A method for identifying a type of scrap steel material

Abstract

The application discloses a kind of scrap steel material type identification method, obtains the reflectivity of scrap steel material surface in each preset spectrum channel and is sequentially ordered according to wavelength Sequence constraint group, carries out first-order difference, normalization and ascending and descending direction symbol extraction, finally determine material type in combination with preset mapping relationship.Because first-order difference processing can eliminate the overall deviation in reflectivity due to common factors such as light intensity, average absolute deviation normalization processing has scale invariance, can eliminate external interference such as detection environment, light source intensity and shooting distance, ascending and descending direction sequence group retains the inherent ascending and descending skeleton form of spectral curve, therefore the application embodiment has good anti-interference ability to collection condition and environmental change, for appearance highly similar but spectral response exists subtle difference Easy-to-confuse material, can be accurately distinguished by ascending and descending direction sequence and relative intensity sequence Gradual matching, can cover the identification needs of 14 kinds of industrial easy-to-confuse materials and corrosion products.

Description

Technical Field

[0001] This invention relates to the fields of multispectral detection and industrial quality inspection technology, and in particular to a method for identifying scrap steel material types, which can be extended to the detection of corrosion in metallic materials. Background Technology

[0002] During the recycling process, scrap steel often contains various non-metallic inclusions such as polyurethane, plastics, rubber, wood, and cement due to its complex origins. Some of these inclusions are highly similar in appearance to the metallic matrix or other non-metallic materials, adversely affecting the control of chemical composition in subsequent smelting processes. Existing detection methods mostly rely on visual inspection or single-sensor technology, making it difficult to accurately distinguish these easily confused materials, resulting in insufficient reliability in scrap steel quality assessment.

[0003] Meanwhile, existing deep learning-based image recognition methods are mostly pure data-driven black-box models, lacking the constraints of prior knowledge in materials science. They are prone to learning non-physical noise features caused by fluctuations in the detection environment and light source, resulting in severe overfitting problems. Furthermore, they require the re-collection of large amounts of labeled data when transferring to different scenarios, leading to high implementation costs. In extended scenarios such as salt spray corrosion detection of cold-rolled galvanized steel sheets, manual visual evaluation suffers from poor consistency and low efficiency, failing to meet the needs of large-scale inspection. Therefore, improving the accuracy, anti-interference ability, and cross-scenario adaptability of easily confused material identification is an urgent problem to be solved in the current industrial quality inspection field. Summary of the Invention

[0004] This application provides a method for identifying scrap steel material types, which solves the technical problem in the prior art where the source of scrap steel is complex, the appearance of inclusions is highly similar, and it is difficult to distinguish them by a single detection method, resulting in insufficient accuracy in identifying easily confused materials. It achieves the technical effect of accurate identification of scrap steel material types based on spectral physical characteristics.

[0005] Firstly, this application provides a method for identifying scrap steel material types, the method comprising: Acquire the spectral image of the material surface of the scrap steel to be identified, and for each pixel in the spectral image of the material surface of the scrap steel to be identified, obtain the reflectance corresponding to X preset spectral channels, and sort the reflectance according to the preset order corresponding to the X preset spectral channels to form a sequence constraint group; Perform first-order difference processing on the sequence constraint group to obtain a difference sequence group; The differential sequence group is normalized to obtain the relative intensity sequence group; Based on the differential sequence group or the relative intensity sequence group, the rising and falling direction sequence group is obtained; Based on the current detection conditions, the material type corresponding to the pixel is determined based on the rising and falling direction sequence group, the relative intensity sequence group, and the preset mapping relationship. The preset mapping relationship includes: the sequential identification order between different material types, and the correspondence between the rising and falling direction sequence features and the relative intensity sequence features and the material type.

[0006] Secondly, this application provides a method for identifying scrap steel material types, characterized in that the method includes: In the preset scenario, a scrap steel material type identification method as provided in the first aspect is used to determine the material type corresponding to the pixel of the scrap steel sample, so as to label the pixel of the scrap steel sample and construct a scrap steel sample dataset; The material type corresponding to the sample pixels of the target scene to be adapted is collected to annotate the sample pixels of the target scene and construct the target scene sample dataset; the preset scene and the target scene are different scenes, and the preset scene and the target scene are any one of the following scenarios: scrap steel inclusion recognition scene, steel defect detection scene, and galvanized cold-rolled steel plate corrosion detection scene; the number of samples in the target scene sample dataset is less than the number of samples in the scrap steel sample dataset. The preset mapping relationship is adjusted based on the distribution statistical characteristics of the target scene sample dataset to obtain the target mapping relationship; Acquire the spectral image of the material surface of the scrap steel to be identified in the target scene, and combine it with the target mapping relationship to determine the material type of each pixel in the spectral image of the material surface of the scrap steel to be identified in the target scene.

[0007] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: This embodiment first obtains the reflectance of the scrap steel surface in each preset spectral channel and sorts them according to wavelength to form a sequence constraint group. By performing first-order difference processing, normalization processing, and extraction of rising and falling direction symbols on the sequence constraint group, the original reflectance values ​​are transformed into physically meaningful dynamic trend features, scale-invariant features, and topological features. Finally, the material type is determined by combining the preset mapping relationship. Since the first-order difference processing can eliminate the overall shift in reflectance caused by common factors such as light intensity, and the mean absolute deviation normalization processing has scale invariance, it can eliminate external interference such as detection environment, light source intensity, and shooting distance. The rising and falling direction sequence group retains the inherent rising and falling skeleton shape of the spectral curve. Therefore, this embodiment has good anti-interference ability against changes in acquisition conditions and environment, and the recognition results have physical interpretability. It avoids the overfitting problem caused by learning non-physical noise features in traditional black box models. At the same time, for easily confused materials with highly similar appearances but subtle differences in spectral response, it can achieve accurate differentiation through step-by-step matching of rising and falling direction sequences and relative intensity sequences, which can cover the identification needs of 14 types of easily confused industrial materials and corrosion products. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 A flowchart illustrating a method for identifying scrap steel material types provided in this application embodiment; Figure 2 A schematic diagram illustrating the principles of first-order difference, normalization processing, and ascending / descending direction sequence provided in the embodiments of this application; Figure 3 The embodiments provided in this application are related to Figure 1 A schematic diagram of another process for a method of identifying scrap steel material types; Figure 4 This is a flowchart illustrating another method for identifying scrap steel material types provided in an embodiment of this application. Detailed Implementation

[0010] This application provides a method for identifying scrap steel material types, which solves the technical problem in the prior art where the source of scrap steel is complex, the appearance of inclusions is highly similar, and it is difficult to distinguish them by a single detection method, resulting in insufficient accuracy in identifying easily confused materials. It achieves the technical effect of accurate identification of scrap steel material types based on spectral physical characteristics.

[0011] The technical solution of this application embodiment is to solve the above-mentioned technical problems, and the general idea is as follows: This embodiment first obtains the reflectance of the scrap steel surface in each preset spectral channel and sorts them according to wavelength to form a sequence constraint group. By performing first-order difference processing, normalization processing, and extraction of rising and falling direction symbols on the sequence constraint group, the original reflectance values ​​are transformed into physically meaningful dynamic trend features, scale-invariant features, and topological features. Finally, the material type is determined by combining the preset mapping relationship. Since the first-order difference processing can eliminate the overall shift in reflectance caused by common factors such as light intensity, and the mean absolute deviation normalization processing has scale invariance, it can eliminate external interference such as detection environment, light source intensity, and shooting distance. The rising and falling direction sequence group retains the inherent rising and falling skeleton shape of the spectral curve. Therefore, this embodiment has good anti-interference ability against changes in acquisition conditions and environment, and the recognition results have physical interpretability. It avoids the overfitting problem caused by learning non-physical noise features in traditional black box models. At the same time, for easily confused materials with highly similar appearances but subtle differences in spectral response, it can achieve accurate differentiation through step-by-step matching of rising and falling direction sequences and relative intensity sequences, which can cover the identification needs of 14 types of easily confused industrial materials and corrosion products.

[0012] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0013] First, it should be clarified that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0014] Based on the same inventive concept, this application provides a method for identifying scrap steel material types, the method including steps S11-S15, which can be referred to in detail. Figure 1 As shown.

[0015] Step S11: Obtain the spectral image of the material surface of the scrap steel to be identified, and for each pixel in the spectral image of the material surface of the scrap steel to be identified, obtain the reflectance corresponding to X preset spectral channels, and sort the reflectance according to the preset order corresponding to the X preset spectral channels to form a sequence constraint group. Step S12: Perform first-order difference processing on the sequence constraint group to obtain a difference sequence group; Step S13: Normalize the differential sequence group to obtain the relative intensity sequence group; Step S14: Based on the differential sequence group or the relative intensity sequence group, obtain the rising and falling direction sequence group; Step S15: Based on the current detection conditions, and based on the lifting direction sequence group, the relative intensity sequence group, and the preset mapping relationship, determine the material type corresponding to the pixel. The preset mapping relationship includes: the sequential identification order between different material types, and the correspondence between the lifting direction sequence features and the relative intensity sequence features and the material type.

[0016] The scrap steel material type identification method provided in this application can be applied to electronic devices, including but not limited to servers, terminal devices, or industrial quality inspection equipment. The server can be an independent physical server, a server cluster consisting of multiple servers, or a cloud server supporting cloud computing services. The terminal device can be an industrial camera, a spectral imager, an embedded edge computing device, or a computer device with data processing capabilities. This application does not limit the specific form of the electronic device. In actual industrial applications, this method can be executed collaboratively by a multispectral acquisition terminal deployed on a scrap steel recycling production line or quality inspection station and a back-end processor connected to it, or it can be executed independently by an integrated detection device that integrates a spectral acquisition module and a data processing module. The electronic device is equipped with a processor and a memory. The memory stores computer programs. When the processor runs the computer program, it executes the steps described in this application, thereby determining the material type of each pixel in the spectral image of the scrap steel surface to be identified. Furthermore, the electronic device also includes or is connected to a multispectral image acquisition device for acquiring the spectral image of the scrap steel surface to be identified. Those skilled in the art will understand that the execution device provided in the embodiments of this application is not limited to the examples above, and any computing device capable of data processing and spectral signal analysis can be used as the execution subject of the embodiments of this application.

[0017] Regarding step S11, obtain the spectral image of the material surface of the scrap steel to be identified, and for each pixel in the spectral image of the material surface of the scrap steel to be identified, obtain the reflectance corresponding to X preset spectral channels, and sort the reflectance according to the preset order corresponding to the X preset spectral channels to form a sequence constraint group.

[0018] Based on the specific implementation of step S11 in this application embodiment, acquiring the spectral image of the scrap steel surface to be identified essentially involves using a multispectral imaging device to image the scrap steel surface. This imaging device is equipped with narrowband filters or spectroscopic elements corresponding to X preset spectral channels, thereby simultaneously acquiring the reflectance response of the material surface under each preset spectral channel within the same spatial field of view. For each pixel in the spectral image of the material surface, its pixel grayscale value or the value after radiometric calibration under different spectral channels represents the relative reflectance of that spatial location at each preset wavelength. Since the reflectance of different materials at different wavelengths has a definite physical response law, and this law strictly depends on the physical evolution order of wavelength from short to long, sorting the above reflectances according to a fixed order of wavelength from short to long can constitute the sequence constraint group.

[0019] The wavelengths of the X preset spectral channels are selected based on the spectral differences between typical interfering materials and scrap steel substrates in the visible to near-infrared band. Following a preset order, the wavelengths of the X preset spectral channels are: 420nm, 530nm, 650nm, 780nm, 850nm, and 940nm. In some embodiments of this application, the multispectral imaging device can be a multispectral camera with six independent spectral channels, or a broadband camera used in conjunction with a continuously adjustable filter wheel, as long as reflectance data at the aforementioned six preset wavelengths can be obtained. For each pixel, each element in the sequence constraint group corresponds one-to-one with each preset spectral channel, thereby ensuring that subsequent processing steps can analyze the spectral response characteristics of each pixel location under a unified wavelength reference.

[0020] In a specific example of this application, for a certain pixel location on the surface of the scrap steel material to be identified, the reflectance of the six preset spectral channels acquired by the multispectral imaging device at that pixel is as follows: 420nm channel reflectance value is 264, 530nm channel reflectance value is 1167, 650nm channel reflectance value is 900, 780nm channel reflectance value is 610, 850nm channel reflectance value is 470, and 940nm channel reflectance value is 112. Arranged in a fixed order from shortest to longest wavelength, the sequence constraint group corresponding to this pixel is [264, 1167, 900, 610, 470, 112]. It should be noted that the above reflectance values ​​are merely exemplary data listed for the purpose of understanding the embodiments of this application. The reflectance values ​​in actual applications depend on factors such as the surface condition of the specific material, the radiation calibration coefficient of the multispectral imaging device, and lighting conditions, which are not limited in this application.

[0021] Regarding step S12, the sequence constraint group is subjected to first-order difference processing to obtain a difference sequence group.

[0022] Based on the specific implementation of step S12 in the embodiments of this application, the first-order difference processing is performed on the sequence constraint group, which is essentially calculating the difference between the reflectance of two adjacent spectral channels in the sequence, thereby transforming the original static reflectance numerical sequence into a dynamic trend feature that reflects the change of spectrum with wavelength.

[0023] Specifically, if the sequence constraint group arranged in wavelength order in step S11 is denoted as data=[x1,x2,x3,x4,x5,x6], where x1-x6 correspond to the reflectance values ​​of the six spectral channels at 420nm, 530nm, 650nm, 780nm, 850nm, and 940nm respectively, then the first-order difference processing involves calculating the difference between the next channel and the previous channel sequentially, i.e., Δ1=x2-x1, Δ2=x3-x2, Δ3=x4-x3, Δ4=x5-x4, Δ5=x6-x5, thus obtaining the difference sequence group Δ=[Δ1,Δ2,Δ3,Δ4,Δ5]. The physical significance of this processing is that different materials have different rates and directions of reflectance change in adjacent wavelength ranges, and this difference can be explicitly extracted through the first-order difference. From a mathematical perspective, the first-order difference corresponds to the discrete approximation of the derivative in calculus, reflecting the trend and rate of change of the spectral curve in each wavelength range. This provides a dynamic characteristic basis for distinguishing different materials in the embodiments of this application. It should be noted that the first-order difference processing can effectively eliminate the overall shift in the absolute value of reflectivity caused by common factors such as light intensity, because the common DC component is naturally canceled out during the difference between adjacent components, making the extracted difference sequence group more reflective of the spectral change properties of the material itself.

[0024] In a specific example of this application, for the sequence constraint group data=[264,1167,900,610,470,112] exemplified in step S11, the differences between adjacent channels are calculated sequentially: Δ1=1167-264=903, Δ2=900-1167=-267, Δ3=610-900=-290, Δ4=470-610=-140, Δ5=112-470=-358, thus obtaining the difference sequence group Δ=[903,-267,-290,-140,-358]. Here, Δ1 is a positive value, indicating that the reflectivity increases from the wavelength range of 420nm to 530nm; Δ2 to Δ5 are all negative values, indicating that the reflectivity continuously decreases from each adjacent wavelength range of 530nm to 940nm. This differential sequence set visually depicts the fluctuation pattern of the spectral curve at the pixel location between six wavelength points, namely, a dynamic evolution pattern of a sharp rise followed by a continuous decline. This pattern is directly related to the identification of specific material types in subsequent steps.

[0025] Regarding step S13, the differential sequence group is normalized to obtain the relative intensity sequence group.

[0026] Based on the specific implementation of step S13 in the embodiments of this application, the differential sequence group is normalized to eliminate the scale interference introduced by the difference in absolute value of spectral reflectance between pixels, so that subsequent processing can focus on the relative fluctuation pattern of the spectral curve.

[0027] Specifically, this embodiment employs the Average Absolute Deviation (AAD) normalization method. First, the average absolute value of each element in the difference sequence group is calculated and used as a scale normalization factor. Then, each element in the difference sequence group is divided by this normalization factor to obtain the relative intensity sequence group. The normalization factor is calculated as M = (|Δ1| + |Δ2| + |Δ3| + |Δ4| + |Δ5|) / 5, where Δ1-Δ5 are the elements in the difference sequence group. The relative intensity sequence group is calculated as s_i = Δ_i / M, where s_i represents the i-th element in the relative intensity sequence group. Through this normalization process, the original difference sequence group is converted into a dimensionless relative intensity sequence group, reflecting the strength of the spectral variation amplitude in each adjacent wavelength interval relative to the overall average variation amplitude, rather than the absolute value of the variation amplitude.

[0028] The above-mentioned mean absolute deviation normalization process exhibits scale invariance. When the spectral reflectance value changes linearly due to external factors such as the detection environment, light source intensity, or shooting distance, i.e., the original reflectance value x_i is transformed into x'_i = k·x_i + b (where k is the scaling factor and b is the translation factor), each element in the difference sequence group will correspondingly become Δ'_i = k·Δ_i. The translation factor b is naturally canceled out in the difference calculation, and the normalization factor M' = k·M. Therefore, the normalized relative intensity s'_i = Δ'_i / M' = (k·Δ_i) / (k·M) = s_i remains unchanged. This mathematical property ensures that the relative intensity sequence group extracted in the embodiments of this application only reflects the inherent spectral fluctuation characteristics of the material itself and is independent of external acquisition conditions, thereby effectively improving the method's anti-interference ability and cross-scene adaptability.

[0029] In a specific example of this application, following the difference sequence group Δ=[903,-267,-290,-140,-358] obtained in step S12, the sum of the absolute values ​​of each element is first calculated as |903|+|-267|+|-290|+|-140|+|-358|=903+267+290+140+358=1958. Dividing this by 5 yields the normalization factor M. =1958 / 5=391.6; then, each element in the difference sequence group is divided by this normalization factor to obtain the relative intensity sequence group s=[903 / 391.6,-267 / 391.6,-290 / 391.6,-140 / 391.6,-358 / 391.6]=[2.306,-0.682,-0.741,-0.357,-0.914]. In this relative intensity sequence group, s1=2.306 indicates that the increase in the wavelength range of 420nm to 530nm is about 2.3 times the average change, that is, this segment shows an extremely strong upward trend; while s2=-0.682 indicates that the decrease in the wavelength range of 530nm to 650nm is about 0.68 times the average change. The sign of each element represents the direction of spectral change in the corresponding wavelength range, while the numerical value represents the significance of the change relative to the overall fluctuation level. Together, they constitute the core quantitative basis for subsequent material type identification.

[0030] Regarding step S14, a rising / falling direction sequence group is obtained based on the differential sequence group or the relative intensity sequence group.

[0031] Based on the specific implementation of step S14 in this embodiment, the ascending / descending direction sequence group is generated according to the positive or negative sign of each element in the relative intensity sequence group. Specifically, for each element si in the relative intensity sequence group s=[s1,s2,s3,s4,s5], if si is greater than 0, the corresponding ascending / descending direction di is denoted as "ascending," indicating that the spectral reflectance in this wavelength range increases with increasing wavelength; if si is less than 0, the corresponding ascending / descending direction di is denoted as "descending," indicating that the spectral reflectance in this wavelength range decreases with increasing wavelength. Thus, the ascending / descending direction sequence group dir=[d1,d2,d3,d4,d5] is obtained. Since the relative intensity sequence group is obtained by dividing each element of the difference sequence group by the same positive normalization factor M, the signs of each element in the difference sequence group are completely consistent with the signs of the corresponding elements in the relative intensity sequence group. In this embodiment, the ascending / descending direction sequence group can also be directly generated based on the positive or negative signs of the difference sequence group. The two methods are equivalent in result.

[0032] The physical significance of this step lies in transforming the continuous numerical form of the relative intensity sequence into a discrete symbolic form of the ascending and descending direction sequence, thereby extracting the morphological topological features of the spectral curve's "ascending" or "descending" within each wavelength range. The ascending and descending direction sequence reflects the overall undulating skeleton of the spectral curve, that is, the directional combination of reflectance changes between different wavelength ranges. This directional combination is determined by the intrinsic physical mechanisms of the material itself, such as electronic transitions and molecular vibrations, and is insensitive to linear scale changes such as external light intensity, exhibiting topological invariance. Compared to retaining continuous numerical values, the discretized symbolic representation compresses the data volume while preserving the most essential ascending and descending skeleton information of the spectral morphology, providing a concise and stable feature input for subsequent rapid matching based on preset mapping relationships.

[0033] In a specific example of this application, following the relative intensity sequence group s=[2.306,-0.682,-0.741,-0.357,-0.914] obtained in step S13, the sign of each element is determined: s1=2.306 is greater than 0, and the corresponding direction is recorded as "rising"; s2=-0.682 is less than 0, and is recorded as "falling"; s3=-0.741 is less than 0, and is recorded as "falling"; s4=-0.357 is less than 0, and is recorded as "falling"; s5=-0.914 is less than 0, and is recorded as "falling". The resulting rising and falling direction sequence group is dir=[rising, falling, falling, falling, falling], which indicates that the spectral curve of this pixel position shows an upward trend in the range of 420nm to 530nm, while it shows a continuous downward trend in the subsequent four adjacent wavelength ranges from 530nm to 940nm. The sequence of rising and falling directions constitutes the morphological skeleton of the pixel's spectral curve, namely, the evolutionary pattern of "rising first and then continuously falling," which can be directly used in subsequent steps to match with the preset mapping relationship to determine the material type. Furthermore, if the differential sequence group Δ=[903,-267,-290,-140,-358] is used to directly determine the sign, with 903 being positive and the rest negative, the result is dir=[rising, falling, falling, falling, falling], which is completely consistent with the rising and falling direction sequence group obtained based on the relative intensity sequence group, verifying the equivalence of the two methods.

[0034] like Figure 2 As shown, the attached figure consists of three sub-figures: left (a), middle (b), and right (c). It fully corresponds to the entire process of operation logic and mathematical invariance verification of spectral feature extraction for scrap steel identification. The overall visualization is based on the example reflectance data [264, 1167, 900, 610, 470, 112] of 6 preset spectral channels (420nm(A), 530nm(B), 650nm(C), 780nm(D), 850nm(E), 940nm(F)).

[0035] Subfigure (a) is titled "Discrete Approximation of the First-Order Difference Derivative." The horizontal axis represents the spectral channel indices A to F, arranged from shortest to longest wavelength. The vertical axis represents the reflectance values. The blue dotted line represents the original spectral signal, with the corresponding reflectance values ​​for each channel labeled: x_A=264, x_B=1167, x_C=900, x_D=610, x_E=470, x_F=112. The first-order difference calculation results are also labeled in each adjacent channel interval. For segment AB, Δ_1 = +903, BC segment Δ_2=-267, CD segment Δ_3=-290, DE segment Δ_4=-140, EF segment Δ_5=-358, the difference operation formula Δi=x{i+1}-xi is marked below, intuitively showing the principle of first-order difference transforming static reflectance sequence into dynamic trend characteristics of spectral change. The curve shape clearly shows the change pattern of reflectance increasing significantly from 420nm to 530nm and continuously decreasing from 530nm to 940nm.

[0036] Subfigure (b) is titled "Proof of AAD Standardized Scale Invariance." The horizontal axis represents the five difference segments AB, BC, CD, DE, and EF. The left vertical axis represents the difference values, and the right vertical axis represents the standardized relative intensity s. In the figure, the blue bars represent the original difference Δ, the red bars represent the scaled difference Δ' after scaling factor k=2, and the green triangular broken line represents the standardized relative intensity s obtained by normalizing the AAD mean absolute deviation. The scale transformation derivation formula x'_i=k is marked below. x_i+b, scaling difference Δ'_i=k Δ_i, normalization factor M'=k M, ultimately leading to the standardized value s'_i=(k Δ_i) / (k M)=s_i, proving that the standardization result is not affected by linear scaling and translation of reflectance and has scale invariance. In the figure, the original difference and the scaled difference of each segment are only related by multiple of each other in terms of magnitude. The shape of the normalized relative intensity broken line is completely consistent, which intuitively verifies this mathematical property.

[0037] Subfigure (c) is titled "Symbol Sequence Encoding of Topological Invariants." The horizontal axis represents spectral channels A to F, and the vertical axis represents reflectance values. The blue curve represents the original spectrum, and the red curve represents the spectrum after scaling and translation transformation. The original curve has an upward inflection point at channel B, and channels C, D, E, and F have downward inflection points in sequence, marked with up and down arrows to indicate the upward and downward trends. After linear transformation, the overall amplitude of the red curve increases significantly, but the order of upward and downward changes corresponding to the curve's inflection points is completely consistent with the original spectrum. The topological feature symbol sequence [↑,↓,↓,↓,↓] is marked below the figure, indicating that regardless of the scaling and overall translation of the spectral reflectance values, the upward and downward symbol sequence representing the skeleton of the curve's fluctuations remains unchanged. This corresponds to the engineering simplification of essential invariants in topology and reflects the stable characteristics of symbol sequence encoding against external linear interference such as illumination and shooting distance.

[0038] The three sub-figures correspond to the three core processing steps in the text: S12 first-order difference, S13 AAD normalization, and S14 ascending and descending direction symbol encoding. They visualize the three core logics of discrete derivative feature extraction, scale invariance mathematical verification, and topological symbol invariance feature retention, respectively. They fully and concretely illustrate the entire feature extraction process from the original multispectral reflectance to the ascending and descending direction sequence group in the text and the core advantages of each step.

[0039] Regarding step S15, based on the current detection conditions, the material type corresponding to the pixel is determined based on the rising and falling direction sequence group, the relative intensity sequence group, and the preset mapping relationship. The preset mapping relationship includes: the sequential identification order between different material types, and the correspondence between the rising and falling direction sequence features and the relative intensity sequence features and the material type.

[0040] Current testing conditions include scrap steel quality inspection mode and cold-rolled galvanized sheet salt spray corrosion testing mode.

[0041] If it is the scrap steel quality inspection mode, then steps S11-S15 can be referred to Figure 3 The process is shown below.

[0042] In the scrap steel quality inspection mode, in this embodiment of the application, step S15 determines the material type of the pixel based on the lifting direction sequence group and the relative strength sequence group, combined with a preset mapping relationship. The preset mapping relationship includes a first mapping relationship, a second mapping relationship, and a third mapping relationship; the first mapping relationship, the second mapping relationship, and the third mapping relationship correspond to different lifting direction sequence features.

[0043] Based on the rising and falling direction sequence group, the relative intensity sequence group, and the preset mapping relationship, the material type corresponding to the pixel is determined, including: The target mapping relationship is determined from the first mapping relationship, the second mapping relationship, and the third mapping relationship based on the ascending and descending direction sequence group; If the target mapping relationship is the first mapping relationship or the third mapping relationship, then the material type corresponding to the pixel is determined based on the relative intensity sequence group and the target mapping relationship; If the target mapping relationship is the second mapping relationship, and the rising and falling direction sequence feature corresponding to the second mapping relationship is rising, falling, falling, falling, rising, then the material type corresponding to the pixel is determined to be a silver-white plastic plate.

[0044] In the specific determination, the mapping relationship to which the current pixel belongs is first determined based on the ascending / descending direction sequence group as the target mapping relationship: if the ascending / descending direction sequence group is "ascending, descending, descending, descending", it corresponds to the first mapping relationship; if it is "ascending, descending, descending, descending, ascending", it corresponds to the second mapping relationship; if it is "ascending, ascending, descending, descending, descending", it corresponds to the third mapping relationship. When the target mapping relationship is the second mapping relationship, the material type of the pixel is directly determined to be a silver-white plastic plate.

[0045] When the target mapping relationship is the first mapping relationship, the values ​​of each element in the relative intensity sequence group need to be matched step by step according to the preset identification order. The order is as follows: smooth gray polyurethane, non-smooth gray polyurethane, white plastic board, white polyurethane, silver-painted wood board, cold-rolled galvanized steel board, silver-painted blue plastic board, and the step-by-step judgment follows the rule of termination upon hit.

[0046] In other words, based on the relative intensity sequence group and the target mapping relationship, determining the material type corresponding to the pixel includes: The smooth gray polyurethane corresponds to the first relative intensity sequence value in the relative intensity sequence group. If the first relative intensity sequence value is greater than or equal to 2.4 and less than or equal to 2.42, then the material type corresponding to the pixel is determined to be smooth gray polyurethane. The non-flat gray polyurethane corresponds to the first relative intensity sequence value in the relative intensity sequence group. If the first relative intensity sequence value is greater than or equal to 2.5 and less than or equal to 2.52, then the material type corresponding to the pixel is determined to be non-flat gray polyurethane. The white plastic board corresponds to the second relative intensity sequence value in the relative intensity sequence group. If the second relative intensity sequence value is greater than or equal to -0.032 and less than or equal to 0.003, then the material type corresponding to the pixel is determined to be white plastic board. The white polyurethane corresponds to the fifth relative intensity sequence value in the relative intensity sequence group. If the fifth relative intensity sequence value is less than or equal to -1.7, then the material type corresponding to the pixel is determined to be white polyurethane. The silver-painted wood panel corresponds to the third relative intensity sequence value in the relative intensity sequence group. If the third relative intensity sequence value is greater than or equal to -0.6, then the material type corresponding to this pixel is determined to be silver-painted wood panel. The cold-rolled galvanized steel sheet corresponds to the second relative strength sequence value in the relative strength sequence group. If the second relative strength sequence value is less than -0.6, then the material type corresponding to the pixel is determined to be cold-rolled galvanized steel sheet. The silver-painted blue plastic board corresponds to the second and third relative intensity sequence values ​​in the relative intensity sequence group. If the second relative intensity sequence value is greater than or equal to -0.5 and less than or equal to -0.2, and the third relative intensity sequence value is less than or equal to -0.6, then the material type corresponding to this pixel is determined to be a silver-painted blue plastic board.

[0047] Specifically, if the first relative intensity sequence value is between 2.4 and 2.42, it is determined to be smooth gray polyurethane; if it is between 2.5 and 2.52, it is determined to be non-smooth gray polyurethane. If neither of the above applies, and the second relative intensity sequence value is between -0.032 and 0.003, it is determined to be white plastic board. If neither of the above applies, and the fifth relative intensity sequence value is less than or equal to -1.7, it is determined to be white polyurethane. If neither of the above applies, and the third relative intensity sequence value is greater than or equal to -0.6, it is determined to be silver-painted wood board. If neither of the above applies, and the second relative intensity sequence value is less than -0.6, it is determined to be cold-rolled galvanized steel sheet. If neither of the above applies, and the second relative intensity sequence value is between -0.5 and -0.2, and the third relative intensity sequence value is less than or equal to -0.6, it is determined to be silver-painted blue plastic board.

[0048] The ascending and descending direction sequence feature corresponding to the third mapping relationship is ascending, ascending, descending, descending, descending; the sequential identification order between different material types corresponding to the third mapping relationship is as follows: black rubber, natural solid wood, yellow rubber, concrete cement block, and the step-by-step judgment follows the rule of termination upon hit.

[0049] Based on the relative intensity sequence group and the target mapping relationship, the material type corresponding to the pixel is determined, including: The black rubber corresponds to the first relative intensity sequence value in the relative intensity sequence group. If the first relative intensity sequence value is greater than or equal to 2.41 and less than or equal to 2.5, then the material type corresponding to the pixel is determined to be black rubber. Natural solid wood corresponds to the second relative intensity sequence value in the relative intensity sequence group. If the second relative intensity sequence value is greater than or equal to 1.5 and less than or equal to 2.5, then the material type corresponding to the pixel is determined to be natural solid wood. The yellow rubber corresponds to the second relative intensity sequence value in the relative intensity sequence group. If the second relative intensity sequence value is greater than or equal to 1.111 and less than or equal to 1.212, then the material type corresponding to this pixel is determined to be yellow rubber. The concrete cement block corresponds to the second relative strength sequence value in the relative strength sequence group. If the second relative strength sequence value is greater than or equal to 0.2 and less than or equal to 0.7, then the material type corresponding to the pixel is determined to be a concrete cement block.

[0050] In other words, if the first relative strength sequence value is between 2.41 and 2.5, it is judged as black rubber; if neither of the above applies, and the second relative strength sequence value is between 1.5 and 2.5, it is judged as natural solid wood; if neither of the above applies, and the second relative strength sequence value is between 1.111 and 1.212, it is judged as yellow rubber; if neither of the above applies, and the second relative strength sequence value is between 0.2 and 0.7, it is judged as concrete block.

[0051] In the above-mentioned step-by-step matching process, if all recognition sequences under the current target mapping relationship fail to match, a fallback judgment logic is entered: silver-sprayed black plastic is the fallback judgment category, with a priority lower than all standard materials; only when none of the aforementioned material types match under the current target mapping relationship, if the second relative intensity sequence value is greater than or equal to -0.77 and less than or equal to 0.43, is the material type corresponding to the pixel determined to be silver-sprayed black plastic. That is to say, when the material type corresponding to the pixel is determined not to be any of the following types: smooth gray polyurethane, non-smooth gray polyurethane, white plastic board, white polyurethane, silver-sprayed wood board, cold-rolled galvanized steel plate, silver-sprayed blue plastic board, black rubber, natural solid wood, yellow rubber, or concrete block, if the second relative intensity sequence value in the relative intensity sequence group is greater than or equal to -0.77 and less than or equal to 0.43, the material type corresponding to the pixel is determined to be silver-sprayed black plastic.

[0052] In the scrap steel quality inspection mode, the object to be identified is the material surface exposed after the scrap steel block is broken, which contains a variety of non-metallic inclusions such as polyurethane, plastic, rubber, wood, and cement. In this embodiment, the accurate determination of the material type of each pixel can be completed through the above steps S11 to S15.

[0053] In the salt spray corrosion detection mode for cold-rolled galvanized steel sheets, the objects to be identified become various corrosion products generated on the surface of the steel sheet after the salt spray test, such as a light zinc oxide corrosion layer and a heavy mixed corrosion product of zinc chloride and basic zinc carbonate. Therefore, the corrosion product type determination logic described below is activated in this mode. In other words, the scrap steel quality inspection mode is one working mode of this application embodiment, while the cold-rolled galvanized steel sheet salt spray corrosion detection mode is another working mode. The two modes can be selectively activated by the working condition selection switch. Both share the feature extraction process described in steps S11 to S14. After step S14, different determination branches are activated according to the corresponding modes. When the corrosion detection branch is activated, the material determination of the corresponding threshold range in the scrap steel quality inspection mode is automatically masked to avoid confusion between two different types of objects corresponding to the same threshold range.

[0054] After obtaining the rising and falling direction sequence group, in the salt spray corrosion detection mode of cold-rolled galvanized steel sheet, the preset mapping relationship also includes a fourth mapping relationship. The rising and falling direction sequence feature corresponding to the fourth mapping relationship is rising, falling, falling, falling, falling. The corresponding identification order between different material types is as follows: light zinc oxide corrosion layer, heavy mixed corrosion of zinc chloride and basic zinc carbonate, non-flat gray polyurethane, white plastic sheet, white polyurethane, silver-painted wood board, cold-rolled galvanized steel sheet, silver-painted blue plastic sheet. Based on the relative intensity sequence group and the fourth mapping relationship, the material type corresponding to the pixel is determined, including: The mild zinc oxide corrosion layer corresponds to the first relative intensity sequence value in the relative intensity sequence group. If the first relative intensity sequence value is greater than or equal to 2.4 and less than or equal to 2.42, then the material type corresponding to the pixel is determined to be a mild zinc oxide corrosion layer. The severe mixed corrosion of zinc chloride and zinc carbonate corresponds to the second relative intensity sequence value in the relative intensity sequence group. If the second relative intensity sequence value is greater than or equal to -0.5 and less than or equal to -0.1, then the material type corresponding to the pixel is determined to be severe mixed corrosion of zinc chloride and zinc carbonate. The non-flat gray polyurethane corresponds to the first relative intensity sequence value in the relative intensity sequence group. If the first relative intensity sequence value is greater than or equal to 2.5 and less than or equal to 2.52, then the material type corresponding to the pixel is determined to be non-flat gray polyurethane. The white plastic board corresponds to the second relative intensity sequence value in the relative intensity sequence group. If the second relative intensity sequence value is greater than or equal to -0.032 and less than or equal to 0.003, then the material type corresponding to the pixel is determined to be white plastic board. The white polyurethane corresponds to the fifth relative intensity sequence value in the relative intensity sequence group. If the fifth relative intensity sequence value is less than or equal to -1.7, then the material type corresponding to the pixel is determined to be white polyurethane. The silver-painted wood panel corresponds to the third relative intensity sequence value in the relative intensity sequence group. If the third relative intensity sequence value is greater than or equal to -0.6, then the material type corresponding to this pixel is determined to be silver-painted wood panel. The cold-rolled galvanized steel sheet corresponds to the second relative strength sequence value in the relative strength sequence group. If the second relative strength sequence value is less than -0.6, then the material type corresponding to the pixel is determined to be cold-rolled galvanized steel sheet. The silver-painted blue plastic board corresponds to the second and third relative intensity sequence values ​​in the relative intensity sequence group. If the second relative intensity sequence value is greater than or equal to -0.5 and less than or equal to -0.2, and the third relative intensity sequence value is less than or equal to -0.6, then the material type corresponding to this pixel is determined to be a silver-painted blue plastic board.

[0055] In this embodiment, the salt spray corrosion detection mode for cold-rolled galvanized steel sheets is a parallel operation of the aforementioned scrap steel quality inspection mode. Under the salt spray corrosion detection mode for cold-rolled galvanized steel sheets, the objects to be identified are mainly various corrosion products generated on the surface of the steel sheet after the salt spray test, as well as any possible residual coatings or inclusions. Therefore, the preset mapping relationship includes a fourth mapping relationship. The rising and falling direction sequence characteristic corresponding to the fourth mapping relationship is "rise, fall, fall, fall, fall," which is the same as the rising and falling direction sequence characteristic of the aforementioned first mapping relationship. However, the corresponding material type sequences and the order of identification are different. The order of identification between different material types corresponding to the fourth mapping relationship is as follows: light zinc oxide corrosion layer, heavy mixed corrosion of zinc chloride and basic zinc carbonate, non-smooth gray polyurethane, white plastic sheet, white polyurethane, silver-sprayed wood board, cold-rolled galvanized steel sheet, and silver-sprayed blue plastic sheet.

[0056] The specific judgment logic of the salt spray corrosion detection mode for cold-rolled galvanized steel sheets is as follows: For pixels with the ascending and descending direction sequence group of "ascend, descend, descend, descend, descend", each element in the relative intensity sequence group is matched with the threshold interval corresponding to each material type in the above-mentioned identification order. When all the threshold conditions corresponding to a certain material type are met, the pixel is determined to belong to that material type, and the judgment is terminated.

[0057] Among them, a mild zinc oxide corrosion layer corresponds to the first relative intensity sequence value; if the value is between 2.40 and 2.42, it is determined to be a mild zinc oxide corrosion layer. A severe mixed corrosion of zinc chloride and basic zinc carbonate corresponds to the second relative intensity sequence value; if the value is between -0.5 and -0.1, it is determined to be a severe mixed corrosion of zinc chloride and basic zinc carbonate. An uneven gray polyurethane corresponds to the first relative intensity sequence value; if the value is between 2.50 and 2.52, it is determined to be an uneven gray polyurethane. A white plastic board corresponds to the second relative intensity sequence value; if the value is between -0.032 and 0.003, it is determined to be a white plastic board. White polyurethane corresponds to the fifth relative intensity sequence value; if this value is less than or equal to -1.7, it is identified as white polyurethane. Silver-painted wood panels correspond to the third relative intensity sequence value; if this value is greater than or equal to -0.6, it is identified as silver-painted wood panels. Cold-rolled galvanized steel sheets correspond to the second relative intensity sequence value; if this value is less than -0.6, it is identified as cold-rolled galvanized steel sheets. Silver-painted blue plastic sheets correspond to both the second and third relative intensity sequence values; if the second relative intensity sequence value is between -0.5 and -0.2 and the third relative intensity sequence value is less than or equal to -0.6, it is identified as silver-painted blue plastic sheets. If none of the recognition sequences under the current mapping relationship are matched, no judgment result is output, or an unrecognizable flag is output, awaiting further processing.

[0058] The judgment logic in the salt spray corrosion detection mode of cold-rolled galvanized steel sheet and the scrap steel quality inspection mode has the same feature extraction process. The only difference is that different mapping relationships are selected for matching according to the working conditions, thereby realizing the reuse of the same feature extraction process in different industrial detection scenarios.

[0059] It should be noted that the relative intensity sequence value ranges corresponding to the above material types were pre-calibrated by collecting and statistically analyzing the spectra of standard samples of each material. The relative intensity values ​​of the same material type may systematically shift under different imaging devices or lighting conditions, but the relative order and distribution pattern between the intervals remain stable. Furthermore, the sequential identification order of the above step-by-step matching reflects the priority of different material types in terms of spectral morphology. The first matching categories are those with strong spectral morphology specificity and are not easily confused with other materials, while the subsequent matching categories are those further distinguished through elimination when the preceding conditions are not met. This order of judgment from specific to general helps improve the accuracy of the final identification results.

[0060] In this embodiment, the feature extraction and determination logic involved in steps S11 to S14 is designed based on the physical laws of spectral response in materials science. Specifically, the curve shape of the reflectance of different materials in the visible to near-infrared bands as a function of wavelength is determined by their intrinsic physical properties such as molecular structure, electronic transition energy levels, and lattice vibrations. This curve shape is deterministic and repeatable. Based on this, this embodiment embeds prior knowledge of materials science into the recognition process, forming a physical constraint framework at three levels: sequence constraint, intensity constraint, and morphological constraint. The sequence constraint is reflected in step S11 by arranging the reflectance of each spectral channel in a fixed order from shortest to longest wavelength, ensuring that the spectral feature extraction of all pixels is completed under a unified wavelength reference. The intensity constraint is reflected in step S13 by using the average absolute deviation normalization method, which makes the relative intensity sequence group scale-invariant and can eliminate numerical drift caused by external acquisition conditions. The morphological constraint is reflected in step S14 by extracting the ascending and descending direction sequence group to obtain the topological skeleton of the spectral curve, and in step S15 by selecting the corresponding mapping relationship according to the current detection mode to complete the material type determination.

[0061] Within the aforementioned physical constraint framework, the physical constraint features extracted in this application consist of three levels: the difference sequence group reflects the rate and direction of spectral change in adjacent wavelength intervals, belonging to dynamic trend features; the relative intensity sequence group eliminates scale interference based on the difference sequence group, belonging to scale-invariant features; and the ascending / descending direction sequence group converts continuous values ​​into discrete symbols, preserving the ascending / descending topological skeleton of the spectral curve, belonging to topological features. These three levels of features progressively advance, collectively constituting a physically interpretable representation of material morphology features, ensuring that the material type determination result for each pixel can be traced back to the specific combination of difference value, relative intensity value, and ascending / descending direction.

[0062] In this embodiment, the preset mapping relationship is not temporarily set, but is obtained in advance by calibrating standard samples of various known materials after spectral acquisition and feature extraction. Specifically, the calibration method is as follows: multispectral images of samples of various known material categories are acquired; for each sample, differential sequence groups, relative intensity sequence groups, and rising / falling direction sequence groups are extracted according to steps S11 to S14; the distribution range of relative intensity sequence values ​​of various materials under the corresponding rising / falling direction sequence features is statistically analyzed; the distribution range corresponding to each material is used as the quantification threshold for that material type; and the preset mapping relationship is stored in the form of a correspondence between "rising / falling direction sequence features plus relative intensity sequence value threshold range" and "material type". This preset mapping relationship includes multiple mapping relationships, where the first, second, and third mapping relationships are used for judgment under the scrap steel quality inspection mode, and the fourth mapping relationship is used for judgment under the cold-rolled galvanized sheet salt spray corrosion detection mode. Different mapping relationships can correspond to the same rising / falling direction sequence features, but the corresponding material type sequences and the order of identification are set separately according to the specific working conditions. In the preset mapping relationship, the order of identification between different material types is set according to the degree of spectral morphology specificity of each material. The more unique the spectral morphology and the less likely it is to be confused with other materials, the higher the sorting order, thus ensuring that the judgment process is efficient and accurate. In addition, the threshold range of the relative intensity sequence value can be fine-tuned in practical applications according to the imaging conditions of the target scene and the statistical distribution of the samples to achieve rapid adaptation across scenes, but the relative order and morphology judgment logic between the material types remain unchanged.

[0063] In summary, this embodiment first obtains the reflectance of the scrap steel surface in each preset spectral channel and sorts them according to wavelength to form a sequence constraint group. By performing first-order difference processing, normalization processing, and extraction of the rising and falling direction symbols on the sequence constraint group, the original reflectance values ​​are transformed into physically meaningful dynamic trend features, scale-invariant features, and topological features. Finally, the material type is determined by selecting the corresponding mapping relationship according to the current detection mode. Since the first-order difference processing can eliminate the overall shift in reflectance caused by common factors such as light intensity, and the mean absolute deviation normalization processing has scale invariance, it can eliminate external interference such as detection environment, light source intensity, and shooting distance. The rising and falling direction sequence group retains the inherent rising and falling skeleton shape of the spectral curve. Therefore, this embodiment has good anti-interference ability against changes in acquisition conditions and environment, and the recognition results have physical interpretability. It avoids the overfitting problem caused by learning non-physical noise features in traditional black box models. At the same time, for easily confused materials with highly similar appearances but subtle differences in spectral response, it can achieve accurate differentiation through step-by-step matching of rising and falling direction sequences and relative intensity sequences. Under the scrap steel quality inspection mode, it can cover the identification needs of 13 categories of easily confused industrial materials, including smooth gray polyurethane, non-smooth gray polyurethane, white plastic board, white polyurethane, silver-painted wood board, cold-rolled galvanized steel sheet, salt spray mixed corrosion products of cold-rolled galvanized sheet, silver-painted blue plastic board, silver-white plastic board, black rubber, natural solid wood, yellow rubber, concrete cement blocks, and silver-painted black plastic. Under the cold-rolled galvanized sheet salt spray corrosion detection mode, it can cover the identification needs of 8 categories of corrosion products and residues, including mild zinc oxide corrosion layer, severe mixed corrosion of zinc chloride and basic zinc carbonate, and some of the above-mentioned non-metallic inclusions.

[0064] Furthermore, by further defining the wavelengths of the six preset spectral channels as 420nm, 530nm, 650nm, 780nm, 850nm, and 940nm, this embodiment of the application can extract the most distinguishable spectral response features of various inclusions in scrap steel in the visible to near-infrared band based on the above wavelength combinations, providing a unified wavelength benchmark for material discrimination under this wavelength combination. This embodiment of the application also achieves rapid material type classification based on the rise-fall, fall, fall, fall, rise sequence by setting a first mapping relationship, a second mapping relationship, and a third mapping relationship corresponding to different rise-fall direction sequence features. Specifically, when the rise-fall direction sequence feature is "rise, fall, fall, fall, rise," it is directly identified as a silver-white plastic plate without the need for subsequent step-by-step threshold matching, effectively improving the judgment efficiency. This application embodiment performs step-by-step threshold matching of relative intensity sequence values ​​under a first mapping relationship, in the order of smooth gray polyurethane, uneven gray polyurethane, white plastic board, white polyurethane, silver-painted wood board, cold-rolled galvanized steel sheet, cold-rolled galvanized sheet salt spray mixed corrosion products, and silver-painted blue plastic board. Under a third mapping relationship, matching is performed in the order of black rubber, natural solid wood, yellow rubber, and concrete blocks. This allows various materials to be distinguished sequentially based on the specificity of their spectral morphology, with materials having more unique spectral morphologies ranking higher, thereby improving processing efficiency while ensuring accuracy. This application embodiment also provides a fallback judgment condition of silver-painted black plastic, offering a compatible judgment path for situations where no of the aforementioned material types are matched under both the first and third mapping relationships, further expanding the coverage of the method. In addition, this application embodiment sets a fourth mapping relationship for the determination in the salt spray corrosion detection mode of cold-rolled galvanized sheet. The determination logic of mild zinc oxide corrosion layer and severe mixed corrosion of zinc chloride and basic zinc carbonate is the same as the determination logic of each material type in the scrap steel quality inspection mode. This realizes the working condition compatibility between scrap steel quality inspection scenario and corrosion detection scenario, so that the same set of feature extraction process can be adapted to different types of industrial testing needs.

[0065] Based on the same inventive concept, this application provides a method for identifying scrap steel material types, the method including steps S41-S44, which can be referred to in detail. Figure 4 As shown.

[0066] Step S41: Under the preset scenario, the scrap steel material type identification method provided above is used to determine the material type corresponding to the scrap steel sample pixels, so as to label the scrap steel sample pixels and construct a scrap steel sample dataset. Step S42: Collect the material type corresponding to the target scene sample pixels to be adapted, so as to label the target scene sample pixels and construct the target scene sample dataset; the preset scene and the target scene are different scenes, and the preset scene and the target scene are any one of the following scenarios: scrap steel inclusion recognition scene, steel defect detection scene, and galvanized cold-rolled steel plate corrosion detection scene; the number of samples in the target scene sample dataset is less than the number of samples in the scrap steel sample dataset; Step S43: Adjust the preset mapping relationship based on the distribution statistical characteristics of the target scene sample dataset to obtain the target mapping relationship. The adjustment only shifts or scales the threshold range of the relative intensity sequence value, without changing the judgment rules of the rising and falling direction sequence or the identification order of various materials. Step S44: Obtain the surface spectral image of the scrap steel to be identified in the target scene, and determine the material type of each pixel in the surface spectral image of the scrap steel to be identified in the target scene by combining the target mapping relationship.

[0067] Regarding step S41, in this embodiment, step S41 is the process of pre-constructing a scrap steel sample dataset under a preset scenario. The preset scenario refers to a source scenario with sufficient labeled data, such as a laboratory environment or a deployed scrap steel quality inspection station. Under this scenario, multiple sets of multispectral images of scrap steel samples of known material types are collected. For each pixel in each sample image, the material type corresponding to that pixel is determined according to the aforementioned steps S11 to S15, and the determination result is used as the labeling information for that pixel, thereby constructing a scrap steel sample dataset covering multiple material types. Based on this, the establishment of this dataset provides a statistical foundation for the initial calibration of subsequent preset mapping relationships. Specifically, for all sample pixels of the same material type in the dataset, the distribution range of their relative intensity sequence values ​​under the corresponding rising and falling direction sequence features is statistically analyzed to determine the quantization threshold corresponding to that material type, thereby completing the initial calibration of the preset mapping relationship. In a specific example of this application, multispectral data of 14 types of samples were collected in the scrap steel inclusion identification scenario, including flat gray polyurethane, non-flat gray polyurethane, cold-rolled galvanized steel sheet, silver-painted wood board, silver-painted blue plastic board, white polyurethane, white plastic board, silver-white plastic board, natural solid wood, concrete cement block, silver-painted black plastic, black rubber, yellow rubber, and salt spray mixed corrosion products of cold-rolled galvanized sheet. The rising and falling direction sequence features and relative intensity sequence values ​​of each type of sample were extracted according to the aforementioned steps, and the quantization threshold ranges corresponding to each type of material were statistically obtained, thus completing the construction of the scrap steel sample dataset.

[0068] Regarding step S42, in this embodiment, step S42 is the process of collecting a small amount of labeled data in the target scene to construct a target scene sample dataset. The target scene refers to a new scene to be adapted, which is different from the preset scene, such as migrating from a laboratory environment to a steel plant workstation, or from one steel plant to another. The preset scene and the target scene are each independently selected from any one of the following: scrap steel inclusion recognition scene, steel defect detection scene, and galvanized cold-rolled steel plate corrosion detection scene, and they are different from each other. In this step, a small amount of multispectral images of samples with known material types are collected for the target scene. The material type of each pixel is determined and labeled according to the same feature extraction process. The number of samples in the constructed target scene sample dataset is less than the number of samples in the scrap steel sample dataset under the preset scene, typically about 10% to 20% of the preset scene data. This small amount of data is used to capture the spectral value shifts introduced by changes in light source color temperature, camera response differences, or changes in ambient lighting conditions in the target scene, rather than for retraining or rebuilding the entire recognition logic. In a specific example of this application, the construction of a scrap steel sample dataset has been completed in a laboratory environment. Now, it is necessary to deploy the implementation of this application to the scrap steel quality inspection station at a steel plant. At the station, approximately 15% of the scrap steel samples from the source scene are collected, and each pixel is labeled to construct the target scene sample dataset for the steel plant. The size of this dataset is much smaller than that of the source scene dataset.

[0069] Regarding step S43, in this embodiment, step S43 is a process of fine-tuning the threshold parameters in the preset mapping relationship based on the distribution statistical characteristics of the target scene sample dataset to obtain the target mapping relationship. Specifically, for all sample pixels of the same material type in the target scene sample dataset, the distribution of their relative intensity sequence values ​​under the corresponding vertical direction sequence features is re-statistically analyzed. For example, the quantiles of each relative intensity threshold are calculated. Based on the statistical results, the threshold interval of the relative intensity sequence values ​​of the corresponding material type in the preset mapping relationship is shifted or scaled to match the actual distribution of this type of material in the target scene, thereby obtaining a target mapping relationship adapted to the target scene. This fine-tuning process only adjusts the relative intensity threshold interval in the form of continuous numerical values, without changing the matching logic of the vertical direction sequence features or the sequential recognition order between various materials, because the vertical direction sequence features reflect the topological skeleton of the spectral curve, which belongs to the physical essential properties of the material itself and remains stable under different scenes. In a specific example of this application, the threshold range for the first relative intensity sequence value corresponding to smooth gray polyurethane in the preset mapping relationship is 2.40 to 2.42. After statistical analysis of the smooth gray polyurethane samples collected in the target scene, the first relative intensity sequence value shifts to the range of 2.50 to 2.52. Therefore, the threshold range is updated to 2.50 to 2.52, and the target mapping relationship is obtained. The threshold ranges for other material types are adjusted accordingly based on the statistical distribution of the target scene sample dataset.

[0070] Regarding step S44, in this embodiment, step S44 applies the adapted target mapping relationship to the actual detection process of the scrap steel to be identified in the target scene. Specifically, a spectral image of the material surface of the scrap steel to be identified in the target scene is acquired. For each pixel in the image, the sequence constraint group, differential sequence group, relative intensity sequence group, and rising / falling direction sequence group of the pixel are extracted according to the aforementioned steps S11 to S14. Then, the initial preset mapping relationship is replaced with the target mapping relationship obtained in step S43. The material type corresponding to each pixel is determined by stepwise matching according to the sequential identification order in the target mapping relationship and the adjusted relative intensity threshold range. Based on this, since the target mapping relationship has been fine-tuned and adapted with a small amount of labeled data in the target scene, its threshold range matches the actual imaging conditions of the target scene. Therefore, when performing inference in the target scene, the spectral value shift introduced by scene changes can be effectively eliminated, and accurate identification results can be obtained without re-collecting large-scale labeled data. In a specific example of this application, a multispectral image of the scrap steel to be identified, collected at the scrap steel quality inspection station of a steel plant, is input into the embodiment of this application. For each pixel in the image, steps S11 to S14 are executed sequentially to extract features. Then, the updated target mapping relationship in step S43 is used to perform step-by-step threshold matching. For example, if the ascending and descending direction sequence of a pixel is "ascend, descend, descend, descend, descend", and the first relative intensity sequence value is 2.51, and the threshold range of non-flat gray polyurethane in the target mapping relationship has been updated to 2.50 to 2.52, then the pixel is determined to be non-flat gray polyurethane, thereby completing the material type determination of each pixel of the scrap steel to be identified in the target scene.

[0071] Further, step S44 includes: Step S441: Obtain the surface spectral image of the scrap steel to be identified in the target scene, and determine the first material type of each pixel in the surface spectral image of the scrap steel to be identified in the target scene by combining the target mapping relationship. Step S442: Input the surface spectral image of the scrap steel to be identified in the target scene into the pre-trained deep convolutional neural network model to obtain the second material type of each pixel in the surface spectral image of the scrap steel to be identified in the target scene. Step S443: Perform adaptive weighted fusion of the first material type and the second material type of each pixel to determine the material type of each pixel in the spectral image of the surface of the scrap steel to be identified in the target scene.

[0072] In this embodiment, step S441 is the process of determining the first material type based on the target mapping relationship. This step adopts the feature extraction and judgment logic based on physical constraints in steps S11 to S15 above. That is, for each pixel in the spectral image of the surface of the scrap steel material to be identified in the target scene, the sequence constraint group, the difference sequence group, the relative intensity sequence group, and the rise and fall direction sequence group are extracted sequentially. Then, the target mapping relationship obtained in step S43 is used to replace the initial preset mapping relationship. According to the sequential identification order in the target mapping relationship and the adjusted relative intensity threshold range, the material type corresponding to the pixel is determined step by step. This result is used as the first material type. On this basis, the determination of the first material type is entirely based on the physical constraints of materials science. Its identification result has clear physical interpretability. Each determination conclusion can be traced back to the specific rise and fall direction sequence features and the relative intensity sequence value threshold matching process.

[0073] In a specific example of this application, for a certain pixel in the target scene, the ascending / descending direction sequence group is extracted as "ascending, descending, descending, descending, descending", and the relative intensity sequence value is s=[2.51, -0.682, -0.741, -0.357, -0.914]. Combined with the fact that the threshold range of the first relative intensity sequence value corresponding to the non-flat gray polyurethane in the target mapping relationship has been updated to 2.50 to 2.52, the first relative intensity sequence value of this pixel, 2.51, falls into this range, thus determining that the first material type is non-flat gray polyurethane.

[0074] In this embodiment, step S442 is the process of determining the second material type based on a deep convolutional neural network model. This step inputs the spectral image of the surface of the scrap steel material to be identified in the target scene into a pre-trained deep convolutional neural network model, which outputs the material type corresponding to each pixel as the second material type. The deep convolutional neural network model is a three-layer cascaded one-dimensional convolutional neural network. The first layer has a kernel size of 3×1, 32 output channels, a stride of 1, and a linear rectified unit activation function; the second layer has a kernel size of 2×1, 64 output channels, a stride of 1, and a linear rectified unit activation function; the third layer has a kernel size of 2×1, 64 output channels, a stride of 1, and a linear rectified unit activation function. Pooling operations are performed sequentially after each convolutional layer, and finally, a 64-dimensional deep feature vector is output through global average pooling. This vector is then passed through a fully connected layer and a normalized exponential function to output the probability distribution of each pixel for each material type. The training process of the deep convolutional neural network model is as follows: A dataset of scrap steel samples with labeled material types under a pre-defined scenario is used as the training set. The 6-channel spectral signal of each sample in the training set is used as input, and the manually labeled material types are used as supervision labels. The cross-entropy loss function is minimized through backpropagation, and the network parameters are iteratively updated until the model converges, resulting in a pre-trained deep convolutional neural network model. Based on this, the deep convolutional neural network model serves as an independent branch parallel to the physical constraint feature extraction branch, used to extract high-dimensional abstract statistical features that are difficult for the human eye to predict directly, thus complementing the physical constraint features.

[0075] In a specific example of this application, after inputting the surface spectral image of the scrap steel material to be identified in the target scene into a pre-trained deep convolutional neural network model, the model outputs the probability distribution of each material type for a certain pixel in the image. The probability of non-flat gray polyurethane is 0.82, and the probabilities of the other material types are all less than 0.10. Therefore, the second material type of the pixel is determined to be non-flat gray polyurethane.

[0076] In this embodiment, step S443 is the process of adaptively weighted fusion of the first material type and the second material type to determine the final material type. This step employs an attention mechanism to achieve collaborative optimization of the two recognition results. Specifically, the physical constraint feature vector corresponding to the first material type and the 64-dimensional depth feature vector corresponding to the second material type are concatenated to form a fused feature vector. Then, a fully connected layer automatically learns the weight allocation coefficients of the two types of features on the current sample. The weighted fused feature is then passed to a classifier to output the final material type determination result. The attention mechanism can dynamically adjust the contribution ratio of the physical constraint features and the depth features according to the specific spectral features of the current input sample. When the discrimination confidence of the physical constraint features is high, they are given a larger weight; when the depth features provide effective supplementary information, their weight is increased accordingly, thereby achieving complementary enhancement of the two types of features. Based on this, the adaptive weighted fusion combines the interpretability of physical constraints with the statistical modeling capabilities of deep learning. Especially when the relative intensity value of the physical constraint features is just near the threshold boundary and a single judgment logic may hesitate, the depth features can provide auxiliary discrimination information to help the fusion layer make a more robust final judgment.

[0077] In a specific example of this application, the first material type of a certain pixel is determined to be non-flat gray polyurethane, and its first relative intensity sequence value of 2.51 is located near the boundary of the threshold interval of 2.50 to 2.52. At the same time, the probability of non-flat gray polyurethane in the second material type output by the deep convolutional neural network model is 0.82 and the probability of flat gray polyurethane is 0.15. The weights learned by the attention mechanism are set to 0.6 for physical constraint features and 0.4 for depth features. After fusion, the comprehensive score of non-flat gray polyurethane is significantly higher than that of other categories, and the material type of the pixel is finally determined to be non-flat gray polyurethane.

[0078] In summary, the cross-scene adaptation method constituted by steps S41 to S44 in this application enables the aforementioned scrap steel material type identification method based on physical constraints to be rapidly deployed across different industrial scenarios. Specifically, the cross-scene adaptation capability of this method stems from the fact that the rise and fall direction sequence features and relative intensity threshold range stored in the preset mapping relationship are independent of the scene. The rise and fall direction sequence features reflect the topological skeleton of the material's spectral curve, which is determined by the material's own physical properties and remains stable under different imaging conditions. Therefore, only a small amount of labeled data from the target scene is needed to fine-tune the threshold parameters to complete the scene adaptation, without retraining the entire recognition model or rebuilding the mapping relationship. The data requirement for this fine-tuning process is far lower than the cost of re-collecting a large-scale labeled dataset, typically only requiring 10% to 20% of the source scene data, effectively reducing the data labeling overhead and adaptation time when deploying this application embodiment in new scenarios. Meanwhile, since the core judgment logic of the preset mapping relationship—namely, the matching rules of the rising and falling direction sequence features and the sequential recognition order between various materials—remains unchanged during the fine-tuning process, only the relative intensity threshold range in the form of continuous numerical values ​​is adjusted, the adapted target mapping relationship still has physical interpretability, and its judgment result can still be traced back to the specific rising and falling direction sequence features and relative intensity sequence values, ensuring the reliability and consistency of the recognition process in different scenarios.

[0079] In this embodiment, step S44 further integrates the first material type determined based on the target mapping relationship with the second material type output by the deep convolutional neural network model through adaptive weighted fusion. This results in a final material type determination that combines the stability of physical constraint features with the high-dimensional abstract representation capability of deep features. The first material type is generated by a physical constraint branch based on ascending / descending direction sequence features and relative intensity threshold matching, exhibiting clear physical interpretability and scale invariance. The second material type is generated by extracting 64-dimensional deep features from a three-layer cascaded one-dimensional convolutional neural network, capable of capturing high-order statistical patterns in spectral signals that are difficult to predefine through manually designed rules. By using an attention mechanism to concatenate the two types of features and automatically learn weight allocation, this embodiment can dynamically adjust the contribution ratio of the physical constraint branch and the deep branch according to the specific spectral features of the current input sample. When the confidence level of the physical constraint features is high, physical determination takes precedence; when the deep features can provide effective supplementary information, their weights are correspondingly increased, thereby achieving synergistic optimization of the two types of features. This fusion mechanism further improves recognition accuracy and generalization ability while maintaining physical interpretability. Especially when the relative intensity sequence value in the physical constraint features is near the threshold boundary, the deep features can provide auxiliary discrimination information to help the fusion layer make a more robust final judgment. Furthermore, this adaptive weighted fusion method does not change the independent working ability of the physical constraint branch. Even when the deep convolutional neural network model cannot function properly, it can still complete the recognition task solely based on the first material type, ensuring the system's operational reliability.

[0080] The above scheme will now be illustrated with several specific examples.

[0081] Example 1 In a specific example of this application, an industrial quality inspection device is used as the execution subject, and the input consists of 6 sets of fixed wavelength spectral channel signals with wavelengths of 420nm, 530nm, 650nm, 780nm, 850nm, and 940nm. For a certain pixel position, the obtained sequence constraint group is data=[264,1167,900,610,470,112]. First, the sequence constraint group is subjected to first-order difference processing to obtain the difference sequence group Δ=[903,-267,-290,-140,-358]; then the mean absolute deviation M=391.6 is calculated, and each element in the difference sequence group is divided by M to obtain the relative intensity sequence group s=[2.306,-0.682,-0.741,-0.357,-0.914]; then the rising and falling direction sequence group dir=[rising, falling, falling, falling, falling] is generated according to the sign of each element; finally, the rising and falling direction sequence group and the relative intensity sequence group are input into the preset mapping relationship for matching. The rising and falling direction sequence group matches the first mapping relationship, and the identification is judged step by step according to the first mapping relationship. s1=2.306 does not fall into the threshold range of flat gray polyurethane and non-flat gray polyurethane, and s2=-0.682 meets the threshold condition of cold-rolled galvanized steel plate, that is, s2 is less than -0.6. Therefore, the material type of this pixel is determined to be cold-rolled galvanized steel plate. In the above process, the first-order difference processing transforms the static reflectance value into a dynamic trend feature, the mean absolute deviation normalization has scale invariance and can eliminate external interference such as detection environment, light source intensity, and shooting distance, and the rise and fall direction sequence group retains the inherent rise and fall topological skeleton of the spectral curve. Together, the three constitute a material morphology feature representation with physical interpretability.

[0082] Example 2 In a specific example of this application, taking the non-metallic inclusion identification scenario of scrap steel briquettes as the preset scenario, multispectral data of 14 types of samples, including galvanized cold-rolled steel plates, silver-painted wood, silver-painted blue plastic plates, raw wood boards, dry cement blocks, and rubber, are first collected in a laboratory environment. Following steps S11 to S15, the material type of each pixel is determined and labeled, constructing a scrap steel sample dataset. Based on the distribution statistics of the rising and falling direction sequence features and relative intensity sequence values ​​of various materials in this dataset, the threshold intervals corresponding to each material are calibrated, completing the initial calibration of the preset mapping relationship. Subsequently, taking the scrap steel briquette dismantling and quality inspection scenario at a steel plant as the target scenario, approximately 15% of the source scenario data is collected from the field and labeled. Based on the distribution statistics of the target scenario data, the relative intensity threshold intervals of various materials in the preset mapping relationship are fine-tuned. For example, the threshold interval of the first relative intensity sequence value corresponding to flat gray polyurethane is adjusted from 2.40 to 2.42 to 2.50 to 2.52, obtaining the target mapping relationship. Finally, the multispectral detection device was deployed at the steel plant's demolition and quality inspection station to collect spectral images of the material surface of the scrap steel to be identified. The material type of each pixel was determined by combining the target mapping relationship, and the type and location information of non-metallic inclusions were output.

[0083] Example 3 In a specific example of this application, the target scenario is a salt spray corrosion detection of cold-rolled galvanized steel sheets. In this scenario, the objects to be identified are various corrosion products generated on the surface of the steel sheet after the salt spray test. Therefore, the pre-defined mapping relationship needs to be supplemented with logic for determining the type of corrosion product. First, multispectral data of corrosion samples such as a light zinc oxide corrosion layer and a severe mixed corrosion product of zinc chloride and basic zinc carbonate are collected. Following steps S11 to S14, the rising and falling direction sequence features and relative intensity sequence values ​​are extracted to determine the threshold ranges for various corrosion products. These are then expanded into the pre-defined mapping relationship according to the sequential identification order in this scenario. Specifically, the determination condition for a light zinc oxide corrosion layer is that the first relative intensity sequence value is between 2.40 and 2.42, and the determination condition for a severe mixed corrosion product of zinc chloride and basic zinc carbonate is that the second relative intensity sequence value is between -0.5 and -0.1. Subsequently, a small amount of labeled data from the target scenario is collected to fine-tune the threshold parameters, completing the scenario adaptation. For cold-rolled galvanized steel sheet samples that have undergone a 72-hour neutral salt spray test, spectral images of the material surface were collected, and the type of corrosion products for each pixel was determined by combining the adapted target mapping relationship. The test results show that the detection time for a single sample is only in the millisecond to second range, while the detection time using the manual visual analysis method is tens of minutes. The detection efficiency of the embodiment of this application is improved by more than tens of times, and the evaluation consistency is stable and unaffected by factors such as personnel fatigue and experience differences.

[0084] Example 4 In a specific example of this application, to verify the cross-scenario generalization ability of the embodiments of this application, comparative experiments on recognition accuracy were conducted in three scenarios: scrap steel inclusion recognition, steel defect detection, and galvanized cold-rolled steel plate corrosion detection. In all three scenarios, the embodiments of this application extracted physical constraint features according to the aforementioned steps S11 to S15, and performed recognition inference after completing scenario adaptation according to steps S41 to S44. In contrast, a traditional pure convolutional neural network model was trained on the same source scenario data, and then fine-tuned with the same small amount of target scenario data before recognition was performed. Experimental results show that in the three different scenarios, the recognition accuracy of the embodiments of this application is significantly higher than that of the traditional pure convolutional neural network model, verifying that the cross-scenario generalization ability of the embodiments of this application is significantly better than that of the pure data-driven model.

[0085] Example 5 In a specific example of this application, to verify the anti-interference capability of the embodiments of this application, robustness tests were conducted in a complex industrial environment with noise. The tests were performed under standard environmental conditions and noisy environmental conditions. The recognition accuracy of the embodiments of this application remained stable under standard conditions, while the recognition accuracy under noisy conditions only decreased slightly. In contrast, the recognition accuracy of traditional pure convolutional neural network models decreased significantly under noisy conditions. The accuracy decrease of the embodiments of this application is significantly smaller than that of traditional pure convolutional neural network models. This is due to the scale invariance of the mean absolute deviation normalization process, which effectively eliminates external interference such as the detection environment, light source intensity, and shooting distance. Simultaneously, the ascending and descending direction sequence group retains the topological skeleton of the spectral curve and is insensitive to local noise. This verifies that the anti-interference capability of the embodiments of this application is significantly better than that of traditional pure data-driven models, and can cover the recognition needs of 14 types of easily confused industrial materials and corrosion products.

[0086] Example 6 In a specific example of this application, to verify the technological advancement of the embodiments of this application, a comparative analysis is conducted from three dimensions: the fineness of the judgment logic, the scope of morphological coverage, and the physical theoretical basis. Compared with the technical architecture of earlier solutions that only supported three levels of judgment logic and eight types of morphological recognition, the embodiments of this application expand the judgment hierarchy to a step-by-step threshold matching based on the multi-mapping relationship of the ascending and descending direction sequence component flow. The judgment logic is more refined, and the ability to distinguish detailed features of spectral evolution is greatly improved. The morphological categories are expanded from eight to fourteen, covering more easily confused morphologies in industrial scenarios, including flat gray polyurethane, non-flat gray polyurethane, silver-painted blue plastic plates, and various corrosion products. The physical theoretical support is expanded from a single theory to a theoretical system integrating spectral analysis, topology, and statistics, making the physical basis of the ascending and descending direction sequence group as a topological feature more solid, and the threshold fine-tuning mechanism during cross-scenario adaptation has statistical theoretical support.

[0087] Example 7 In a specific example of this application, a deep convolutional neural network branch can be optionally introduced based on the aforementioned physical constraint feature extraction branch, forming a dual-drive recognition architecture. The deep convolutional neural network branch employs a three-layer cascaded one-dimensional convolutional neural network. The first layer has a 3×1 kernel and outputs 32 dimensions; the second layer has a 2×1 kernel and outputs 64 dimensions; and the third layer has a 2×1 kernel and outputs 64 dimensions, all with a stride of 1. The activation function is a linear rectified unit (RCU). Each convolutional layer is followed by a pooling operation, ultimately outputting a 64-dimensional deep feature vector through global average pooling. This branch works in parallel with the physical constraint feature extraction branch. The physical constraint branch outputs 15 dimensions of physical constraint features: a 5-dimensional first-order difference sequence, a 5-dimensional relative intensity sequence, and a 5-dimensional ascent / descend direction sequence. The deep branch outputs 64-dimensional deep features. The adaptive weighted fusion layer uses an attention mechanism, concatenating the two types of features through channels and automatically learning weight allocation through a fully connected layer to output fused features. The attention mechanism dynamically adjusts the contribution ratio of physical constraint features and deep features based on the specific spectral characteristics of the current input sample. Optionally, in scenarios requiring pixel-level material category probability distribution output, an encoder-decoder network is introduced for upsampling to restore the spatial resolution of the input image. Classification prediction is then performed independently for each pixel location, outputting the material category probability distribution for each pixel, used for area proportion statistics and distribution heatmap generation. Optionally, during model training, a materials science constraint regularization term is superimposed on the traditional cross-entropy loss function. When the model predicts a sample as a certain material type, it verifies whether the sample's ascending / descending direction sequence features and relative intensity sequence values ​​conform to the feature rules in the mapping relationship corresponding to that material type. If not, a penalty term is added to force the model's output to conform to the physical laws of materials science.

[0088] Since the electronic device described in this embodiment is an electronic device used to implement the information processing method in the embodiments of this application, those skilled in the art can understand the specific implementation methods and various variations of the electronic device in this embodiment based on the information processing method described in the embodiments of this application. Therefore, how the electronic device implements the method in the embodiments of this application will not be described in detail here. Any electronic device used by those skilled in the art to implement the information processing method in the embodiments of this application falls within the scope of protection of this application.

[0089] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0090] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0091] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0092] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0093] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0094] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for identifying scrap steel material types, characterized in that, The method includes: Acquire the spectral image of the material surface of the scrap steel to be identified, and for each pixel in the spectral image of the material surface of the scrap steel to be identified, obtain the reflectance corresponding to X preset spectral channels, and sort the reflectance according to the preset order corresponding to the X preset spectral channels to form a sequence constraint group; Perform first-order difference processing on the sequence constraint group to obtain a difference sequence group; The differential sequence group is normalized to obtain the relative intensity sequence group; Based on the differential sequence group or the relative intensity sequence group, the rising and falling direction sequence group is obtained; Based on the current detection conditions, the material type corresponding to the pixel is determined based on the rising and falling direction sequence group, the relative intensity sequence group, and the preset mapping relationship. The preset mapping relationship includes: the sequential identification order between different material types, and the correspondence between the rising and falling direction sequence features and the relative intensity sequence features and the material type.

2. The method for identifying scrap steel material types as described in claim 1, characterized in that, According to the preset order, the wavelengths of the X preset spectral channels are respectively: 420nm, 530nm, 650nm, 780nm, 850nm, and 940nm.

3. The method for identifying scrap steel material types as described in claim 1, characterized in that, If the current inspection mode is scrap steel quality inspection mode, the preset mapping relationship includes at least a first mapping relationship, a second mapping relationship, and a third mapping relationship; the first mapping relationship, the second mapping relationship, and the third mapping relationship correspond to different lifting direction sequence features; Based on the rising and falling direction sequence group, the relative intensity sequence group, and the preset mapping relationship, the material type corresponding to the pixel is determined, including: The target mapping relationship is determined from the first mapping relationship, the second mapping relationship, and the third mapping relationship based on the ascending and descending direction sequence group; If the target mapping relationship is the first mapping relationship or the third mapping relationship, then the material type corresponding to the pixel is determined based on the relative intensity sequence group and the target mapping relationship; If the target mapping relationship is the second mapping relationship, and the rising and falling direction sequence feature corresponding to the second mapping relationship is rising, falling, falling, falling, rising, then the material type corresponding to the pixel is determined to be a silver-white plastic plate.

4. The method for identifying scrap steel material types as described in claim 3, characterized in that, The sequence of rising and falling directions corresponding to the first mapping relationship is rising, falling, falling, falling, falling; the order of identification between different material types corresponding to the first mapping relationship is as follows: flat gray polyurethane, non-flat gray polyurethane, white plastic board, white polyurethane, silver-painted wood board, cold-rolled galvanized steel board, silver-painted blue plastic board. Based on the relative intensity sequence group and the target mapping relationship, the material type corresponding to the pixel is determined, including: The smooth gray polyurethane corresponds to the first relative intensity sequence value in the relative intensity sequence group. If the first relative intensity sequence value is greater than or equal to 2.4 and less than or equal to 2.42, then the material type corresponding to the pixel is determined to be smooth gray polyurethane. The non-flat gray polyurethane corresponds to the first relative intensity sequence value in the relative intensity sequence group. If the first relative intensity sequence value is greater than or equal to 2.5 and less than or equal to 2.52, then the material type corresponding to the pixel is determined to be non-flat gray polyurethane. The white plastic board corresponds to the second relative intensity sequence value in the relative intensity sequence group. If the second relative intensity sequence value is greater than or equal to -0.032 and less than or equal to 0.003, then the material type corresponding to the pixel is determined to be white plastic board. The white polyurethane corresponds to the fifth relative intensity sequence value in the relative intensity sequence group. If the fifth relative intensity sequence value is less than or equal to -1.7, then the material type corresponding to the pixel is determined to be white polyurethane. The silver-painted wood panel corresponds to the third relative intensity sequence value in the relative intensity sequence group. If the third relative intensity sequence value is greater than or equal to -0.6, then the material type corresponding to this pixel is determined to be silver-painted wood panel. The cold-rolled galvanized steel sheet corresponds to the second relative strength sequence value in the relative strength sequence group. If the second relative strength sequence value is less than -0.6, then the material type corresponding to the pixel is determined to be cold-rolled galvanized steel sheet. The silver-painted blue plastic board corresponds to the second and third relative intensity sequence values ​​in the relative intensity sequence group. If the second relative intensity sequence value is greater than or equal to -0.5 and less than or equal to -0.2, and the third relative intensity sequence value is less than or equal to -0.6, then the material type corresponding to this pixel is determined to be a silver-painted blue plastic board.

5. The method for identifying scrap steel material types as described in claim 3, characterized in that, The ascending / descending direction sequence feature corresponding to the third mapping relationship is ascending, ascending, descending, descending, descending; the sequential identification order between different material types corresponding to the third mapping relationship is as follows: black rubber, natural solid wood, yellow rubber, concrete cement block; Based on the relative intensity sequence group and the target mapping relationship, the material type corresponding to the pixel is determined, including: The black rubber corresponds to the first relative intensity sequence value in the relative intensity sequence group. If the first relative intensity sequence value is greater than or equal to 2.41 and less than or equal to 2.5, then the material type corresponding to the pixel is determined to be black rubber. Natural solid wood corresponds to the second relative intensity sequence value in the relative intensity sequence group. If the second relative intensity sequence value is greater than or equal to 1.5 and less than or equal to 2.5, then the material type corresponding to the pixel is determined to be natural solid wood. The yellow rubber corresponds to the second relative intensity sequence value in the relative intensity sequence group. If the second relative intensity sequence value is greater than or equal to 1.111 and less than or equal to 1.212, then the material type corresponding to this pixel is determined to be yellow rubber. The concrete cement block corresponds to the second relative strength sequence value in the relative strength sequence group. If the second relative strength sequence value is greater than or equal to 0.2 and less than or equal to 0.7, then the material type corresponding to the pixel is determined to be a concrete cement block.

6. A method for identifying scrap steel material types as described in claim 4 or 5, characterized in that, Based on the relative intensity sequence group and the target mapping relationship, the material type corresponding to the pixel is determined, including: If none of the aforementioned material types match under the current target mapping relationship, and the second relative intensity sequence value in the relative intensity sequence group is greater than or equal to -0.77 and less than or equal to 0.43, then the material type corresponding to the pixel is determined to be silver-sprayed black plastic; silver-sprayed black plastic is a fallback determination type, with a priority lower than all standard material types.

7. The method for identifying scrap steel material types as described in claim 1, characterized in that, If the current testing condition is the salt spray corrosion testing mode for cold-rolled galvanized steel sheets, the preset mapping relationship also includes a fourth mapping relationship. The rising and falling direction sequence feature corresponding to the fourth mapping relationship is rising, falling, falling, falling, falling. The corresponding identification order between different material types is as follows: light zinc oxide corrosion layer, heavy mixed corrosion of zinc chloride and basic zinc carbonate, non-flat gray polyurethane, white plastic sheet, white polyurethane, silver-painted wood board, cold-rolled galvanized steel sheet, and silver-painted blue plastic sheet. The method further includes: Based on the relative intensity sequence group and the fourth mapping relationship, the material type corresponding to the pixel is determined, including: The mild zinc oxide corrosion layer corresponds to the first relative intensity sequence value in the relative intensity sequence group. If the first relative intensity sequence value is greater than or equal to 2.4 and less than or equal to 2.42, then the material type corresponding to the pixel is determined to be a mild zinc oxide corrosion layer. The severe mixed corrosion of zinc chloride and zinc carbonate corresponds to the second relative intensity sequence value in the relative intensity sequence group. If the second relative intensity sequence value is greater than or equal to -0.5 and less than or equal to -0.1, then the material type corresponding to the pixel is determined to be severe mixed corrosion of zinc chloride and zinc carbonate. The non-flat gray polyurethane corresponds to the first relative intensity sequence value in the relative intensity sequence group. If the first relative intensity sequence value is greater than or equal to 2.5 and less than or equal to 2.52, then the material type corresponding to the pixel is determined to be non-flat gray polyurethane. The white plastic board corresponds to the second relative intensity sequence value in the relative intensity sequence group. If the second relative intensity sequence value is greater than or equal to -0.032 and less than or equal to 0.003, then the material type corresponding to the pixel is determined to be white plastic board. The white polyurethane corresponds to the fifth relative intensity sequence value in the relative intensity sequence group. If the fifth relative intensity sequence value is less than or equal to -1.7, then the material type corresponding to the pixel is determined to be white polyurethane. The silver-painted wood panel corresponds to the third relative intensity sequence value in the relative intensity sequence group. If the third relative intensity sequence value is greater than or equal to -0.6, then the material type corresponding to this pixel is determined to be silver-painted wood panel. The cold-rolled galvanized steel sheet corresponds to the second relative strength sequence value in the relative strength sequence group. If the second relative strength sequence value is less than -0.6, then the material type corresponding to the pixel is determined to be cold-rolled galvanized steel sheet. The silver-painted blue plastic board corresponds to the second and third relative intensity sequence values ​​in the relative intensity sequence group. If the second relative intensity sequence value is greater than or equal to -0.5 and less than or equal to -0.2, and the third relative intensity sequence value is less than or equal to -0.6, then the material type corresponding to this pixel is determined to be a silver-painted blue plastic board.

8. A method for identifying scrap steel material types, characterized in that, The method includes: In a preset scenario, the scrap steel material type identification method as described in any one of claims 1-7 is used to determine the material type corresponding to the scrap steel sample pixels, so as to label the scrap steel sample pixels and construct a scrap steel sample dataset; The material type corresponding to the sample pixels of the target scene to be adapted is collected to annotate the sample pixels of the target scene and construct the target scene sample dataset; the preset scene and the target scene are different scenes, and the preset scene and the target scene are any one of the following scenarios: scrap steel inclusion recognition scene, steel defect detection scene, and galvanized cold-rolled steel plate corrosion detection scene; the number of samples in the target scene sample dataset is less than the number of samples in the scrap steel sample dataset. The preset mapping relationship is adjusted based on the distribution statistical characteristics of the target scene sample dataset to obtain the target mapping relationship; Acquire the spectral image of the material surface of the scrap steel to be identified in the target scene, and combine it with the target mapping relationship to determine the material type of each pixel in the spectral image of the material surface of the scrap steel to be identified in the target scene.

9. The method for identifying scrap steel material types as described in claim 8, characterized in that, Acquire the spectral image of the material surface of the scrap steel to be identified in the target scene, and determine the material type of each pixel in the spectral image of the material surface of the scrap steel to be identified in the target scene by combining it with the target mapping relationship, including: Acquire the surface spectral image of the scrap steel to be identified in the target scene, and combine it with the target mapping relationship to determine the first material type of each pixel in the surface spectral image of the scrap steel to be identified in the target scene; The surface spectral image of the scrap steel to be identified in the target scene is input into a pre-trained deep convolutional neural network model to obtain the second material type of each pixel in the surface spectral image of the scrap steel to be identified in the target scene. Adaptive weighted fusion of the first and second material types of each pixel is performed to determine the material type of each pixel in the spectral image of the surface of the scrap steel to be identified in the target scene.

Images