Garbage classification system based on visual recognition

Abstract

The application discloses a garbage classification system based on visual recognition and belongs to the technical field of garbage classification. A spectrum recognition module acquires a multispectral image of a garbage target, constructs a spectral reflectance curve, and determines a garbage type. A structured light parameter determination module determines the illumination wavelength and spatial modulation frequency of a structured light fringe according to the garbage type. A structured light projection and image processing module constructs a spectral reflectance curve according to the illumination wavelength and spatial modulation frequency to obtain a deformed fringe image. A three-dimensional topography and feature extraction module reconstructs three-dimensional topography data of the internal structural layer of the garbage target according to the deformed fringe image, extracts periodic texture period parameters and interlayer distance features, and a material classification module associates and maps the absorption peak position of the spectral reflectance curve with the periodic texture period parameters and interlayer distance features to construct a cross-layer physical property feature vector and obtain a category recognition result. The application solves the problem of low garbage classification and recognition accuracy.

Description

Technical Field

[0001] This invention relates to the field of waste sorting technology, and in particular to a waste sorting system based on visual recognition. Background Technology

[0002] In recent years, with the rapid development of deep learning technology, visual recognition-based waste sorting systems have been widely applied. Existing technologies typically employ the following technical approach: acquiring visible light images of waste targets using image acquisition devices, and then utilizing deep learning models such as convolutional neural networks to detect and classify the waste targets in the images. Some solutions further incorporate multispectral imaging technology, using spectral information of waste targets in near-infrared and other bands to assist in material identification; others employ multimodal data fusion strategies, fusing visible light images with spectral, weight, and other sensor data to improve classification accuracy. Furthermore, addressing the challenge of detecting transparent objects in waste identification, existing research has improved network structures to enhance the model's feature extraction capabilities for transparent objects.

[0003] However, the existing technologies still have the following drawbacks: Current visual recognition methods for waste sorting rely on the surface appearance features or surface spectral information of the waste as the basis for classification. This means they can only acquire information about the color, texture, shape, or surface material of the waste, but cannot obtain the three-dimensional morphology and structural features of the internal structural layers. For transparent and semi-transparent waste, their surface appearance features are easily affected by the background environment, transmitted light, and internal residues, leading to a significant decrease in the accuracy of traditional visual classification methods. Experimental data shows that the recognition rate of pure visual solutions drops below 70% in scenarios involving transparent plastic bags and oily waste. Research indicates that glass and transparent plastic are highly prone to misclassification due to their similar optical properties. Although there are existing technologies that use near-infrared spectral analysis for plastic material classification and research on using structured light technology for three-dimensional reconstruction of transparent objects, these technologies are all independent applications. No solution has yet been found that combines spectral analysis with structured light three-dimensional reconstruction to obtain information about the internal structural layers of transparent and semi-transparent waste to assist in classification and recognition. Therefore, how to overcome optical interference from the surface of transparent and semi-transparent waste and obtain information about its internal structural layers to achieve accurate classification is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0004] The technical problem solved by this invention is that existing visual waste sorting technologies suffer from low sorting and recognition accuracy because they cannot obtain information about the internal structural layers of waste.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: As a preferred embodiment of the visual recognition-based waste sorting system described in this invention, wherein: The spectral recognition module acquires multispectral images of the garbage target under multiple narrowband spectral illumination conditions. Based on the pixel intensity response of the garbage target region in the multispectral image, it constructs the spectral reflectance curve of the garbage target at each pixel point and determines the garbage type of the garbage target based on the waveform characteristics of the spectral reflectance curve. The structured light parameter determination module determines the illumination wavelength of the structured light stripes that penetrate the surface layer of the waste target and project onto the internal structural layer based on the waste type of the waste target and the transmission peak wavelength and reflection valley wavelength in the spectral reflectance curve. It also determines the spatial modulation frequency of the structured light stripes based on the thickness estimate of the waste target. The structured light projection and image processing module projects multi-layer coded structured light onto the garbage target according to the illumination wavelength and spatial modulation frequency of the structured light stripes. It then acquires a composite structured light deformation image formed by superposition of reflection from the front surface of the garbage target, internal scattering, and transmission from the rear surface. Based on the absorption coefficient in the spectral reflectance curve, it performs frequency domain filtering and scattering component striping on the composite structured light deformation image to obtain a deformation stripe image. The three-dimensional morphology and feature extraction module reconstructs the three-dimensional morphology data of the internal structural layers of the waste target based on the deformed stripe image, and extracts the periodic texture period parameters and interlayer distance features of the internal structural layers of the waste target. The material classification module associates and maps the absorption peak position of the spectral reflectance curve with the periodic texture period parameter and the interlayer distance feature to construct a cross-layer physical attribute feature vector of the waste target. This vector is then input into the trained material classifier, which outputs the category recognition result of the waste target.

[0006] Furthermore, multispectral images of the garbage target under multiple narrowband spectral illumination conditions are acquired. Based on the pixel intensity response of the garbage target region in the multispectral images, a spectral reflectance curve of the garbage target at each pixel is constructed. The garbage type of the garbage target is determined based on the waveform characteristics of the spectral reflectance curve, including: The multispectral light source is controlled to illuminate the garbage target with different narrowband wavelengths in sequence, and monochrome images of the garbage target under each narrowband wavelength are acquired simultaneously to obtain multiple frames of single-channel spectral images corresponding to different narrowband wavelengths; Pixel-level image registration is performed on the single-channel spectral images corresponding to different narrowband wavelengths in the multiple frames to eliminate pixel offsets caused by the movement of garbage targets or differences in light source angles between the single-channel spectral images, and the registered multi-channel spectral image stacked data is obtained. Based on the registered multi-channel spectral image stacking data, the pixel gray values ​​of the garbage target area in each channel are extracted, and the continuous spectral reflectance curve of each pixel of the garbage target is generated by fitting the wavelength as the horizontal axis and the normalized gray value as the vertical axis. By performing waveform analysis on the continuous spectral reflectance curve, the characteristics of the number of transmission peaks and the depth of absorption valleys in the continuous spectral reflectance curve within a preset full-band range are extracted. Based on the transmission peak quantity characteristics and absorption valley depth characteristics, the waste type is determined using a preset dual-threshold decision tree to obtain the waste type of the waste target; The types of waste targeted include opaque waste, semi-transparent waste, and transparent waste.

[0007] Furthermore, based on the transmission peak quantity characteristics and the absorption valley depth characteristics, a waste type determination is performed using a preset dual-threshold decision tree to obtain the waste type of the waste target, including: When the number of transmission peaks is 0, the garbage target is determined to be opaque garbage, and a first type of physical parameter set is output. The first type of physical parameter set includes a first illumination wavelength range and a first spatial modulation frequency. When the number of transmission peaks is 1, the garbage target is determined to be semi-transparent garbage, and a second type of physical parameter set is output. The second type of physical parameter set includes a second illumination wavelength range and a second spatial modulation frequency, and the second illumination wavelength range is negatively correlated with the center wavelength of the absorption valley depth feature. When the number of transmission peaks is greater than or equal to 2, the garbage target is determined to be transparent garbage, and a third type of physical parameter set is output. The third type of physical parameter set includes a third illumination wavelength range and a third spatial modulation frequency, and the third spatial modulation frequency is higher than the second spatial modulation frequency.

[0008] Furthermore, based on the type of waste target, and according to the transmission peak wavelength and reflection valley wavelength in the spectral reflectance curve, the illumination wavelength for structured light stripes to penetrate the surface layer of the waste target and project onto the internal structural layer is determined, and the spatial modulation frequency of the structured light stripes is determined based on the estimated thickness of the waste target, including: When the garbage target's garbage type is opaque garbage: The illumination wavelength of the structured light stripes is determined to be the first fixed wavelength in the visible light band; The spatial modulation frequency is determined as the first fixed frequency; When the target's trash type is semi-transparent trash: Extract the valley wavelength corresponding to the main absorption valley with the largest absorption valley depth in the spectral reflectance curve, calculate the illumination wavelength offset based on the valley wavelength and absorption valley depth value of the main absorption valley, and determine the illumination wavelength of the structured light stripe as the sum of the valley wavelength and the illumination wavelength offset, so that the illumination wavelength of the structured light stripe deviates from the valley wavelength of the main absorption valley. Based on the estimated thickness of the waste target, the spatial modulation frequency of the structured light stripes is calculated using a preset thickness frequency mapping function; When the garbage target's garbage type is transparent garbage: Extract the peak wavelengths corresponding to all transmission peaks in the spectral reflectance curve, select the peak wavelength with the highest transmittance as the basic illumination wavelength, and fine-tune the basic illumination wavelength according to the valley depth weighting coefficient in the absorption valley depth feature to obtain the illumination wavelength of the structured light stripe. Based on the estimated thickness of the waste target, the initial spatial modulation frequency is calculated using a preset thickness frequency mapping function, and the initial spatial modulation frequency is multiplied by a preset multiplication factor to obtain the spatial modulation frequency of the structured light stripe, wherein the multiplication factor is greater than 1. The illumination wavelength and spatial modulation frequency of the structured light stripes are used to control the structured light projector to project multi-layer coded structured light onto the garbage target.

[0009] Further, the valley wavelength corresponding to the main absorption valley with the largest absorption valley depth in the spectral reflectance curve is extracted. An illumination wavelength offset is calculated based on the valley wavelength and absorption valley depth of the main absorption valley. The illumination wavelength of the structured light stripe is determined as the sum of the valley wavelength and the illumination wavelength offset, so that the illumination wavelength of the structured light stripe deviates from the valley wavelength of the main absorption valley. This includes: Extract the valley wavelength, left half-peak wavelength, and right half-peak wavelength corresponding to the main absorption valley with the largest absorption valley depth in the spectral reflectance curve, and calculate the asymmetry coefficient of the main absorption valley based on the valley wavelength, the left half-peak wavelength, and the right half-peak wavelength. The asymmetry coefficient is the ratio of the distance from the valley wavelength to the left half-peak wavelength and the right half-peak wavelength, respectively. Obtain the absorption valley depth value of the main absorption valley, and calculate the illumination wavelength offset based on the absorption valley depth value and the asymmetry coefficient. The illumination wavelength offset is biased towards the right half-peak direction when the asymmetry coefficient is greater than 1, and biased towards the left half-peak direction when the asymmetry coefficient is less than 1. The illumination wavelength of the structured light stripe is determined as the sum of the valley wavelength and the illumination wavelength offset; The spatial modulation frequency of the structured light stripe is adjusted according to the asymmetry coefficient. The greater the deviation of the asymmetry coefficient from 1, the greater the adjustment of the spatial modulation frequency towards the lower frequency direction.

[0010] Further, the peak wavelengths corresponding to all transmission peaks in the spectral reflectance curve are extracted, and the peak wavelength with the highest transmittance is selected as the basic illumination wavelength. The basic illumination wavelength is then fine-tuned according to the valley depth weighting coefficient in the absorption valley depth feature to obtain the illumination wavelength of the structured light stripes, including: Extract the peak wavelengths corresponding to all transmission peaks in the spectral reflectance curve and the transmittance value of each transmission peak, and select the peak wavelength with the highest transmittance as the basic illumination wavelength. Extract the valley wavelength and valley depth value corresponding to the absorption valley closest to the basic illumination wavelength in the spectral reflectance curve. Calculate the valley depth weighting coefficient based on the wavelength spacing between the valley wavelength and the basic illumination wavelength, and the valley depth value. The valley depth weighting coefficient is positively correlated with the valley depth value and negatively correlated with the wavelength spacing. The difference between the product of the base illumination wavelength and the valley depth weighting coefficient and the preset fine-tuning reference step size is used as the illumination wavelength of the structured light stripe, so that the illumination wavelength of the structured light stripe is shifted away from the absorption valley based on the base illumination wavelength. Where there is no absorption valley in the spectral reflectance curve, the basic illumination wavelength is directly used as the illumination wavelength of the structured light stripe.

[0011] Further, based on the illumination wavelength and spatial modulation frequency of the structured light stripes, multi-layer coded structured light is projected onto the waste target. A composite structured light deformation image is acquired, formed by superposition of reflection from the front surface, internal scattering, and transmission from the rear surface of the waste target. Based on the absorption coefficient in the spectral reflectance curve, the composite structured light deformation image is subjected to frequency domain filtering and scattering component stripping processing to obtain a deformed stripe image, including: Based on the illumination wavelength of the structured light stripes, the output light source wavelength of the structured light projector is controlled to be consistent with the illumination wavelength, and a first layer of coded structured light and a second layer of coded structured light with corresponding stripe densities are generated according to the spatial modulation frequency of the structured light stripes. The stripe phases of the first layer of coded structured light and the second layer of coded structured light are 90 degrees apart. The structured light projector is controlled to sequentially project the first layer of coded structured light and the second layer of coded structured light onto the garbage target, and the first composite structured light deformation image and the second composite structured light deformation image are simultaneously acquired after reflection from the front surface of the garbage target, internal scattering and transmission from the rear surface of the garbage target are superimposed. Two-dimensional Fourier transforms are performed on the first composite structured light deformation image and the second composite structured light deformation image respectively to transform the first composite structured light deformation image and the second composite structured light deformation image from the spatial domain to the spatial frequency domain, thereby obtaining the first composite frequency domain spectrum and the second composite frequency domain spectrum. Based on the absorption coefficient in the spectral reflectance curve, an adaptive frequency domain filter kernel function is constructed. The passband cutoff frequency of the filter kernel function is negatively correlated with the absorption coefficient, and the stopband attenuation slope is positively correlated with the absorption coefficient. Based on the spatial modulation frequency of the structured light stripes, the positions of the fundamental frequency component and harmonic components are determined in the first composite frequency domain spectrum and the second composite frequency domain spectrum, respectively. The filtering kernel function is used to perform passband filtering on the spectral components at the fundamental frequency component positions in each composite frequency domain spectrum, while the spectral components at the harmonic component positions are subjected to stopband suppression, to obtain the first filtered fundamental frequency spectrum component and the second filtered fundamental frequency spectrum component. Inverse two-dimensional Fourier transforms are performed on the first filtered fundamental frequency spectrum component and the second filtered fundamental frequency spectrum component respectively to restore them to the spatial domain, resulting in a first deformed stripe image and a second deformed stripe image. The deformed stripe image is then synthesized based on the phase difference between the first deformed stripe image and the second deformed stripe image.

[0012] Furthermore, based on the absorption coefficient in the spectral reflectance curve, an adaptive frequency domain filtering kernel function for the absorption coefficient is constructed, including: Obtain the absorption coefficient value in the spectral reflectance curve corresponding to the illumination wavelength of the structured light stripe, and use it as the target absorption coefficient; Based on the target absorption coefficient, the passband cutoff frequency is calculated using a preset cutoff frequency mapping function, wherein the cutoff frequency mapping function is a monotonically decreasing function so that the passband cutoff frequency decreases as the target absorption coefficient increases. Based on the target absorption coefficient, the stopband attenuation slope is calculated using a preset attenuation slope mapping function, wherein the attenuation slope mapping function is a monotonically increasing function so that the stopband attenuation slope increases as the target absorption coefficient increases. Extract the local slope value of the spectral curve in the neighborhood of the illumination wavelength from the spectral reflectance curve, and calculate the filter kernel smoothing coefficient based on the local slope value. The filter kernel smoothing coefficient is positively correlated with the absolute value of the local slope value. A frequency domain filter kernel function is constructed based on the passband cutoff frequency, the stopband attenuation slope, and the filter kernel smoothing coefficient. The transfer function value of the frequency domain filter kernel function approaches 1 in the passband and approaches 0 in the stopband. The width of the transition band between the passband and the stopband is controlled by the filter kernel smoothing coefficient. Wherein, when the target absorption coefficient exceeds the preset upper limit threshold of the absorption coefficient, the passband cutoff frequency is set to the preset lower limit value of the minimum cutoff frequency.

[0013] Further, based on the deformed stripe image, the three-dimensional morphological data of the internal structural layers of the waste target are reconstructed, and the periodic texture period parameters and interlayer distance features of the internal structural layers of the waste target are extracted, including: When the waste target is opaque waste, phase demodulation processing is performed on the deformed stripe image to obtain the three-dimensional surface topography data of the opaque waste, and the surface roughness features and surface texture periodic parameters of the three-dimensional surface topography data are extracted; the surface texture periodic parameters are used as the periodic texture periodic parameters of the opaque waste, and the interlayer distance feature is set to a preset zero value to indicate that the opaque waste does not have an internal structural layer; When the waste type of the waste target is semi-transparent waste, phase demodulation processing is performed on the deformed stripe image to obtain the three-dimensional topography data of the internal structure layer of the semi-transparent waste. Frequency domain transformation is performed on the three-dimensional topography data of the internal structure layer to extract the periodic texture period parameters of the internal structure layer. Based on the phase difference between the phase value of each pixel in the deformed stripe image and the calibrated reference phase value, the average depth position of the internal structure layer is calculated. The difference between the average depth position and the thickness estimate of the waste target is used as the interlayer distance feature. When the waste type of the target waste is transparent waste, phase demodulation processing is performed on the deformed stripe image to obtain the three-dimensional morphological data of the internal structural layer of the transparent waste. Two-dimensional autocorrelation analysis is performed on the three-dimensional morphological data of the internal structural layer. The periodic texture periodic parameters of the internal structural layer are extracted based on the peak position of the autocorrelation function. Based on the mapping relationship between the local deformation degree of the stripes in the deformed stripe image and the local depth value of the internal structural layer, a depth distribution map of the internal structural layer is constructed. The variance of the depth value of each local region in the depth distribution map is extracted as the interlayer distance feature.

[0014] Furthermore, the absorption peak positions of the spectral reflectance curve are correlated and mapped with the periodic texture periodic parameter and the interlayer distance feature to construct a cross-layer physical attribute feature vector of the waste target. This vector is then input into a trained material classifier, which outputs the waste target category identification result, including: Extract the wavelength positions of the N absorption valleys with the largest absorption valley depths from the spectral reflectance curve, and use them as the absorption peak position feature vectors, where N is a positive integer greater than or equal to 2; The periodic texture periodic parameter and the interlayer distance feature are normalized respectively to obtain normalized texture periodic parameter and normalized interlayer distance feature; Based on the type of waste in the target waste, a preset association mapping matrix corresponding to the type of waste is retrieved. The absorption peak position feature vector, the normalized texture period parameter, and the normalized interlayer distance feature are projected onto a cross-layer physical attribute feature space of a preset dimension using the preset association mapping matrix to obtain the cross-layer physical attribute feature vector. The cross-layer physical property feature vector is input into a trained material classifier. The material classifier outputs the probability values ​​of the garbage target belonging to each category based on the cross-layer physical property feature vector, and outputs the category with the highest probability value as the category recognition result of the garbage target. The preset association mapping matrix is ​​a diagonal matrix, and the diagonal elements of the diagonal matrix correspond to the mapping weights of the absorption peak position feature vector, the normalized texture period parameter, and the normalized interlayer distance feature in the cross-layer physical attribute feature space.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs the spectral reflectance curve of a waste target and determines the waste type through a spectral recognition module. A structured light parameter determination module adaptively determines the illumination wavelength and spatial modulation frequency of the structured light stripes based on the transmission peak wavelength and reflection valley wavelength in the spectral reflectance curve. This allows the structured light to adaptively penetrate the surface layer of the waste target and project onto its internal structural layer according to the material's optical properties. A structured light projection and image processing module acquires composite structured light deformation images and performs frequency domain filtering and scattering component striping based on the absorption coefficient to obtain deformed stripe images. A three-dimensional morphology and feature extraction module reconstructs the three-dimensional morphology data of the waste target's internal structural layer and extracts periodic texture period parameters and interlayer distance features. A material classification module correlates the absorption peak position of the spectral reflectance curve with the periodic texture period parameters and interlayer distance features to construct a cross-layer physical attribute feature vector for classification and identification. This achieves effective acquisition and utilization of information about the internal structural layer of transparent and semi-transparent waste targets, solving the problem of low classification accuracy in existing visual waste classification technologies due to the inability to acquire information about the internal structural layer of waste, and improving the classification accuracy of transparent and semi-transparent waste.

[0016] 2. This invention constructs the spectral reflectance curve of the waste target through a spectral recognition module and extracts the transmission peak quantity feature and absorption valley depth feature. Based on a preset dual-threshold decision tree, it determines the waste type as opaque waste, semi-transparent waste, or transparent waste. The structured light parameter determination module adaptively determines the illumination wavelength of the structured light stripes according to the transmission peak wavelength and reflection valley wavelength in the spectral reflectance curve. This allows the structured light to adaptively select the optimal penetration wavelength according to the spectral transmission characteristics of the waste target, avoiding the problem of missing internal structural layer information due to the inability of a fixed wavelength to effectively penetrate waste targets with different light transmission characteristics. At the same time, it adaptively determines the spatial modulation frequency according to the thickness estimate of the waste target, avoiding the problem of severe attenuation of a fixed frequency in thick media or insufficient resolution in thin media. This enables the effective acquisition of internal structural layer information of waste targets with different light transmission characteristics and different thicknesses.

[0017] 3. This invention acquires a composite structured light deformation image formed by the superposition of reflection from the front surface, internal scattering, and transmission from the rear surface of a waste target through a structured light projection and image processing module. Based on the absorption coefficient in the spectral reflectance curve, an adaptive frequency domain filtering kernel function is constructed. This kernel function is used to perform passband filtering on the fundamental frequency component of the composite structured light deformation image and stopband suppression on the harmonic components, resulting in a deformed stripe image after stripping the rear surface transmission component and the front surface reflection component. A three-dimensional morphology and feature extraction module reconstructs the three-dimensional morphology data of the internal structural layers of the waste target based on the deformed stripe image and extracts periodic texture period parameters and interlayer distance features. This achieves effective separation of the front surface reflection component, internal scattering component, and rear surface transmission component in the composite structured light deformation image, avoiding the problem of decreased phase demodulation accuracy due to the aliasing of various optical signals, which prevents accurate reconstruction of the three-dimensional morphology of the internal structural layers. This ensures that the acquired periodic texture period parameters and interlayer distance features can truly reflect the internal structural properties of the waste target.

[0018] 4. This invention uses a material classification module to correlate and map the absorption peak position of the spectral reflectance curve with periodic texture period parameters and interlayer distance features. Based on the waste type of the target waste, a corresponding preset correlation mapping matrix is ​​retrieved. Using this preset correlation mapping matrix, the absorption peak position feature vector, normalized texture period parameters, and normalized interlayer distance features are projected onto a cross-layer physical attribute feature space to construct a cross-layer physical attribute feature vector. This cross-layer physical attribute feature vector is then input into a trained material classifier to output the category recognition result. This achieves comprehensive characterization and classification of waste targets from three physical attribute dimensions: absorption peak position (chemical composition dimension), periodic texture period parameters (internal structure dimension), and interlayer distance features (spatial geometry dimension). This avoids the problem of difficulty in distinguishing waste targets with similar spectral characteristics due to insufficient information from a single modality, and improves the classification accuracy for transparent and semi-transparent waste. Attached Figure Description

[0019] Figure 1 A schematic diagram of the basic structure of a vision-based waste sorting system provided in one embodiment of the present invention; Figure 2 A schematic diagram illustrating the workflow of a vision-based waste sorting system according to an embodiment of the present invention; Figure 3 This is a basic flowchart illustrating the process of determining the type of waste target according to an embodiment of the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Example 1

[0021] like Figure 1 As shown in the figure, this embodiment introduces a visual recognition-based waste sorting system, including a spectral recognition module, a structured light parameter determination module, a structured light projection and image processing module, a three-dimensional morphology and feature extraction module, and a material classification module.

[0022] The spectral recognition module acquires multispectral images of waste targets under multiple narrowband spectral illumination conditions. Based on the pixel intensity response of the waste target region in the multispectral image, it constructs the spectral reflectance curve of the waste target at each pixel point, and determines the waste type of the waste target based on the waveform characteristics of the spectral reflectance curve. The output of the spectral recognition module is connected to the input of the structured light parameter determination module. The output of the spectral recognition module is also connected to the input of the structured light projection and image processing module and the material classification module.

[0023] The structured light parameter determination module determines the illumination wavelength of structured light stripes that penetrate the surface layer of the waste target and project onto the internal structural layer, based on the transmission peak wavelength and reflection valley wavelength in the spectral reflectance curve, according to the waste target's waste type. It also determines the spatial modulation frequency of the structured light stripes based on the estimated thickness of the waste target. The output of the structured light parameter determination module is connected to the input of the structured light projection and image processing module.

[0024] The structured light projection and image processing module projects multi-layered coded structured light onto the waste target based on the illumination wavelength and spatial modulation frequency of the structured light stripes. It then acquires a composite structured light deformation image formed by the superposition of reflections from the front surface, internal scattering, and transmission from the rear surface of the waste target. Based on the absorption coefficient in the spectral reflectance curve, the module performs frequency domain filtering and scattering component striping on the composite structured light deformation image to obtain the deformed stripe image. The output of the structured light projection and image processing module is connected to the input of the 3D topography and feature extraction module.

[0025] The 3D morphology and feature extraction module is used to reconstruct the 3D morphology data of the internal structural layers of the waste target based on the deformed stripe image, and to extract the periodic texture period parameters and interlayer distance features of the internal structural layers of the waste target. The output of the 3D morphology and feature extraction module is connected to the input of the material classification module.

[0026] The material classification module is used to associate and map the absorption peak position of the spectral reflectance curve with the periodic texture period parameters and interlayer distance features to construct a cross-layer physical attribute feature vector of the waste target. This vector is then input into the trained material classifier, which outputs the category recognition result of the waste target.

[0027] The specific structure and working process of each module are described in detail below.

[0028] I. Spectral Recognition Module.

[0029] The spectral recognition module includes a multispectral light source, an image acquisition device, and a spectral data processing unit.

[0030] A multispectral light source is used to sequentially irradiate the garbage target with different narrowband wavelengths. In one specific embodiment, the multispectral light source is implemented using a light-emitting diode array, which includes multiple light-emitting diodes with different center wavelengths, covering a range of 400 nm to 2500 nm, and the full width at half maximum (FWHM) of each light-emitting diode is no greater than 20 nm. In another specific embodiment, the multispectral light source is implemented using a broadband light source in conjunction with a tunable filter. The broadband light source is a halogen tungsten lamp or a xenon lamp, and the tunable filter is a liquid crystal tunable filter or an acousto-optic tunable filter. By adjusting the passband center wavelength of the filter, illumination light of different narrowband wavelengths is output sequentially. The angle between the illumination direction of the multispectral light source and the plane where the garbage target is located is 30 degrees to 60 degrees to ensure that the surface of the garbage target receives uniform illumination.

[0031] The image acquisition device is used to acquire monochrome images of the garbage target at various narrowband wavelengths. The device employs a charge-coupled device (CCD) image sensor or a complementary metal-oxide-semiconductor (CMOS) image sensor, whose spectral response range covers the output wavelength range of the multispectral light source. The optical axis of the image acquisition device is perpendicular to the plane containing the garbage target to ensure that the acquired images do not suffer from perspective distortion. Each time the multispectral light source outputs illumination at a narrowband wavelength, the image acquisition device triggers an acquisition, obtaining one frame of a single-channel spectral image corresponding to that narrowband wavelength. After sequentially completing the illumination and acquisition at all narrowband wavelengths, multiple frames of single-channel spectral images corresponding to different narrowband wavelengths are obtained.

[0032] The spectral data processing unit processes multiple frames of single-channel spectral images corresponding to different narrowband wavelengths. First, the spectral data processing unit performs pixel-level image registration on the multiple frames of single-channel spectral images. Because the waste target may be in motion on the conveyor, or the light source angles of different narrowband wavelengths may differ, pixel offsets may exist between the single-channel spectral images. Pixel-level image registration specifically employs a registration algorithm based on mutual information or a registration algorithm based on scale-invariant feature transformation feature point matching. In one specific implementation, a single-channel spectral image frame is selected as a reference image, and the remaining single-channel spectral images are registered to the coordinate system of the reference image through affine transformation or thin-plate spline transformation, eliminating pixel offsets between the single-channel spectral images and obtaining registered multi-channel spectral image stacked data. The multi-channel spectral image stacked data is a three-dimensional data block, where two dimensions are spatial coordinates and the third dimension is the wavelength channel.

[0033] The spectral data processing unit extracts the pixel grayscale values ​​of the garbage target region in each channel based on the registered multi-channel spectral image stacked data. The segmentation of the garbage target region can be achieved using threshold segmentation, edge detection, or a deep learning-based target detection network. In one specific implementation, a deep learning-based instance segmentation network is used to perform target detection and segmentation on a specific channel of the multi-channel spectral image to obtain an instance segmentation mask for the garbage target. This mask is then applied to all channels to extract the pixel grayscale values ​​of the garbage target region in each channel.

[0034] The spectral data processing unit uses wavelength as the abscissa and normalized grayscale value as the ordinate, employing spline interpolation or polynomial fitting methods to generate continuous spectral reflectance curves for each pixel of the waste target. The normalization process involves dividing the grayscale value of each channel's pixel by the grayscale value of the standard white board in that channel to obtain a dimensionless relative reflectance value. The standard white board is made of polytetrafluoroethylene (PTFE), and its reflectance is close to 100% in the 400 nm to 2500 nm wavelength range.

[0035] The spectral data processing unit performs waveform analysis on the continuous spectral reflectance curve, extracting the transmission peak quantity and absorption valley depth characteristics across a preset full-band range. Based on these characteristics and a preset dual-threshold decision tree, it determines the type of waste, identifying the target waste. Waste types include opaque waste, semi-transparent waste, and transparent waste.

[0036] II. Structured Light Parameter Determination Module.

[0037] The structured light parameter determination module receives the garbage type and spectral reflectance curve of the garbage target output by the spectral recognition module, whereby... Figure 3 This is a basic flowchart illustrating the process of determining the type of waste target according to an embodiment of the present invention.

[0038] For opaque waste, the structured light parameter determination module determines the illumination wavelength of the structured light stripes as a first fixed wavelength within the visible light band and the spatial modulation frequency as a first fixed frequency. In one specific embodiment, the first fixed wavelength is 550 nanometers, and the first fixed frequency is one stripe per millimeter. Opaque waste does not require structured light to penetrate its interior; a clear image of surface deformation stripes can be obtained using the visible light band and low-frequency stripes.

[0039] For translucent waste, the structured light parameter determination module extracts the valley wavelength corresponding to the main absorption valley with the largest absorption valley depth in the spectral reflectance curve. Based on the valley wavelength and valley depth, it calculates the illumination wavelength offset and determines the illumination wavelength of the structured light stripes as the sum of the valley wavelength and the illumination wavelength offset, thus ensuring that the illumination wavelength of the structured light stripes deviates from the valley wavelength of the main absorption valley. Simultaneously, the structured light parameter determination module calculates the spatial modulation frequency of the structured light stripes based on the estimated thickness of the waste target using a preset thickness frequency mapping function. The thickness estimate can be obtained through binocular visual parallax or laser ranging. The thickness frequency mapping function is a monotonically decreasing function; that is, the larger the estimated thickness, the lower the spatial modulation frequency.

[0040] For transparent waste, the structured light parameter determination module extracts the peak wavelengths corresponding to all transmission peaks in the spectral reflectance curve, selects the peak wavelength with the highest transmittance as the basic illumination wavelength, and fine-tunes the basic illumination wavelength according to the valley depth weighting coefficient in the absorption valley depth feature to obtain the illumination wavelength of the structured light stripes. Simultaneously, the structured light parameter determination module calculates the initial spatial modulation frequency based on the estimated thickness of the waste target using a preset thickness frequency mapping function, and multiplies the initial spatial modulation frequency by a preset multiplication factor to obtain the spatial modulation frequency of the structured light stripes. The multiplication factor is greater than 1. In one specific embodiment, the multiplication factor is any value between 1.5 and 3.0.

[0041] III. Structured Light Projection and Image Processing Module.

[0042] The structured light projection and image processing module includes a structured light projector, an image acquisition unit, and an image processing unit.

[0043] The structured light projector determines the illumination wavelength and spatial modulation frequency of the structured light stripes output by the module based on the structured light parameters, and projects multi-layered coded structured light onto the waste target. The structured light projector is implemented using a digital micromirror device or a liquid crystal projector, and its output wavelength and spatial modulation frequency are adjustable. In one specific embodiment, the structured light projector controls the output wavelength of its internal light source to match the illumination wavelength of the structured light stripes, and generates a coded structured light pattern with a corresponding stripe density based on the spatial modulation frequency of the structured light stripes. This coded structured light pattern is then projected onto the surface of the waste target using a digital micromirror device.

[0044] The image acquisition unit is used to acquire composite structured light deformation images formed by the superposition of reflections from the front surface of the waste target, internal scattering, and transmission from the rear surface. A synchronous trigger signal line is set between the image acquisition unit and the structured light projector. For each frame of coded structured light projected by the structured light projector, the image acquisition unit synchronously acquires one frame of composite structured light deformation image.

[0045] The image processing unit performs frequency domain filtering and scattering component stripping on the composite structured light deformation image to obtain a deformed fringe image. Specifically, the image processing unit performs a two-dimensional Fourier transform on the composite structured light deformation image, converting it from the spatial domain to the spatial frequency domain, resulting in a composite frequency domain spectrum. The image processing unit constructs an adaptive frequency domain filtering kernel function based on the absorption coefficient in the spectral reflectance curve. The passband cutoff frequency of this filtering kernel function is negatively correlated with the absorption coefficient, while the stopband attenuation slope is positively correlated with the absorption coefficient. The image processing unit determines the positions of the fundamental frequency component and harmonic components in the composite frequency domain spectrum based on the spatial modulation frequency of the structured light fringes. It then performs passband filtering on the spectral components at the fundamental frequency component positions and stopband suppression on the spectral components at the harmonic component positions using the filtering kernel function, obtaining the filtered fundamental frequency spectrum component. The image processing unit performs an inverse two-dimensional Fourier transform on the filtered fundamental frequency spectrum component, restoring it from the spatial frequency domain to the spatial domain, resulting in a deformed fringe image striped by removing the back surface transmission component and the front surface reflection component.

[0046] IV. 3D Shape and Feature Extraction Module.

[0047] The 3D morphology and feature extraction module receives the deformed stripe image output by the structured light projection and image processing module, reconstructs the 3D morphology data of the internal structural layer of the waste target based on the deformed stripe image, and extracts the periodic texture period parameters and interlayer distance features of the internal structural layer of the waste target.

[0048] The 3D topography and feature extraction module first performs phase demodulation processing on the deformed stripe image. Phase demodulation can be achieved using Fourier transform profilometry or phase shifting. In one specific implementation, Fourier transform profilometry is used: a two-dimensional Fourier transform is performed on the deformed stripe image to extract the fundamental frequency component. After performing an inverse Fourier transform on the fundamental frequency component, its phase is extracted to obtain a wrapping phase map. The wrapping phase map is then phase-unfolded to obtain a continuous phase distribution map. Based on the phase difference between the continuous phase distribution map and the calibration reference phase map, and combined with the principle of triangulation, the depth values ​​of each pixel in the internal structural layer of the waste target are calculated to obtain the 3D topography data.

[0049] The 3D topography and feature extraction module extracts periodic texture periodic parameters from the 3D topography data. Specifically, it performs a frequency domain transformation on the depth value matrix of the 3D topography data, extracts the peak frequency components in the frequency domain, and calculates the periodic texture periodic parameters based on the frequency values ​​corresponding to the peak frequency components. The units of the periodic parameters are millimeters or pixels.

[0050] The 3D topography and feature extraction module also extracts interlayer distance features based on the 3D topography data. Specifically, based on the phase difference between the phase value of each pixel in the deformed stripe image and the calibrated reference phase value, the average depth position of the internal structural layers is calculated, and the difference between this average depth position and the estimated thickness of the waste target is used as the interlayer distance feature. The unit of the interlayer distance feature is millimeters.

[0051] V. Material Classification Module.

[0052] The material classification module receives the spectral reflectance curve output by the spectral recognition module, and the periodic texture period parameters and interlayer distance features output by the three-dimensional morphology and feature extraction module.

[0053] The material classification module extracts the wavelength positions of the N absorption valleys with the largest absorption valley depths from the spectral reflectance curve, using these as feature vectors for the absorption peak positions. N is a positive integer greater than or equal to 2. In one specific implementation, N equals 3.

[0054] The material classification module normalizes the periodic texture period parameters and interlayer distance features to obtain normalized texture period parameters and normalized interlayer distance features. The normalization process employs either min-max normalization or Z-score normalization.

[0055] The material classification module retrieves a preset association mapping matrix corresponding to the waste type of the target waste. Using this matrix, it projects the absorption peak position feature vector, normalized texture period parameter, and normalized interlayer distance feature onto a preset-dimensional cross-layer physical attribute feature space, obtaining a cross-layer physical attribute feature vector. The preset association mapping matrix is ​​a diagonal matrix, where the diagonal elements correspond to the mapping weights of the absorption peak position feature vector, normalized texture period parameter, and normalized interlayer distance feature in the cross-layer physical attribute feature space. These mapping weights vary depending on the waste type. For opaque waste, the mapping weight of spectral features is greater than that of texture features and interlayer distance features; for semi-transparent waste, the mapping weights of each feature are balanced; for transparent waste, the mapping weights of texture features and interlayer distance features are greater than that of spectral features.

[0056] The material classification module inputs cross-layer physical attribute feature vectors into a trained material classifier. Based on these feature vectors, the classifier outputs probability values ​​for each category to which the waste target belongs, and outputs the category with the highest probability value as the classification result. The material classifier is implemented using a Support Vector Machine (SVM), Random Forest, or Lightweight Gradient Boosting Machine. In one specific implementation, the material classifier uses an SVM, which is pre-trained using a training sample set labeled with material categories. The training sample set includes several broad categories such as plastics, glass, metals, paper, textiles, and rubber, each further subdivided into multiple subcategories. Example 2

[0057] like Figure 2 As shown in the figure, this embodiment provides an overall description of the workflow of a visual recognition-based waste sorting system.

[0058] The waste target is transported to the detection station by the conveyor. The spectral recognition module first starts multispectral imaging to acquire multispectral images of the waste target under multiple narrowband spectral illumination conditions, constructs spectral reflectance curves and determines the waste type of the waste target.

[0059] The structured light parameter determination module determines the illumination wavelength and spatial modulation frequency of the structured light stripes based on the type of waste, the transmission peak wavelength and reflection valley wavelength in the spectral reflectance curve, and the estimated thickness of the waste target.

[0060] The structured light projection and image processing module projects multi-layer coded structured light onto the garbage target according to the determined illumination wavelength and spatial modulation frequency, acquires composite structured light deformation images, and performs frequency domain filtering and scattering component stripping processing based on the absorption coefficient to obtain deformed stripe images.

[0061] The 3D morphology and feature extraction module reconstructs the 3D morphology data of the internal structural layers of the waste target based on the deformed stripe image, and extracts the periodic texture period parameters and interlayer distance features.

[0062] The material classification module associates and maps the absorption peak position of the spectral reflectance curve with the periodic parameters of the periodic texture and the interlayer distance features to construct a cross-layer physical property feature vector, which is then output by the material classifier as the category identification result of the waste target.

[0063] The vision-based waste sorting system described in this embodiment can be applied to scenarios such as automatic sorting lines for household waste, recycling stations, and pre-processing lines for kitchen waste. The system can be installed above the waste conveyor belt and fixed in place by a bracket. A light shield is installed at the detection station to reduce ambient light interference.

[0064] 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 implemented on one or more computer-usable storage media containing computer-usable program code. It should be noted that the above embodiments are merely illustrative of the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A waste sorting system based on visual recognition, characterized in that, include: The spectral recognition module acquires multispectral images of the garbage target under multiple narrowband spectral illumination conditions. Based on the pixel intensity response of the garbage target region in the multispectral image, it constructs the spectral reflectance curve of the garbage target at each pixel point and determines the garbage type of the garbage target based on the waveform characteristics of the spectral reflectance curve. The structured light parameter determination module determines the illumination wavelength of the structured light stripes that penetrate the surface layer of the waste target and project onto the internal structural layer based on the waste type of the waste target and the transmission peak wavelength and reflection valley wavelength in the spectral reflectance curve. It also determines the spatial modulation frequency of the structured light stripes based on the thickness estimate of the waste target. The structured light projection and image processing module projects multi-layer coded structured light onto the garbage target according to the illumination wavelength and spatial modulation frequency of the structured light stripes. It then acquires a composite structured light deformation image formed by superposition of reflection from the front surface of the garbage target, internal scattering, and transmission from the rear surface. Based on the absorption coefficient in the spectral reflectance curve, it performs frequency domain filtering and scattering component striping on the composite structured light deformation image to obtain a deformation stripe image. The three-dimensional morphology and feature extraction module reconstructs the three-dimensional morphology data of the internal structural layers of the waste target based on the deformed stripe image, and extracts the periodic texture period parameters and interlayer distance features of the internal structural layers of the waste target. The material classification module associates and maps the absorption peak position of the spectral reflectance curve with the periodic texture period parameter and the interlayer distance feature to construct a cross-layer physical attribute feature vector of the waste target. This vector is then input into the trained material classifier, which outputs the category recognition result of the waste target.

2. The visual recognition-based waste sorting system as described in claim 1, characterized in that, Acquire multispectral images of the garbage target under multiple narrowband spectral illumination conditions. Based on the pixel intensity response of the garbage target region in the multispectral images, construct the spectral reflectance curve of the garbage target at each pixel point. Determine the garbage type of the garbage target based on the waveform characteristics of the spectral reflectance curve, including: The multispectral light source is controlled to illuminate the garbage target with different narrowband wavelengths in sequence, and monochrome images of the garbage target under each narrowband wavelength are acquired simultaneously to obtain multiple frames of single-channel spectral images corresponding to different narrowband wavelengths; Pixel-level image registration is performed on the single-channel spectral images corresponding to different narrowband wavelengths in the multiple frames to eliminate pixel offsets caused by the movement of garbage targets or differences in light source angles between the single-channel spectral images, and the registered multi-channel spectral image stacked data is obtained. Based on the registered multi-channel spectral image stacking data, the pixel gray values ​​of the garbage target area in each channel are extracted, and the continuous spectral reflectance curve of each pixel of the garbage target is generated by fitting the wavelength as the horizontal axis and the normalized gray value as the vertical axis. By performing waveform analysis on the continuous spectral reflectance curve, the characteristics of the number of transmission peaks and the depth of absorption valleys in the continuous spectral reflectance curve within a preset full-band range are extracted. Based on the transmission peak quantity characteristics and absorption valley depth characteristics, the waste type is determined using a preset dual-threshold decision tree to obtain the waste type of the waste target; The types of waste targeted include opaque waste, semi-transparent waste, and transparent waste.

3. The visual recognition-based waste sorting system as described in claim 2, characterized in that, Based on the transmission peak quantity characteristics and absorption valley depth characteristics, a waste type determination is performed using a preset dual-threshold decision tree to obtain the waste type of the waste target, including: When the number of transmission peaks is 0, the garbage target is determined to be opaque garbage, and a first type of physical parameter set is output. The first type of physical parameter set includes a first illumination wavelength range and a first spatial modulation frequency. When the number of transmission peaks is 1, the garbage target is determined to be semi-transparent garbage, and a second type of physical parameter set is output. The second type of physical parameter set includes a second illumination wavelength range and a second spatial modulation frequency, and the second illumination wavelength range is negatively correlated with the center wavelength of the absorption valley depth feature. When the number of transmission peaks is greater than or equal to 2, the garbage target is determined to be transparent garbage, and a third type of physical parameter set is output. The third type of physical parameter set includes a third illumination wavelength range and a third spatial modulation frequency, and the third spatial modulation frequency is higher than the second spatial modulation frequency.

4. The visual recognition-based waste sorting system as described in claim 1, characterized in that, Based on the type of waste target, the illumination wavelengths of structured light stripes used to penetrate the surface layer of the waste target and project onto the internal structural layer are determined according to the transmission peak wavelength and reflection valley wavelength in the spectral reflectance curve. The spatial modulation frequency of the structured light stripes is then determined based on the estimated thickness of the waste target, including: When the garbage target's garbage type is opaque garbage: The illumination wavelength of the structured light stripes is determined to be the first fixed wavelength in the visible light band; The spatial modulation frequency is determined as the first fixed frequency; When the target's trash type is semi-transparent trash: Extract the valley wavelength corresponding to the main absorption valley with the largest absorption valley depth in the spectral reflectance curve, calculate the illumination wavelength offset based on the valley wavelength and absorption valley depth value of the main absorption valley, and determine the illumination wavelength of the structured light stripe as the sum of the valley wavelength and the illumination wavelength offset, so that the illumination wavelength of the structured light stripe deviates from the valley wavelength of the main absorption valley. Based on the estimated thickness of the waste target, the spatial modulation frequency of the structured light stripes is calculated using a preset thickness frequency mapping function; When the garbage target's garbage type is transparent garbage: Extract the peak wavelengths corresponding to all transmission peaks in the spectral reflectance curve, select the peak wavelength with the highest transmittance as the basic illumination wavelength, and fine-tune the basic illumination wavelength according to the valley depth weighting coefficient in the absorption valley depth feature to obtain the illumination wavelength of the structured light stripe. Based on the estimated thickness of the waste target, the initial spatial modulation frequency is calculated using a preset thickness frequency mapping function, and the initial spatial modulation frequency is multiplied by a preset multiplication factor to obtain the spatial modulation frequency of the structured light stripe, wherein the multiplication factor is greater than 1. The illumination wavelength and spatial modulation frequency of the structured light stripes are used to control the structured light projector to project multi-layer coded structured light onto the garbage target.

5. The visual recognition-based waste sorting system as described in claim 4, characterized in that, Extract the valley wavelength corresponding to the main absorption valley with the largest absorption valley depth from the spectral reflectance curve. Calculate the illumination wavelength offset based on the valley wavelength and absorption valley depth of the main absorption valley. Determine the illumination wavelength of the structured light stripe as the sum of the valley wavelength and the illumination wavelength offset, so that the illumination wavelength of the structured light stripe deviates from the valley wavelength of the main absorption valley. This includes: Extract the valley wavelength, left half-peak wavelength, and right half-peak wavelength corresponding to the main absorption valley with the largest absorption valley depth in the spectral reflectance curve, and calculate the asymmetry coefficient of the main absorption valley based on the valley wavelength, the left half-peak wavelength, and the right half-peak wavelength. The asymmetry coefficient is the ratio of the distance from the valley wavelength to the left half-peak wavelength and the right half-peak wavelength, respectively. Obtain the absorption valley depth value of the main absorption valley, and calculate the illumination wavelength offset based on the absorption valley depth value and the asymmetry coefficient. The illumination wavelength offset is biased towards the right half-peak direction when the asymmetry coefficient is greater than 1, and biased towards the left half-peak direction when the asymmetry coefficient is less than 1. The illumination wavelength of the structured light stripe is determined as the sum of the valley wavelength and the illumination wavelength offset; The spatial modulation frequency of the structured light stripe is adjusted according to the asymmetry coefficient. The greater the deviation of the asymmetry coefficient from 1, the greater the adjustment of the spatial modulation frequency towards the lower frequency direction.

6. The visual recognition-based waste sorting system as described in claim 4, characterized in that, The peak wavelengths corresponding to all transmission peaks in the spectral reflectance curve are extracted, and the peak wavelength with the highest transmittance is selected as the basic illumination wavelength. The basic illumination wavelength is then fine-tuned according to the valley depth weighting coefficient in the absorption valley depth feature to obtain the illumination wavelength of the structured light stripes, including: Extract the peak wavelengths corresponding to all transmission peaks in the spectral reflectance curve and the transmittance value of each transmission peak, and select the peak wavelength with the highest transmittance as the basic illumination wavelength. Extract the valley wavelength and valley depth value corresponding to the absorption valley closest to the basic illumination wavelength in the spectral reflectance curve. Calculate the valley depth weighting coefficient based on the wavelength spacing between the valley wavelength and the basic illumination wavelength, and the valley depth value. The valley depth weighting coefficient is positively correlated with the valley depth value and negatively correlated with the wavelength spacing. The difference between the product of the base illumination wavelength and the valley depth weighting coefficient and the preset fine-tuning reference step size is used as the illumination wavelength of the structured light stripe, so that the illumination wavelength of the structured light stripe is shifted away from the absorption valley based on the base illumination wavelength. Where there is no absorption valley in the spectral reflectance curve, the basic illumination wavelength is directly used as the illumination wavelength of the structured light stripe.

7. The visual recognition-based waste sorting system as described in claim 1, characterized in that, Based on the illumination wavelength and spatial modulation frequency of the structured light stripes, multi-layer coded structured light is projected onto the waste target. A composite structured light deformation image is acquired, formed by the superposition of reflection from the front surface, internal scattering, and transmission from the rear surface of the waste target. Based on the absorption coefficient in the spectral reflectance curve, the composite structured light deformation image is subjected to frequency domain filtering and scattering component stripping to obtain a deformed stripe image, including: Based on the illumination wavelength of the structured light stripes, the output light source wavelength of the structured light projector is controlled to be consistent with the illumination wavelength, and a first layer of coded structured light and a second layer of coded structured light with corresponding stripe densities are generated according to the spatial modulation frequency of the structured light stripes. The stripe phases of the first layer of coded structured light and the second layer of coded structured light are 90 degrees apart. The structured light projector is controlled to sequentially project the first layer of coded structured light and the second layer of coded structured light onto the garbage target, and the first composite structured light deformation image and the second composite structured light deformation image are simultaneously acquired after reflection from the front surface of the garbage target, internal scattering and transmission from the rear surface of the garbage target are superimposed. Two-dimensional Fourier transforms are performed on the first composite structured light deformation image and the second composite structured light deformation image respectively to transform the first composite structured light deformation image and the second composite structured light deformation image from the spatial domain to the spatial frequency domain, thereby obtaining the first composite frequency domain spectrum and the second composite frequency domain spectrum. Based on the absorption coefficient in the spectral reflectance curve, an adaptive frequency domain filter kernel function is constructed. The passband cutoff frequency of the filter kernel function is negatively correlated with the absorption coefficient, and the stopband attenuation slope is positively correlated with the absorption coefficient. Based on the spatial modulation frequency of the structured light stripes, the positions of the fundamental frequency component and harmonic components are determined in the first composite frequency domain spectrum and the second composite frequency domain spectrum, respectively. The filtering kernel function is used to perform passband filtering on the spectral components at the fundamental frequency component positions in each composite frequency domain spectrum, while the spectral components at the harmonic component positions are subjected to stopband suppression, to obtain the first filtered fundamental frequency spectrum component and the second filtered fundamental frequency spectrum component. Inverse two-dimensional Fourier transforms are performed on the first filtered fundamental frequency spectrum component and the second filtered fundamental frequency spectrum component respectively to restore them to the spatial domain, resulting in a first deformed stripe image and a second deformed stripe image. The deformed stripe image is then synthesized based on the phase difference between the first deformed stripe image and the second deformed stripe image.

8. The visual recognition-based waste sorting system as described in claim 7, characterized in that, Based on the absorption coefficient in the spectral reflectance curve, an adaptive frequency domain filtering kernel function for the absorption coefficient is constructed, including: Obtain the absorption coefficient value in the spectral reflectance curve corresponding to the illumination wavelength of the structured light stripe, and use it as the target absorption coefficient; Based on the target absorption coefficient, the passband cutoff frequency is calculated using a preset cutoff frequency mapping function, wherein the cutoff frequency mapping function is a monotonically decreasing function so that the passband cutoff frequency decreases as the target absorption coefficient increases. Based on the target absorption coefficient, the stopband attenuation slope is calculated using a preset attenuation slope mapping function, wherein the attenuation slope mapping function is a monotonically increasing function so that the stopband attenuation slope increases as the target absorption coefficient increases. Extract the local slope value of the spectral curve in the neighborhood of the illumination wavelength from the spectral reflectance curve, and calculate the filter kernel smoothing coefficient based on the local slope value. The filter kernel smoothing coefficient is positively correlated with the absolute value of the local slope value. A frequency domain filter kernel function is constructed based on the passband cutoff frequency, the stopband attenuation slope, and the filter kernel smoothing coefficient. The transfer function value of the frequency domain filter kernel function approaches 1 in the passband and approaches 0 in the stopband. The width of the transition band between the passband and the stopband is controlled by the filter kernel smoothing coefficient. Wherein, when the target absorption coefficient exceeds the preset upper limit threshold of the absorption coefficient, the passband cutoff frequency is set to the preset lower limit value of the minimum cutoff frequency.

9. The visual recognition-based waste sorting system as described in claim 1, characterized in that, Based on the deformed stripe image, the three-dimensional morphological data of the internal structural layers of the waste target are reconstructed, and the periodic texture period parameters and interlayer distance features of the internal structural layers of the waste target are extracted, including: When the waste target is opaque waste, phase demodulation processing is performed on the deformed stripe image to obtain the three-dimensional surface topography data of the opaque waste, and the surface roughness features and surface texture periodic parameters of the three-dimensional surface topography data are extracted; the surface texture periodic parameters are used as the periodic texture periodic parameters of the opaque waste, and the interlayer distance feature is set to a preset zero value to indicate that the opaque waste does not have an internal structural layer; When the waste type of the waste target is semi-transparent waste, phase demodulation processing is performed on the deformed stripe image to obtain the three-dimensional topography data of the internal structure layer of the semi-transparent waste. Frequency domain transformation is performed on the three-dimensional topography data of the internal structure layer to extract the periodic texture period parameters of the internal structure layer. Based on the phase difference between the phase value of each pixel in the deformed stripe image and the calibrated reference phase value, the average depth position of the internal structure layer is calculated. The difference between the average depth position and the thickness estimate of the waste target is used as the interlayer distance feature. When the waste type of the target waste is transparent waste, phase demodulation processing is performed on the deformed stripe image to obtain the three-dimensional morphological data of the internal structural layer of the transparent waste. Two-dimensional autocorrelation analysis is performed on the three-dimensional morphological data of the internal structural layer. The periodic texture periodic parameters of the internal structural layer are extracted based on the peak position of the autocorrelation function. Based on the mapping relationship between the local deformation degree of the stripes in the deformed stripe image and the local depth value of the internal structural layer, a depth distribution map of the internal structural layer is constructed. The variance of the depth value of each local region in the depth distribution map is extracted as the interlayer distance feature.

10. The visual recognition-based waste sorting system as described in claim 1, characterized in that, The absorption peak positions of the spectral reflectance curve are correlated and mapped with the periodic texture periodicity parameter and the interlayer distance feature to construct a cross-layer physical attribute feature vector of the waste target. This vector is then input into a trained material classifier, which outputs the waste target category identification result, including: Extract the wavelength positions of the N absorption valleys with the largest absorption valley depths from the spectral reflectance curve, and use them as the absorption peak position feature vectors, where N is a positive integer greater than or equal to 2; The periodic texture periodic parameter and the interlayer distance feature are normalized respectively to obtain normalized texture periodic parameter and normalized interlayer distance feature; Based on the type of waste in the target waste, a preset association mapping matrix corresponding to the type of waste is retrieved. The absorption peak position feature vector, the normalized texture period parameter, and the normalized interlayer distance feature are projected onto a cross-layer physical attribute feature space of a preset dimension using the preset association mapping matrix to obtain the cross-layer physical attribute feature vector. The cross-layer physical property feature vector is input into a trained material classifier. The material classifier outputs the probability values ​​of the garbage target belonging to each category based on the cross-layer physical property feature vector, and outputs the category with the highest probability value as the category recognition result of the garbage target. The preset association mapping matrix is ​​a diagonal matrix, and the diagonal elements of the diagonal matrix correspond to the mapping weights of the absorption peak position feature vector, the normalized texture period parameter, and the normalized interlayer distance feature in the cross-layer physical attribute feature space.

Images