the cooperative machine learning process. This training and testing can be exemplified in Figure 9.

[085] The result of this process is how well the model is able to accurately segment for each set of test data provided to it.

[086] Object detection in images with CNNs requires more elaborate metrics for calculating results. One of these metrics is the Mean Average Precision (mAP), which is the result of the arithmetic mean of the Average Precision (AP) of all images in the analyzed set. The AP of each image is linked to the terms Intersection over Union (IoU) and to the hits and misses of the objects found in the image.

[087] For each object identified in an image, the predicted bounding box is superimposed on the actual bounding box. The IoU is calculated from this overlap, which is the result of dividing the area of ​​the intersection by the area of ​​the union of the bounding boxes. The IoU can range from 0 to 1, with 0 indicating completely different bounding boxes and 1 indicating identical bounding boxes.

[088] The AP of the image is calculated using the IoU of each object. To do this, it is necessary to define a threshold and compare it with the IoU. If the IoU is greater than the threshold, it is a true positive; otherwise, it is a false positive. Petition 870250012538, dated 02 / 14 / 2025, page 24 / 61 16 / 33

[089] With metrics defined, such as AP and mAP, ensuring accuracy in object detection and segmentation in images, it becomes possible to apply advanced methods for processing and analyzing the data obtained.

[090] Analyzing the mask metrics AP (Average Precision) and mAP (mean Average Precision), Table 1 shows the results, in which it can be seen that the best result was for the 50x magnification with a mAP of 74.45 and that the difference between the Mixed and Iron APs was also the smallest, with a value of 2.05 points (IRON AP - MIXED AP), while the other magnifications have differences greater than eight points. Table 1 - Image magnification comparison table Number of images Mixed quantity Iron quantity Magnification AP mixed AP iron mAP 19 196 27 16x 68.74 60.32 64.53 19 82 17 32x 71.81 62.62 67.21 19 58 7 50x 73.42 75.47 74.45

[091] For the final process of generating new images, the CML (cooperative machine learning) technique was used, in which part of the work is done by a model obtained by transfer learning and adjusted by a human and fed back into the model. Twenty initial images were created with the chosen magnification of 50x. These images were classified and segmented in the MAKESENSE.AI tool and after that submitted to the training process and added to the training and test image database, in this case in the proportion of 90% / 10%.

[092] Following the training process described at the beginning of this subsection, a new model is created and made available as a template to be applied to the next images.

[093] The next images created went through the CML process, that is, when the images are generated, they are first submitted to the available model, resulting in a pre-classification and segmentation, and then returned for the necessary adjustments made by the specialists. After the adjustments, they were added Petition 870250012538, dated 02 / 14 / 2025, page 25 / 61 17 / 33 to the training and test image database and a new training was performed with all the images in the database and a new model was made available.

[094] At the end of the process, two hundred classified and segmented images were generated. In addition, two hundred new images that did not go through this process are available in the database of images without classification and segmentation and which were subsequently used as a basis for validating the method.

[095] For use in the training and testing process, only the first two hundred were used, with 90% for training and 10% for testing.

[096] In this stage, the focus is on the implementation and analysis of a Mask RCNN model using the Detectron2 framework. The central objective is to evaluate the performance and effectiveness of this model when trained with different backbones.

[097] Through this invention, the aim is not only to understand the nuances of these architectures, but also to identify which backbone best optimizes object detection and segmentation. Careful analysis of these variants has allowed for a deeper understanding of the capabilities and limitations of Mask R-CNN within the proposed context.

[098] To perform the training using transfer learning, it was necessary to choose the backbones that would be tested from those available, as shown in Table 2. Table 2 - List of available backbones Name IR Schedule Training Time (without iteration) Interference Time (without image) Training Memory (GB) Box AP Mask AP R50-C4 1x 0.584 0.110 5.2 36.8 32.2 R50DC5 1x 0.471 0.076 6.5 38.3 34.2 R50-FPN 1x 0.261 0.043 3.4 38.6 35.2 R50-C4 3x 0.575 0.111 5.2 39.8 34.4 Petition 870250012538, dated 02 / 14 / 2025, page 26 / 61 18 / 33 R50DC5 3x 0.470 0.076 6.5 40.0 35.9 R50- FPN 3x 0.261 0.043 3.4 41.0 37.2 R101- C4 3x 0.652 0.145 6.3 42.6 36.7 R101- DC5 3x 0.545 0.092 7.6 41.9 37.3 R101FPN 3x 0.340 0.056 4.6 42.9 38.6

[099] All 1x models in the 'lrsched' column were eliminated, meaning that these models were trained with 12 epochs on the standard MS COCO dataset. On the other hand, the 3x models were trained with 37 epochs. Thus, it can be observed that the latter present a larger mask mAP, maintaining the same inference times, training time, and memory used. Therefore, these were the models chosen. The term “lrsched” is an abbreviation for learning rate schedule and indicates the number of epochs used in training. An epoch is a complete pass of an algorithm over the entire dataset.

[100] Of the remaining models, it is possible to observe that the use of resources, such as training memory, inference time and training time, have higher averages when compared to the R50-FPN and R101-FPN models for similar or smaller mAP, therefore, these two models (or backbones) were used for the choice of models.

[101] The X101-FPN model has a slightly higher mAP mask value than the R50-FPN and R101-FPN models, but uses more memory and has more than double the training and inference times. Therefore, this model was also discarded because the cost of using the resources is not advantageous compared to the gain in the mAP mask value.

[102] For model training, the Colab environment (GOOGLECOLAB, sd) was selected with an Nvidia Tesla T4 GPU that has 16GB of memory, 320 Turing Tensor cores and 2560 CUDA cores. Petition 870250012538, dated 02 / 14 / 2025, page 27 / 61 19 / 33

[103] After selecting the execution environment, the parameters for running the model training and inference tests on the images were defined. One of the parameters is the choice of the execution framework itself, which is Detectron2[7]. The version of Detectron2 used was 0.6.

[104] Another parameter was the definition of the Mask R-CNN architecture with the chosen backbone, that is, for each execution one of the two chosen backbones was configured: R50-FPN and R101-FPN. Along with the backbone, the locations where the already trained data would be used, using the transfer learning technique, were also configured. Each backbone was configured as follows: • R50-FPN: “COCO-InstanceSegmentation / mask_rcnn_R_50_FPN_3x.yaml”. • R101 -FPN: “COCO-InstanceSegmentation / mask_rcnn_R_101_FPN_3x.yaml”.

[105] Other parameters are the number of training epochs: 1000, number of classes: 2, training json files with 178 images and tests with 22 images.

[106] The training was carried out with the settings defined in the previous subsections. Of the 178 images selected, 448 iron 4 particles and 2127 mixed 6 particles were classified and segmented. Of the 22 images used for the tests, 66 particles were iron 4 particles and 222 were mixed 6 particles.

[107] Table 3 shows the results of the training and testing of the two chosen backbones with their processing times and AP and mAP metrics. It can be seen that the model with the R101-FPN backbone obtained better results. After choosing the best model, it was saved in a file for use in the next steps. Table 3 - Training and testing results table for the models Backbone Training Time Test Time AP Iron AP Mixed mAP R50-FPN 20 minutes 24.6 s 83.77 81.77 82.77 R101-FPN 36 minutes 34.5 s 87.5 83.26 85.38 Petition 870250012538, dated 02 / 14 / 2025, page 28 / 61 20 / 33

[108] For the segmentation of each particle to determine its iron content, the SLIC algorithm was chosen, which demonstrated better results in the initial tests. Based on this selection, experiments were prepared for the ideal parameterization of the algorithm in the context of this invention.

[109] Simple Linear Iterative Clustering (SLIC) is an image segmentation algorithm that is an adaptation of K-means to generate superpixels, which is the process of dividing a digital image into multiple segments or sets of similar pixels.

[110] For this purpose, 117 particle images were extracted and separated from the segmented images with the chosen model and submitted to mining experts so that the percentage of iron in each of these particles could be estimated.

[111] The need to estimate the percentage of iron in the particle is related to the type of particle, iron, quartz or mixed. In this example, the particles are under a grid to help in the process of estimating the percentage of iron. For the experiment used, the amount of iron in an iron particle 4 is 100%, being calculated as the total of squares defined in the grid minus the blank squares. For quartz particles 5 the amount of iron is 0%. To visually estimate the amount of iron in each mixed particle 6, a method was defined, in which the specialist defines the grid over the image of the particle, with the aid of image visualization software, as shown in Figure 14.

[112] For each square the specialist defined a value of 0 if it is the background of the image, not containing any part of the particle, or 1 if the square contains the particle, even if only partially. Figure 15 demonstrates this scheme for Figure 14.

[113] The calculation of the percentage estimate of iron in an image is defined by Equation 1: Petition 870250012538, dated 02 / 14 / 2025, page 29 / 61 21 / 33 PFP = NQFE * ( ^) * 100 (Eq. 1)

[114] Where PFP is the estimate of the percentage of iron in the particle; NQFE is the number of iron squares in the particle; NQP is the number of squares of the particle, which can be understood as the area of ​​the particle and (1 / NQP) as the equivalent area of ​​a square.

[115] To define the particle area value (NQP), the figure represented by 1s and 0s is used, as in the example in Figure 15, where the particle area can be defined by counting the 1s in the figure or by Equation 2: NQP = NQT - NQFU (Eq. 2)

[116] Where NQT is the total number of squares in the Figure, or the total area of ​​the figure and NQFU is the number of background squares in the figure.

[117] The total area of ​​the figure (NQT) can be obtained by counting all the squares in the figure or by multiplying the number of squares in a row by the number of squares in a column, that is, it can be represented by Equation 3: NQP = NQT * NQFU (Eq. 3)

[118] Where NQL is the number of squares in a row and NQC is the number of squares in a column.

[119] Since the number of iron squares in the particle (NQFE) is important information, the expert defines it as 1 when a square is completely filled with iron and estimates a fraction of iron when the square is not completely filled. Petition 870250012538, dated 02 / 14 / 2025, page 30 / 61 22 / 33

[120] For the computational calculation of the iron estimate in the particles, experiments were carried out using the SLIC algorithm, processing the 117 images of particles separated for this stage, defining parameters for the number of segments and threshold values ​​for the iron color.

[121] The SLIC (Simple Linear Iterative Clustering) algorithm is an unsupervised algorithm that groups pixels of an image into regions called superpixels based on their spatial proximity and color similarity, where a mandatory parameter is the number of clusters, or segments, desired.

[122] To define the amount of iron, it is necessary to define a cutoff threshold for the separation between the darker and lighter superpixels because iron ore is not entirely defined by the color black, and can assume colors of different shades, which only a human specialist can define.

[123] To determine the percentage of iron in a particle, an experiment was carried out that transformed the image into grayscale and for each segment returned by SLIC the average of the pixel values ​​contained in that segment was calculated, where 0 corresponds to entirely iron and 1 to entirely quartz.

[124] To calculate the percentage of iron, a value between 0 and 1 (threshold) was defined, so that average values ​​lower than 1 were considered iron, otherwise quartz. By summing the number of all iron pixels per segment (defined by the threshold) and dividing the result by the total number of pixels in all segments of the image, it was possible to obtain the percentage of iron in the particle.

[125] To visualize the iron segmentation in the particle, after defining the threshold, all segments representing iron in Figure 16 were colored green and this is an example of a particle with the percentage of iron calculated with segment values ​​equal to 500 and another with 100 and a threshold of 0.55 for the iron color, along with the original image, obtaining the values ​​of 11.4% and 9.34% respectively. Petition 870250012538, dated 02 / 14 / 2025, page 31 / 61 23 / 33

[126] To find the best combination of segment number and threshold values ​​that minimizes the error in relation to manual estimation per particle with the SLIC algorithm, as shown in Figure 17.

[127] As shown in Figure 17, the experiment was carried out by choosing the following threshold values: 0.1, 0.2, 0.3, 0.4, 0.5, 0.55, 0.6 and 0.65, and the segment values ​​were 100, 200, 300, 400, 500. Table 4 shows the result of all the combinations performed. Table 4 - SLIC result with RMSE and combinations of Segments and Thresholds Experiment Threshold Number of Segments RMSE Exp01 0.1 100 51.81 Exp02 0.2 100 48.78 Exp03 0.3 100 31.86 Exp04 0.4 100 18.52 Exp05 0.5 100 11.28 Exp06 0.55 100 9.25 Exp07 0.6 100 12.32 Exp08 0.65 100 19.46 Exp09 0.1 200 51.81 Exp10 0.2 200 48.48 Exp11 0.3 200 31.53 Exp12 0.4 200 18.39 Exp13 0.5 200 11.6 Exp14 0.55 200 9.76 Exp15 0.6 200 12.17 Exp16 0.65 200 19.36 Exp17 0.1 300 51.81 Exp18 0.2 300 48.14 Exp19 0.3 300 31.52 Exp20 0.4 300 18.73 Exp21 0.5 300 11.83 Exp22 0.55 300 9.78 Exp23 0.6 300 12.24 Exp24 0.65 300 19.41 Exp25 0.1 400 51.81 Exp26 0.2 400 48.09 Exp27 0.3 400 31.5 Exp28 0.4 400 18.85 Exp29 0.5 400 11.92 Exp30 0.55 400 9.96 Petition 870250012538, dated 02 / 14 / 2025, page 32 / 61 24 / 33 Exp31 0.6 400 12.36 Exp32 0.65 400 19.54 Exp33 0.1 500 51.81 Exp34 0.2 500 48.12 Exp35 0.3 500 31.46 Exp36 0.4 500 18.94 Exp37 0.5 500 12.16 Exp38 0.55 500 10.07 Exp39 0.6 500 12.44 Exp40 0.65 500 19.39 Table 5 - Table showing the results of the 0.55 threshold values. Segments Threshold RMSE 100 0.55 9.25 200 0.55 9.76 300 0.55 9.78 400 0.55 9.96 500 0.55 10.07

[128] To choose the best threshold, all the results were plotted on graphs and the one with the lowest average error was chosen, which was the threshold of 0.55 for all segments.

[129] Figure 18 shows the graphs with Root Mean Square Error (RMSE) values, which is a metric used to check the difference between predicted and actual values, from the comparison between the values ​​defined by the manual method and the computational method performed by SLIC with segment values ​​of 100, 200, 300, 400 and 500 and with thresholds of 0.1, 0.2, 0.3, 0.4, 0.5, 0.55, 0.6 and 0.65.

[130] The manual method used by the specialist described below is in accordance with the guidelines proposed by Gaudin, according to which the particles are identified one by one, recording the number of free particles of a given species and evaluating the quantity of this species contained in the mixed particles. In Gaudin's method, the particles are arbitrarily considered to be composed of twenty area units and the grids in the particles are considered so that each quadrant represents five units, making it easier to assess the content of the particles. Petition 870250012538, dated 02 / 14 / 2025, page 33 / 61 25 / 33

[131] In order for the comparison process of the manual method and the computational method to be compatible with each other, some steps were followed by the specialist at the time of visual estimation of the degree of liberation of an iron ore image.

[132] These steps relate to particles that touch the edge of the image, particles that do not contain iron, the division of the particle into four quadrants and the value given in each quadrant.

[133] The comparison image dataset in this experiment consisted of 60 images that underwent the manual method for estimating the degree of mineral liberation. The manual method has four sub-steps, which will be described below.

[134] The first sub-step of the method was the elimination of particles that touch the edges of the image. Since it is not possible to define their real shape, they are eliminated from the particle counting process.

[135] Subsequently, the second sub-step was carried out, which was the elimination of particles that do not contain iron, since they are not part of the calculation of the degree of mineral liberation, as they are gangue minerals. Figure 11 shows an example of signaling this sub-step, in which the eliminated particles are marked with a yellow X.

[136] In the third sub-stage, a signaling system was used in which the remaining particles had some amount of iron, therefore they were signaled in four quadrants, representing the Cartesian plane. Then a sequential numbering was defined for each particle to facilitate its identification in this sub-stage. An example of this signaling is shown in Figure 11, where the particles that have iron are signaled with the color blue and a number.

[137] In the fourth and final sub-step of the method, the estimated percentage of ore in each particle and the degree of mineral liberation in each image were calculated. Petition 870250012538, dated 02 / 14 / 2025, page 34 / 61 26 / 33

[138] For the calculation in each particle, a maximum area value of five was defined in each quadrant, resulting in a value of twenty for the total maximum area of ​​the particle.

[139] The expert then estimated the amount of iron for each quadrant and added the four quadrants together as the total area value of the particle.

[140] To calculate the degree of mineral liberation in each image, the values ​​of particles with a maximum value of twenty, i.e., entirely iron (or liberated), were summed and divided by the total sum of the values ​​of all particles, i.e., the sum of the liberated particles with the iron calculated for the mixed particles 6.

[141] This result is the representative fraction of the degree of mineral liberation of the image, which multiplied by one hundred becomes the percentage of the degree of liberation, as exemplified in Equation 4: (Σ PRTL) * 20GL (%) = (Σ PRTL) * 20 + Σ FPM *100(Eq. 4)

[142] Where GL is the degree of liberation; PRTL represents the iron particles 4, released and FPM represents the iron values ​​contained in the mixed particles 6.

[143] As described in the previous section, to obtain the Gaudin liberation degree it is important to identify particles with 100% iron, which is simple for the human analyst, but is never achieved by the proposed method, since the colors are affected by light reflection and the fact that they are not two-dimensional. Figure 19 is an example of an image with light reflection.

[144] The data in Table 6 show that no iron 4 particles were found with a configuration of 100 segments and a threshold of 0.55, so all particles are recognized as mixed particles 6. Petition 870250012538, dated 02 / 14 / 2025, pp. 35 / 61 27 / 33 Table 6 - Table of particle quantities by manual and computational method Method Iron particles Mixed particles Total particles Manual 124 372 496 Computational 100 seconds, Threshold = 0.55 0 579 579

[145] Through the analysis of these results, a process was proposed so that a fraction of the mixed particles 6 from the computational method would be eliminated from the process because they presented low values ​​not measurable by the operator and that others would become iron particles 4, given the impossibility of recognition by the computational method of iron particles 4, as evaluated by the operator (Table 6).

[146] Thus, lower and upper cutoff values ​​for the percentage of iron were defined, in which particles with little iron were eliminated, below the lower value, and particles with 100% iron were counted above an upper value, being used as released particles. With this, several tests were carried out with different combinations of lower and upper cutoffs that came close to the manually defined value.

[147] The upper values ​​defined for testing were: 70%, 75%, 80%, 85% and 90% and the lower values ​​were 0.05%; 0.1%; 0.2% and 1.0%. To verify which combination of values ​​had the smallest error, the RMSE (Root Mean Square Error) and Mean Absolute Error (MAE) metrics were used, which is a measure representing the average absolute difference between predicted and actual values.

[148] For the execution of the first part of the experiment the procedure was: • 1. Define the upper and lower cutoff values, where particles with iron content below the lower cutoff were discarded and particles above the upper cutoff were defined with an area value of twenty, equivalent to the maximum value defined in the manual procedure; Petition 870250012538, dated 02 / 14 / 2025, pp. 36 / 61 28 / 33 • 2. The calculated values ​​for the mixed particles 6 defined between the lower and upper cutoff values ​​were multiplied by twenty to adjust to the operator's value; • 3. After multiplication, if the particle value fell below one, it would be rounded down to one, which is the smallest value assigned by the operator (1 / 20).

[149] This result can be seen in Figure 20, where the degree of liberation value was 0.909 or 90.9%, with a total of ten particles found, two defined as iron (segm_8, segm_9) with a value of 20, 4 defined as mixed (segm_[0,1,4,5]) with a value of 1 and 4 particles discarded (segm_[2,3,6,7]).

[150] In this example, the amount of iron in mixed particles 6 was less than 5%, so it was adjusted to the lowest value, which is 1, equivalent to 5% (1 / 20). Other values ​​defined in the experiment were: lower cutoff equal to 0.001 (0.1%), upper cutoff equal to 0.7 (70%), threshold equal to 0.55 and segments equal to 500.

[151] For the execution of the second part of the experiment, the procedure was to put in a table all the results of the combinations of lower and upper cut along with the variation of the segments, verifying them with the result of the manual procedure, generating the calculation of the RMSE error for 60 images. Table 7 shows part of the results obtained and which will be detailed in the next section. Table 7 - Table of partial computational results of the degree of image release with RMSE. Lower Upper No. of segments IMG 1 IMG 2 IMG 3 *** IMG 60 MAE DP RMSE Manual Manual 71.42 89.88 54.79 *** 86.95 - - - 0.1%70% 100 83.33 92.81 62.15 *** 96.50 7.01 5.74 9.03 0.1%75% 100 83.33 92.81 62.15 *** 96.50 7.34 6.40 9.70 0.1%80% 100 83.33 92.81 62.15 *** 96.50 10.39 13.52 16.96 0.1%85% 100 83.33 92.81 62.15 *** 96.50 16.29 18.98 24.90 0.1% 90% 100 0.00 71.73 62.15 *** 66.53 29.02 26.64 39.24 Petition 870250012538, dated 02 / 14 / 2025, pp. 37 / 61 29 / 33 0.1% 70% 500 76.92 91.27 61.72 *** 93.44 3.38 2.98 4.49 0.1% 75% 500 0.00 91.27 61.72 *** 93.44 6.56 11.87 13.48 0.1% 80% 500 0.00 91.27 61.72 *** 93.44 10.89 16.99 20.06 0.1% 85% 500 0.00 71.36 61.72 *** 93.44 22.12 24.15 32.60 0.1% 90% 500 0.00 49.31 31.87 *** 64.33 40.82 27.88 49.31

[152] Table 7 shows the results of the degree of release, where the values ​​of 500 segments, lower cutoff of 0.1% and upper cutoff of 70% obtained the best RMSE result, with RMSE values ​​of 4.49 and MAE of 3.38%. With this result, the parameters chosen with the best results were: threshold equal to 0.55; segments equal to 500; lower cutoff equal to 0.1% and upper cutoff equal to 70%.

[153] The first line of Table 7 shows the result of the manual procedure, which was used as a basis for comparison to determine the reference values ​​used for the release calculation.

[154] To compare the results between Gaudin's manual method and the computational method, comparison experiments were carried out and the result of two images was obtained.

[155] A first example, in which the proposal adheres to Gaudin's method, can be seen in Figure 13, which has two images, (a) representing the manual analysis performed by the specialist using Gaudin's method and (b) representing the analysis performed by the computational method. The two images were obtained from the same image and processed to represent each method.

[156] In both images, it was possible to observe that the particles that were eliminated by touching the edge, represented with a red X, were the same, both by the manual and computational methods.

[157] Of the particles eliminated because they do not contain or practically do not contain iron (within the defined limits), represented by the yellow X, only the one that Petition 870250012538, dated 02 / 14 / 2025, pp. 38 / 61 30 / 33 number five was not eliminated by the computational method. Four particles remained for analysis in the image using the manual method and five using the computational method.

[158] Table 8 presents the results of the analyses performed for the example, it can be seen that the degree of mineral liberation of method (a) was 80% and of method (b) was 81.15%, that is, a difference, in this case, of 1.15 points from the Gaudin measurement, or 1.4%. Table 8 - Table with analysis results Item Manual Iron % Computational Iron % Manual Points Computational Points Total % Particle 1 5 2.33 1 1 - Particle 2 5 1.82 1 1 - Particle 3 15 8.22 3 1.64 - Particle 4 100 92.7 20 20 - Particle 5 - 0.22 - 1 - Sum of iron - - 20 20 - Sum of mixed - - 5 4.64 - Manual mineral liberation % - - - - 80.00 Computational mineral liberation % - - - - 81.15

[159] In a second example, where the methods obtained different results, it is demonstrated in Figure 21. The greatest differences were found in particles 1, 2, and 6, where there was an estimate in the sum of points of 3.53 by the computational method and 7 by the manual method. All other particles were eliminated or analyzed equally in both methods.

[160] Table 9 presents the results of the analyses performed for the example; it can be seen that the degree of mineral liberation of method (a) was 68.96% and of method (b) was 78.32%, that is, a difference, in this case, of 9.36 points from the measurement. Petition 870250012538, dated 02 / 14 / 2025, pp. 39 / 61 31 / 33 of Gaudin, or 13.57%. This contributed to the increase in the final result of the overall average error of the images. Table 9 - Table with analysis results Item Manual Iron % Computational Iron % Manual Points Computational Points Total % Particle 1 15 7.67 3 1.53 - Particle 2 10 2.78 2 1 - Particle 3 5 0.23 1 1 - Particle 4 5 0.93 1 1 - Particle 5 100 92.61 20 20 - Particle 6 10 0.92 2 1 - Sum of iron - - 20 20 - Sum of mixed - - 9 5.53 - Manual mineral liberation % - - - - 69.96 Computational mineral liberation % - - - - 78.32

[161] The comparative analysis between the result of the computational method and the result obtained by the expert, using Gaudin's manual method, revealed a difference of 4.49% in the mean squared error for a set of 60 samples. This squared error value highlights the ability of the computational method to provide results in accordance with the expert's estimate and consistency with the manual method.

[162] The graph in Figure 22 illustrates this difference, highlighting the effectiveness of the computational method in approximating the results obtained by the expert and providing an automated and standardized analysis compared to manual analysis, which can be crucial in applications that require standardization and processing speed.

[163] Table 10 shows the processing time, in minutes and seconds, of 10 images using manual and computational methods. The relationship between the Petition 870250012538, dated 02 / 14 / 2025, pp. 40 / 61 32 / 33 average computational (42s) and manual (242s) is a value of 17%, indicating a processing time approximately 5 times shorter than the manual time. Table 10 - Table showing the manual and computational processing times for 10 images. Image Manual time (min's”) Computational time (min's”) Time saving (times) 1 3'29” 0'34” 6 2 4'18” 0'42” 6 3 4'52” 0'49” 5 4 3'49” 0'53” 4 5 3'01” 0'33” 5 6 4'27” 0'59” 4 7 3'21” 0'42” 4 8 4'10” 0'53” 4 9 4'22” 1'08” 3 10 3'30” 0'28” 7

[164] It should be noted that the drawings presented are not necessarily to scale, being merely conceptual in nature. Nevertheless, it is expressly provided that all combinations of elements that perform the same function substantially in the same way to achieve the same results as the elements now claimed are within the scope of the present invention. Finally, it should be noted that the scope of protection of the present invention covers other possible variations, not being limited solely by the content of the claims alone, including possible equivalents. References [1] LUZ, Adão Benvindo da; SAMPAIO, João Alves; FRANÇA, Silvia Cristina Alves. Tratamento de minérios. [S.l.]: CETEM / MCT, 2010. [2] FANDRICH, Rolf et al. Modern SEM-based mineral liberation analysis. International Journal of Mineral Processing, Elsevier, v. 84, n. 1-4, p. 310-320, 2007. [3] SYLVESTER, Paul J. Use of the mineral liberation analyzer (MLA) for mineralogical studies of sediments and sedimentary rocks. Mineralogical Association of Canada Short Course, v. 42, p. 1-16, 2012. Petição 870250012538, de 14 / 02 / 2025, pág. 41 / 61 33 / 33 [4] WARLO, Mathis et al. Automated quantitative mineralogy optimized for simultaneous detection of (precious / critical) rare metals and base metals in a production-focused environment. Minerals, MDPI, v. 9, n. 7, p. 440-459, 2019. [5] HRSTKA, Tomas et al. Automated mineralogy and petrology-applications of TESCAN Integrated Mineral Analyzer (TIMA). Journal of Geosciences, Czech Geological Society, v. 63, n. 1, p. 47-63, 2018. [6] HE, Kaiming; GKIOXARI, Georgia et al. Mask r-cnn. In: PROCEEDINGS of the IEEE international conference on computer vision. [S.l.: s.n.], 2017. P. 2961-2969. [7] WU, Yuxin et al. Detectron2. [S.l.: s.n.], 2019. https: / / github.com / facebookresearch / detectron2. [8] LIN, Tsung-Yi et al. Microsoft coco: Common objects in context. In: SPRINGER. EUROPEAN conference on computer vision. [S.l.: s.n.], 2014. P. 740-755. Petição 870250012538, de 14 / 02 / 2025, pág. 42 / 61

Claims

1 / 3 CLAIMS 1. Method for analyzing the degree of liberation of iron ore using computer vision, characterized by comprising at least the following steps: a) capturing images of particle samples using a microscope (2); b) segmenting the images to identify and isolate each particle, separating it from the background and neighboring particles, creating the dataset for model validation; c) creating and validating the model used for segmentation; d) calculating the percentage of iron present in the segmented particles using analysis algorithms based on clustering; and e) calculating the degree of liberation for each image taken from a microscope composed of a set of segmented particles using analysis algorithms based on image recognition and clustering to determine the percentage of iron present, calculating the degree of liberation by the Gaudin method;wherein, mixed particles (6) correspond to particles containing iron and quartz, quartz particles (5) comprise less than a lower limit of iron and iron particles (4) comprise more than an upper limit of iron; wherein, in step e), particles that are quartz particles (5) are excluded; and wherein, the degree of liberation is calculated as the ratio between the liberation index of iron particles (4) and the sum of the indices of iron particles (4) and mixed particles (6).

2. Method for analyzing the degree of liberation of iron ore using computer vision, according to claim 1, characterized in that step a) Petition 870250012538, dated 02 / 14 / 2025, page 43 / 61 2 / 3 comprises the sub-step of preparing the iron ore samples, including separation by shaking to avoid overlapping.

3. Method for analyzing the degree of liberation of iron ore using computer vision, according to claim 1, characterized in that step a) comprises the sub-steps: setting up the microscope magnification (2); setting up the microscope illumination (2); and adjusting the sample position; wherein, the microscope magnification (2) is obtained using at least one eyepiece lens and at least one objective lens; and wherein, images with inadequate illumination, noise or tilt are discarded.

4. Method for analyzing the degree of liberation of iron ore using computer vision, according to claim 1, characterized in that, in step b), Convolutional Neural Networks (Mask R-CNN) are used for precise segmentation of mineral particles in the image.

5. Method for analyzing the degree of liberation of iron ore using computer vision, according to claim 1, characterized in that, in step c), the Mask R-CNN architecture with transfer learning is used on the MS COCO dataset, to create the dataset for model validation.

6. Method for analyzing the degree of liberation of iron ore using computer vision, according to claim 5, characterized in that, in step c), the Mask R-CNN uses pre-trained backbones for the creation and validation of the model used for particle segmentation.

7. Method for analyzing the degree of liberation of iron ore using computer vision, according to claim 1, characterized by, in step d), using the SLIC clustering algorithm for particle identification and calculation of the amount of iron in the mineral particles. Petition 870250012538, dated 02 / 14 / 2025, page 44 / 61 3 / 3 8. Method for analyzing the degree of liberation of iron ore using computer vision, according to claim 1, characterized by, in step d), certain upper and lower thresholds being applied during segmentation to correct for reflection or noise effects and calculation of the percentage of iron ore in each particle.

9. Method for analyzing the degree of liberation of iron ore using computer vision, according to claim 1, characterized in that, in step e), the lower limit of the amount of iron in the quartz particles (5) is 0.1% and the upper limit of iron in the iron particles (4) is 70% of the percentage of iron ore measured in each particle in step d).

10. Method for analyzing the degree of liberation of iron ore using computer vision, according to claim 1, characterized by being applied to an ore sample collected on a conveyor belt. Petition 870250012538, dated 02 / 14 / 2025, pp. 45 / 61

Images