Different-grade iron ore hyperspectral image classification method based on differential characteristics

Abstract

The invention relates to the technical field of metal smelting, in particular to a different-grade iron ore hyperspectral image classification method based on differential characteristics. The different-grade iron ore hyperspectral image classification method based on differential characteristics comprises the following steps: S1, preprocessing: collecting iron ore hyperspectral images, and preprocessing; s2, feature extraction: obtaining a hyperspectral wave band of the iron ore, and extracting an average spectral feature and a proportion feature; and S3, classification: taking the combination of the average spectral characteristics and the proportion characteristics obtained in the step S2 as input characteristics of a random forest classifier, and performing multi-grade classification on the iron ores, so that reasonable classification of different grades of iron ores can be realized under the condition of attachment of mine dust and water. And high-precision separation of each iron ore in the iron ore production process is helped to be completed.

Description

Technical Field

[0001] The invention relates to the technical field of metal smelting, and in particular to a method for classifying hyperspectral images of iron ores of different grades based on differential features. Background Art

[0002] Iron ore is the core raw material of the steel industry. Steel products obtained from iron ore smelting have long become essential materials for applications in industrial fields such as shipbuilding, motor vehicle manufacturing, railway construction, bridge construction, heavy industry, and machinery manufacturing. Therefore, iron ore is considered to be the cornerstone of the industrial field and is very important to the global economy. In the processing of iron ore, grading iron ore in advance before grinding and other processes can significantly reduce the economic loss caused by wear and tear, energy consumption, dust pollution, and promote waste utilization. Traditional metallurgical methods are difficult to perform this task because they cannot accurately assess the iron content range of each piece of ore. At present, hyperspectral imaging technology can accurately and non-destructively estimate the grade range of each piece of iron ore due to the full use of spectral information (spectral information has been shown to be highly correlated with iron content), thereby helping to grade each piece of iron ore. Therefore, this technology has great potential to be truly applied to the iron ore sorting industry. In general, the development of this technology has a significant positive effect on a green and sustainable environment.

[0003] For hyperspectral imaging technology, an issue that cannot be ignored is that when a certain material is to be spectrally analyzed or identified, the water and dust on the surface of the material may affect the spectral response of the material. For dust, some studies have pointed out that when dust is attached to the surface of a material or rock, the dust can mask the spectral characteristics of the underlying material in visible light and infrared (Johnson et al., 2004). Two studies have found that in the wavelength range of 750nm-1000nm, the spectral reflectance of dusty leaves is significantly higher than that of dust-free leaves. Based on this characteristic, it is possible to estimate whether the leaf surface is dusted (Lin et al., 2021). Some studies have found through spectral unmixing that there is a difference between the end members of the sampling point in the mining area with obvious dust accumulation and the end members of other locations (Cleaver et al., 2022). For water, some studies have pointed out that the water content in the tailings will cause changes in the spectrum. At 1450nm and 1900nm, the tailings containing water have obvious troughs, while the tailings without water do not have such obvious characteristics (Maurais et al., 2021). Studies have found that there are also obvious spectral differences between water-containing soil and water-free soil. Water content will lead to a significant decrease in spectral reflectance, and the trough will be more obvious around 1900nm (Krzyszczak et al., 2023). In the wine industry and the wood industry, studies have shown that the spectra of pit mud and wood are affected by water content (Chen and Li, 2020; Zhu et al., 2020). In general, the influence of water and mineral dust on the sample spectrum does exist.

[0004] In view of this, there is an urgent need to provide a method for classifying hyperspectral images of iron ores of different grades based on differential features to solve the problem that the current iron ore hyperspectral image classification model is disturbed by mine dust and water. Summary of the invention

[0005] The purpose of the present invention is to solve at least one of the technical problems existing in the prior art and to provide a method for classifying hyperspectral images of iron ores of different grades based on differential features.

[0006] To achieve the above object, the technical solution adopted by the present invention is as follows: a hyperspectral image classification method for different grades of iron ore based on differential features, comprising the following steps:

[0007] S1. Preprocessing: Collecting iron ore hyperspectral images and performing preprocessing;

[0008] S2, feature extraction: obtain the hyperspectral band of iron ore and extract the average spectral features and proportion features;

[0009] S3, classification: The average spectral feature and the proportion feature obtained in step S2 are combined as input features of the random forest classifier to perform multi-level classification on the iron ore.

[0010] Furthermore, the hyperspectral bands of iron ore include 7, and the data of the 7 hyperspectral bands are 959, 1073, 1324, 1770, 2228, 2271 and 2331 nm respectively.

[0011] Further, extracting the average spectral feature includes extracting the average spectrum of the spectral image data of the hyperspectral band, and then calculating the spectral difference thereof;

[0012] The calculation of the spectral difference includes calculating the first-order difference D1[d 1,2 ,d 1,3 ,…,d 1,n-1 ]:

[0013] [d 1,2 ,d 1,3 ,…,d 1,n-1 ]=[b2,b3,…,b n ]-[b1,b2,…,b n-1 ]

[0014] Where n is the number of selected bands, [b1,b2,…,b n ] represents the reflectance of the selected bands from the 1st to the nth in ascending order; then calculate the level II difference D2[d 2,1 ,d 2,2 ,…,d 2,n-2 ]:

[0015] D2=[d 1,2 ,d 1,3 ,…,d 1,n-1 ]-[d 1,1 ,…,d 1,n-2 ]

[0016] Among them, d p,q represents the qth sub-feature in order of difference at level p;

[0017] Finally, D1 and D2 are combined as the differential spectral feature, which is then used as the average spectral feature.

[0018] Furthermore, extracting the proportional feature includes calculating the spectral difference based on the spectral image data of the acquired hyperspectral band, then combining the obtained spectral differences to construct a differential spectral feature, then classifying it using a support vector machine to obtain the pixel distribution information in the iron ore, and finally calculating the pixel ratio, and integrating the calculated pixel ratio as the proportional feature.

[0019] Furthermore, the pixel distribution information includes any one of three types of pixels: A, B and C; the A-type pixels refer to "pixels of high iron content materials", the B-type pixels refer to "pixels that are difficult to determine whether they are high iron content materials", and the C-type pixels refer to "pixels of non-high iron content materials".

[0020] Furthermore, the pixel ratio is calculated and the calculated pixel ratio is integrated as a ratio feature including

[0021] Calculate the proportion of pixels of type A, B and C in the pixels of this ore:

[0022]

[0023] N=N A +N B +N C

[0024] Where P A , P B and P C Respectively represent the proportion of A, B and C pixels in iron ore; N A , N B and N C Respectively represent the total number of A, B and C type pixels in the iron ore; N represents the total number of pixels contained in the iron ore;

[0025] Finally, P A , P B and P C Integration as a proportional feature.

[0026] Furthermore, step S3 performs multi-level classification of the iron ore, including classifying the iron ore into four levels according to the average spectral features and the proportion features using a random forest classifier: iron ore with a TFe content of less than 10% is marked as Class IV, indicating a very low TFe content; iron ore with a TFe content of 10% ≤ TFe < 48% is marked as Class III, indicating a relatively low TFe content; iron ore with a TFe content of 48% ≤ TFe < 64% is marked as Class II, indicating a relatively high TFe content; iron ore with a TFe content of ≥ 64% is marked as Class I, indicating a very high TFe content.

[0027] Compared with the prior art, the technical solution of this application has the following beneficial effects:

[0028] This application removes the bands that are severely disturbed by water and supplements the bands that are less affected by mineral dust as the band set used for prediction, constructs the secondary difference of the spectrum as the key feature of model judgment, and classifies iron ore in combination with random forest. Using this method, it is possible to reasonably classify iron ore of different grades under the conditions of mineral dust and water attachment, and help complete the high-precision sorting of each piece of iron ore in the iron ore production process. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a flow chart of a method for classifying hyperspectral images of iron ores of different grades based on differential features in a preferred embodiment of the present invention;

[0030] Figure 2 It is a diagram showing the computing structure of the present invention;

[0031] Figure 3 This is a qualitative evaluation diagram of the accuracy of the calculation results of the present invention. DETAILED DESCRIPTION

[0032] The technical solutions in the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0033] Reference Figure 1-Figure 3 As shown, a preferred embodiment of the present invention is a method for classifying hyperspectral images of iron ores of different grades based on differential features, comprising the following steps:

[0034] S1. Preprocessing: Collecting iron ore hyperspectral images and performing preprocessing;

[0035] Preprocess the hyperspectral image. This process includes removing noise from the spectral curve and masking out the background pixels in the image, which will not be described in detail here.

[0036] S2, feature extraction: obtain the hyperspectral band of iron ore and extract the average spectral features and proportion features;

[0037] The hyperspectral bands of the iron ore include 7, and the data of the 7 hyperspectral bands are 959, 1073, 1324, 1770, 2228, 2271 and 2331nm respectively; these 7 bands are found through analysis and are less affected by ore or water. In addition, most of them are at the peak and valley positions, and the information of the peak and valley is very helpful for distinguishing iron ores of different grades.

[0038] Extracting the average spectrum feature includes extracting the average spectrum of the spectrum image data of the hyperspectral band, and then calculating the spectrum difference thereof;

[0039] The calculation of the spectral difference includes calculating the first-order difference D1[d 1,2 ,d 1,3 ,…,d 1,n-1 ]:

[0040] [d 1,2 ,d 1,3 ,…,d 1,n-1 ]=[b2,b3,…,bn ]-[b1,b2,…,b n-1 ]

[0041] Where n is the number of selected bands, [b1,b2,…,b n ] represents the reflectance of the selected bands from the 1st to the nth in ascending order; then calculate the level II difference D2[d 2,1 ,d 2,2 ,…,d 2,n-2 ]:

[0042] D2=[d 1,2 ,d 1,3 ,…,d 1,n-1 ]-[d 1,1 ,…,d 1,n-2 ]

[0043] Among them, d p,q Represents the qth sub-feature arranged in order at the p-level difference; finally, D1 and D2 are combined as the differential spectral feature, and then this feature is used as the average spectral feature.

[0044] Extracting the proportion feature includes calculating the spectral difference according to the spectral image data of the acquired hyperspectral band, then combining the obtained spectral differences to construct the differential spectral feature, and then using the support vector machine to classify it, obtaining the pixel distribution information in the iron ore, and finally calculating the pixel ratio, and integrating the calculated pixel ratio as the proportion feature. The pixel distribution information includes any one of the three types of pixels A, B and C; the A-type pixels refer to "pixels of high-iron content materials", the B-type pixels refer to "pixels that are difficult to judge whether they are high-iron content materials", and the C-type pixels refer to "pixels of non-high-iron content materials".

[0045] The calculation of pixel ratio and integration of the calculated pixel ratio as a ratio feature includes calculating the ratio of the three types of pixels A, B and C to the pixels of the ore:

[0046]

[0047] N=N A +N B +N C

[0048] Where P A , P B and P C Respectively represent the proportion of A, B and C pixels in iron ore; N A , N B and N CRespectively represent the total number of A, B and C pixels in the iron ore; N represents the total number of pixels contained in the iron ore; finally, P A , P B and P C Integration as a proportional feature.

[0049] S3, classification: The average spectral feature and the proportion feature obtained in step S2 are combined as input features of the random forest classifier to perform multi-level classification on the iron ore.

[0050] The multi-level classification of iron ore includes dividing the iron ore into four levels using a random forest classifier based on average spectral features and proportion features: iron ore with a TFe content of less than 10% is marked as Class IV, indicating a very low TFe content; iron ore with a TFe content of 10% ≤ TFe < 48% is marked as Class III, indicating a lower TFe content; iron ore with a TFe content of 48% ≤ TFe < 64% is marked as Class II, indicating a higher TFe content; iron ore with a TFe content of ≥ 64% is marked as Class I, indicating a very high TFe content.

[0051] The present invention will be further described below in conjunction with specific embodiments.

[0052] Data Introduction:

[0053] (1) Experimental materials: In order to ensure that both high-grade and low-grade ores exist in the experimental samples, low-grade and high-grade raw ores (particle size: 200-500mm) were purchased from an iron ore mining area in Shandong (the mining area mainly produces magnetite), amounting to hundreds of kilograms. First, the low-grade and high-grade raw ores were crushed using a jaw crusher (model: PE-1500*200) as shown in Figure (b). Secondly, 150 samples that met the particle size standards were manually selected from the crushed low-grade ore and high-grade ore using sieves with diameters of 20mm and 40mm, respectively. This process produced a total of 300 samples, and finally the samples were numbered to facilitate the subsequent crushing and testing of each ore.

[0054] (2) Hyperspectral data acquisition: As shown in the figure, after the iron ore samples are prepared, the hyperspectral images of these samples are first obtained without washing, which are recorded as dust-free and water-free data. Then, the mineral dust attached to the surface of the samples is washed off with clean water to ensure that the samples are in a wet state, and the hyperspectral images of these samples are immediately obtained, which are recorded as dust-free and water-free data. Finally, the wet samples are dried under normal conditions and their hyperspectral images are obtained, which are recorded as dust-free and water-free data. In short, a total of hyperspectral image data of the same batch of ore under three conditions (dust-free and water-free, dust-free and water-free, and dust-free and water-free) are obtained. Note that generally speaking, mineral dust and water are not easy to exist at the same time (because water washing can wash away mineral dust), so this experiment does not consider the situation of dust and water. Note that each piece of ore was imaged once in three different postures.

[0055] The present invention proposes to quickly and accurately predict and classify iron ore based on differential features and hyperspectral images, as shown in the technical flow chart, which is mainly divided into two steps:

[0056] First, the hyperspectral image should be preprocessed, which includes noise removal, masking and other operations. Then the classification of iron ore is divided into two steps:

[0057] (1) Feature extraction

[0058] a. Band extraction

[0059] First, extract the data of 7 hyperspectral bands: 959, 1073, 1324, 1770, 2228, 2271 and 2331nm. These 7 bands were found through analysis and are less affected by ore or water. In addition, most of them are located at the peak and valley positions, and the information of peak and valley is very helpful for distinguishing iron ore of different grades.

[0060] b. Average spectral feature extraction

[0061] First, the average spectrum of the 7-band spectral image data is extracted, and then the first-order difference D1[d 1,2 ,d 1,3 ,…,d 1,n-1 ]:

[0062] [d 1,2 ,d 1,3 ,…,d 1,n-1 ]=[b2,b3,…,b n ]-[b1,b2,…,b n-1 ]

[0063] Where n is the number of selected bands (here 7), [b1,b2,…,b n] represents the reflectance of the selected bands from the 1st to the nth in ascending order. Then calculate the second-level difference D2[d 2,1 ,d 2,2 ,…,d 2,n-2 ]:

[0064] D2=[d 1,2 ,d 1,3 ,…,d 1,n-1 ]-[d 1,1 ,…,d 1,n-2 ]

[0065] Among them, d p,q represents the qth sub-feature in the order of p-level differences, for example, the first sub-feature of level 1 difference is denoted as d 1,1 Finally, D1 and D2 are combined as the difference spectrum feature, and then this feature is used as the average spectrum feature.

[0066] c. Proportional feature extraction

[0067] First, for each pixel-level spectrum in the iron ore hyperspectral image, extract its first-level difference and second-level difference as the average spectrum, then combine them to construct the differential spectrum feature, and then use the support vector machine to classify it into any one of the three types of pixels A, B and C. Among them, the A-type pixel refers to "the pixel of the high-iron content material", the B-type pixel refers to "the pixel that is difficult to judge whether it is a high-iron content material", and the C-type pixel refers to "the pixel of the non-high-iron content material". After that, the distribution of the three types of pixels A, B and C in this piece of iron ore is obtained, and then the proportion of the three types of pixels A, B and C in the pixels of this piece of ore is calculated:

[0068]

[0069] N=N A +N B +N C

[0070] Where P A , P B and P C Represents the proportion of A, B and C pixels in iron ore. A , N B and N C Represent the total number of A, B and C pixels in the iron ore. N represents the total number of pixels contained in the iron ore. A , P B and P C Integration as a proportional feature.

[0071] After steps b and c are completed, the average spectral features and proportion features obtained respectively are combined as input features to the next step random forest classifier.

[0072] (2) Classification

[0073] Based on these characteristics, the iron ore is classified into one of the following categories: I, II, III, and IV using a random forest classifier (which has been trained using iron ore samples). Ores with a TFe content of <10% are labeled as Class IV, indicating a very low TFe content. Ores with a TFe content of <48% but ≥10% are labeled as Class III, indicating a low TFe content. Ores with a TFe content of <64% but ≥48% are classified as Class II, indicating a high TFe content. Ores with a TFe content of ≥64% are classified as Class I, indicating a very high TFe content. The Chinese standard specifies 48% as the minimum value for industrial iron ore. Therefore, in this classification standard, both Class I and Class II are considered to be ores that can be used for industry, but Class II ores are not yet able to meet the standards for furnace entry, so some additional beneficiation processes are recommended. Class III and Class IV iron ores are not considered to be ores that can be used for industry. However, for countries with more poor ores, such as China, since most of the ores mined are less than 48% grade, there is almost no ore available if this standard is strictly followed. Therefore, Class III iron ore can also be considered for further processing to extract some of the iron. Class IV iron ore contains too little iron and can hardly bring economic value to the steel industry, so it is recommended to be used in other fields or directly discarded as waste.

[0074] This method realizes the classification of iron ore under different conditions. Figure 2 As shown in the figure, the algorithm results of the Pearl River Delta are shown as an example. Under the three conditions of "no dust and no water", "dust and no water" and "dust and water", the classification method achieved 87.7%, 85% and 85.3% classification accuracy respectively. The accuracy difference between the three is not large, indicating that the model has good universality and will not be significantly affected by water or mineral dust. In order to better illustrate the classification results, the classification effect is visualized as follows Figure 3 As shown in the figure, it can be seen that when iron ores of different grades are placed in an interlaced manner, the classification model can classify these ores relatively accurately, and the number of misclassifications under the three conditions is not much different. It should be noted that iron ore is mainly used in the steel smelting industry, and the misidentification of iron ore during the classification process will not have a significant impact on the steel industry. This is mainly because the sorting of iron ore still has a blending operation at the end, and this step will ensure that all materials used for smelting meet reasonable content requirements. In summary, this classification method has great potential to be used in the actual industrial sorting of iron ore under different conditions.

[0075] Under the premise that no conflict occurs, those skilled in the art may freely combine and superimpose the above-mentioned additional technical features.

[0076] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.

Claims

1. A hyperspectral image classification method for different grades of iron ore based on differential features, characterized in that: The following steps are involved: S1. Preprocessing: Collecting iron ore hyperspectral images and performing preprocessing; S2, feature extraction: obtain the hyperspectral band of iron ore and extract the average spectral features and proportion features; S3, classification: The average spectral feature and the proportion feature obtained in step S2 are combined as input features of the random forest classifier to perform multi-level classification on the iron ore.

2. The method for classifying hyperspectral images of different grades of iron ores based on differential features according to claim 1 is characterized in that: The iron ore has seven hyperspectral bands, and the data of the seven hyperspectral bands are 959, 1073, 1324, 1770, 2228, 2271 and 2331 nm respectively.

3. The method for classifying hyperspectral images of different grades of iron ores based on differential features according to claim 1 is characterized in that: The extracting of the average spectrum feature comprises extracting the average spectrum of the spectrum image data of the hyperspectral band, and then calculating the spectrum difference thereof; The calculation of the spectral difference includes calculating the first-order difference D1[d 1,2 ,d 1,3 ,…,d 1,n-1 ]: [d 1,2 ,d 1,3 ,…,d 1,n-1 ]=[b2,b3,…,b n ]-[b1,b2,…,b n-1 ] Where n is the number of selected bands, [b1,b2,…,b n ] represents the reflectance of the selected bands from the 1st to the nth in ascending order; then calculate the level II difference D2[d 2,1 ,d 2,2 ,…,d 2,n-2 ]: D2=[d 1,2 ,d 1,3 ,…,d 1,n-1 ]-[d 1,1 ,…,d 1,n-2 ] Among them, d p,q represents the qth sub-feature in order of difference at level p; Finally, D1 and D2 are combined as the differential spectral feature, which is then used as the average spectral feature.

4. The method for classifying hyperspectral images of iron ores of different grades based on differential features according to claim 1 is characterized in that: The extraction of proportional features includes calculating the spectral difference based on the spectral image data of the acquired hyperspectral band, then combining the obtained spectral differences to construct differential spectral features, then classifying them using a support vector machine to obtain pixel distribution information in the iron ore, and finally calculating the pixel ratio, and integrating the calculated pixel ratio as the proportional feature.

5. The method for classifying hyperspectral images of different grades of iron ores based on differential features according to claim 4 is characterized in that: The pixel distribution information includes any one of three types of pixels: A, B and C; the A-type pixels refer to "pixels of high-iron content materials", the B-type pixels refer to "pixels that are difficult to determine whether they are high-iron content materials", and the C-type pixels refer to "pixels of non-high-iron content materials".

6. The method for classifying hyperspectral images of different grades of iron ores based on differential features according to claim 5 is characterized in that: The calculating pixel ratio and integrating the calculated pixel ratio as a ratio feature includes Calculate the proportion of pixels of type A, B and C in the pixels of this ore: N=N A +N B +N C Where P A , P B and P C Respectively represent the proportion of A, B and C pixels in iron ore; N A , N B and N C Respectively represent the total number of A, B and C type pixels in the iron ore; N represents the total number of pixels contained in the iron ore; Finally, P A , P B and P C Integration as a proportional feature.

7. The method for classifying hyperspectral images of different grades of iron ores based on differential features according to claim 1 is characterized in that: The step S3 performs multi-level classification of the iron ore, including classifying the iron ore into four levels according to the average spectral features and the proportion features using a random forest classifier: iron ore with a TFe content of less than 10% is marked as Class IV, indicating that the TFe content is very low; iron ore with a TFe content of 10%≤TFe<48% is marked as Class III, indicating a lower TFe content; iron ore with a TFe content of 48%≤TFe<64% is marked as Class II, indicating a higher TFe content; iron ore with a TFe content of ≥64% is marked as Class I, indicating that the TFe content is very high.

Images