Color sorter efficient sorting method based on multispectral feature fusion
By combining multispectral feature fusion and dynamic feature weight adjustment with a multimodal attention mechanism, the problem of rigid feature fusion in color sorters is solved, achieving efficient and stable material sorting.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-06
- Publication Date
- 2026-03-17
AI Technical Summary
Existing multi-feature fusion color sorting methods suffer from rigid feature fusion strategies, leading to information loss and poor adaptability to complex scenarios, which affects sorting efficiency and stability.
A multispectral feature fusion method is adopted, which dynamically adjusts feature weights and strengthens key features through a dynamic feature fusion strategy and a multimodal attention mechanism. Combined with multi-band imaging, color and texture feature extraction, an improved support vector machine is used for sorting decision.
It improves the sorting accuracy and stability of color sorters, reduces the fluctuation of sorting accuracy in complex scenarios, and significantly improves sorting efficiency and processing speed.
Smart Images
Figure CN121669575A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of color sorting machines, in particular to a high-efficiency sorting method for a color sorting machine based on multi-spectral feature fusion. BACKGROUND
[0002] As the core equipment for material sorting, the sorting efficiency and accuracy of the color sorting machine directly affect the quality and cost of industrial production. With the increasing demand for sorting in industrial production, the traditional color sorting machine based on single feature (such as color) sorting method has been unable to meet the demand, and multi-feature fusion (such as color + texture) has become the key direction to improve the sorting performance.
[0003] However, the existing multi-feature fusion color sorting machine method has the core problem of rigid feature fusion strategy, which is specifically manifested in:
[0004] Information loss or conflict: The traditional method usually adopts weighted summation, principal component analysis (PCA) or simple linear discriminant for feature fusion. The fixed weight distribution method cannot intelligently adapt to the importance difference of different features. For example, in the scene where the material surface is clean and unobstructed, the color feature contributes more to the sorting decision; while in the scene where the material has wear and tear, the texture feature is more important. Fixed weight will cause key features to be submerged or irrelevant information to be amplified, reducing the decision robustness.
[0005] Poor adaptability to complex scenes: In actual production, materials often overlap, obstruct or have shadows, etc. At this time, the stability of the extracted feature parameters decreases. The rigid fusion strategy cannot dynamically adjust the feature weight, which further amplifies the feature instability, resulting in reduced sorting stability of the color sorting machine. The fluctuation range of sorting accuracy can reach 5%-10% in long-term operation, which seriously affects the sorting efficiency.
[0006] Therefore, there is an urgent need for a color sorting machine sorting method that can dynamically adjust the feature fusion strategy and adapt to complex scenes to solve the problem of reduced efficiency in the prior art. SUMMARY
[0007] The purpose of the present application is to provide a high-efficiency sorting method for a color sorting machine based on multi-spectral feature fusion, which solves the problems of information loss and poor adaptability to complex scenes caused by rigid feature fusion in traditional methods through dynamic feature fusion strategy and multi-modal attention mechanism, and improves the sorting efficiency and stability of the color sorting machine.
[0008] To achieve the above purpose, the present application is implemented by the following technical solutions:
[0009] The present application provides a high-efficiency sorting method for a color sorting machine based on multi-spectral feature fusion, comprising the following steps:
[0010] Multi-spectral data acquisition: through the multi-band imaging assembly of the color sorter, image data of the material to be sorted under at least two different wavebands is collected, and the wavebands include a visible light waveband and a near-infrared waveband;
[0011] Multi-feature extraction: based on the multi-band image data collected in the multi-spectral data acquisition step, color feature parameters of the material are extracted through a color feature extraction submodule, and texture feature parameters of the material are extracted through a texture feature extraction submodule;
[0012] Dynamic feature fusion: the color feature parameters and the texture feature parameters extracted in the multi-feature extraction step are input into a dynamic weight distribution module, the dynamic weight distribution module calculates dynamic weight values of different features according to real-time state parameters of the material, and key features are strengthened in combination with a multi-modal attention mechanism module to output a fusion feature vector;
[0013] Sorting decision: the fusion feature vector output in the dynamic feature fusion step is input into a sorting decision module, the sorting decision module determines the category of the material through a preset classification model, generates a control signal of an execution mechanism of the color sorter, and completes sorting of the material.
[0014] Further, the multi-band imaging assembly in step 1 includes a visible light camera and a near-infrared camera, the resolution of the visible light camera is not less than 1920*1080, the frame rate is not less than 30 fps, the resolution of the near-infrared camera is not less than 1280*720, the frame rate is not less than 30 fps, and the shooting trigger signals of the two cameras are synchronized, and the synchronization error is less than 1 ms.
[0015] Further, the color feature extraction submodule in step 2 extracts color feature parameters using an HSV color space model, specifically including hue mean, saturation variance, and lightness gradient, wherein the value range of hue is 0°-360°, and the value range of saturation and lightness is 0-1; the texture feature extraction submodule extracts texture feature parameters using a gray level co-occurrence matrix, specifically including contrast, correlation, energy, and uniformity, and the calculation distance of the gray level co-occurrence matrix is 1 pixel and the gray level is 256 levels.
[0016] Further, the acquisition method of the real-time state parameters of the material in step 3 is: identifying the material region through an image segmentation algorithm based on semantic segmentation, calculating the material overlap degree O and the occlusion area ratio C; calculating the shadow intensity S through a shadow detection algorithm based on double-waveband gray difference, wherein the gray difference is the difference between the near-infrared waveband image gray value and the visible light waveband image gray value.
[0017] Further, the weight calculation process of the dynamic weight distribution module in step 3 includes:
[0018] Constructing state evaluation function: taking material overlap degree, shielding area proportion, shadow intensity as input, calculating state evaluation value E, formula is E = a x O + b x C + g x S, wherein a = 0.4, b = 0.3, g = 0.3, and a + b + g = 1;
[0019] Dynamic weight mapping: preset low state threshold E1 = 0.3, high state threshold E2 = 0.6, color feature weight Wc and texture feature weight Wt are obtained by piecewise function mapping according to state evaluation value E, when E < 0.3, Wc = 0.6-0.2 x E, Wt = 0.4+0.2 x E; when 0.3 <= E <= 0.6, Wc = 0.4-0.4 x (E-0.3) / (0.6-0.3), Wt = 0.6+0.4 x (E-0.3) / (0.6-0.3); when E > 0.6, Wc = 0.2, Wt = 0.8.
[0020] Further, the multi-modal attention mechanism module in step 3 includes a self-attention network and a feature adjustment unit, the input of the self-attention network is the dynamically weighted color feature vector and texture feature vector, the network structure includes 2 fully connected layers and 1 activation layer, and the attention weight matrix is obtained by training; the feature adjustment unit adjusts the weight of each dimension of the feature vector element by element according to the attention weight matrix, and the adjustment formula is F_i = F_i x A_i.
[0021] Further, the training sample of the self-attention network includes at least 5000 groups of multi-spectral feature data of materials in different states, each sample contains color feature vector, texture feature vector, material real-time state parameter and artificial labeled key feature label, the training process uses Adam optimizer, the initial value of learning rate is set to 0.001, and the learning rate is attenuated to 0.8 times of the previous one every 100 epochs, and the training iteration number is not less than 1000 times.
[0022] Further, the classification model of the sorting decision module in step 4 adopts an improved support vector machine (SVM) model, the kernel function of the model is a radial basis function, and the kernel function formula is K(x, y) = exp(-g||x-y||2), wherein g is the kernel function parameter; the model parameters are optimized by 5-fold cross-validation method, and the optimization target is the highest classification accuracy, and the finally determined penalty coefficient C is in the range of 1-10, and the kernel function parameter g is in the range of 0.1-1.
[0023] Further, the control signal of the actuator in step 4 includes a jet valve control signal and a conveying belt speed regulation signal, the response delay of the jet valve control signal is less than 5ms, and the jet time can be adaptively adjusted in the range of 10-50ms according to the size of the material; the conveying belt speed regulation signal can adjust the conveying belt speed in the range of 0.5-2m / s, and the adjustment accuracy is 0.01m / s.
[0024] Further, it comprises a multi-band imaging assembly, a data processing unit, an actuator and a human-computer interaction unit; the data processing unit adopts an industrial-grade CPU and GPU, and is used for running multi-spectral feature extraction, dynamic weight distribution, multi-modal attention mechanism and sorting decision algorithm; the human-computer interaction unit comprises a touch screen and a state indicating lamp, can display running parameters such as sorting accuracy and processing speed in real time, and supports manual adjustment of threshold parameters; the multi-band imaging assembly, the actuator and the human-computer interaction unit all communicate with the data processing unit through an industrial Ethernet, and the communication rate is not less than 100Mbps.
[0025] Compared with the known prior art, the technical scheme provided by the application has the following beneficial effects:
[0026] Through the dynamic weight distribution module, the feature weight is adjusted according to the real-time state of the material, so that the key features (such as texture features in complex scenes) can obtain higher weights, the key information is prevented from being submerged due to fixed weights, and the effectiveness of feature fusion is improved;
[0027] The multi-modal attention mechanism module can further strengthen the key information in unstable features, reduce the influence of overlapping, occlusion and shadow on feature quality, reduce the fluctuation range of the sorting accuracy of the color sorter in complex scenes to 1%-2%, and improve the stability of long-time operation;
[0028] The color sorter adopting the method can improve the sorting accuracy by 3%-5% compared with the traditional method, the processing speed can reach 1000-1500kg / h (taking grain sorting as an example), the sorting efficiency is significantly improved, and the material waste rate in industrial production is reduced. BRIEF DESCRIPTION OF DRAWINGS
[0029] Figure 1 The working system flowchart of the application. DETAILED DESCRIPTION
[0030] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments but not all of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.
[0031] The present application will be further described below with reference to the embodiments.
[0032] Embodiment one: rice sorting (grain material, solving the problem of sorting mildewed grains and stone impurities)
[0033] 1. Experimental background and target
[0034] Rice sorting is a key link in grain processing. The traditional method is easily affected by material overlap and shadow, resulting in low sorting accuracy of mildewed grains (yellowish color) and stone impurities (hard texture). The target of this embodiment is to achieve a sorting accuracy of qualified rice ≥ 99%, a mildewed grain recognition rate ≥ 98%, and a stone removal rate ≥ 99% under the conditions of a conveying belt speed of 1.5 m / s and a material thickness of 3-5 mm.
[0035] 2. Experimental conditions
[0036] 2.1 Hardware configuration
[0037] Multi-band imaging assembly: visible light camera (Hikvision MV-CA050-10GM, resolution 2448x2048, frame rate 25fps, lens focal length 16mm), near-infrared camera (Daheng image MER-132-43GM, resolution 1280x1024, frame rate 30fps, waveband 760-1100nm, lens focal length 12mm), two cameras are installed 50cm above the conveying belt, and the trigger signals are synchronized (synchronization error ≤0.5ms);
[0038] Data processing unit: industrial computer (CPU: Intel Core i7-13700E, main frequency 3.8GHz; GPU: NVIDIA RTX A5000, video memory 24GB; memory 32GB DDR5);
[0039] Actuator: air jet valve (Yadea 4V210-08, response time ≤3ms, air jet duration adjustable range 10-30ms), conveying belt (width 800mm, speed adjustment range 0.5-2m / s, speed adjustment accuracy 0.01m / s);
[0040] Human-computer interaction unit: 10.1-inch touch screen (1920×1200 resolution), supporting real-time display of sorting accuracy, processing speed, and jet valve status.
[0041] 2.2 Material and Parameter Presets
[0042] Material to be sorted: Mixed rice (95% qualified rice, 3% moldy rice, 2% stone impurities, total mass 100kg);
[0043] Threshold parameters: Low-state threshold E1 = 0.3, high-state threshold E2 = 0.6; State evaluation function coefficients α = 0.4, β = 0.3, γ = 0.3;
[0044] SVM model parameters: penalty coefficient C = 5, kernel function parameter γ = 0.5 (optimized through 5-fold cross-validation, training samples consist of 10,000 sets of rice feature data);
[0045] Attention network parameters: initial learning rate 0.001, 1500 iterations, training samples containing at least 5000 sets of rice feature data with different overlapping / occlusion states (key features are labeled as "color of moldy grains" and "texture of pebbles").
[0046] 3. Implementation Steps
[0047] 3.1 Multispectral Data Acquisition
[0048] The color sorter is started, the conveyor belt runs at a speed of 1.5m / s, and the multi-band imaging components simultaneously capture images:
[0049] A visible light camera captures the color information of rice (focusing on capturing the yellow hues of moldy grains);
[0050] Near-infrared cameras capture texture information of rice (with a focus on distinguishing the high-density texture of pebbles from the loose texture of rice);
[0051] Image data is transmitted to the data processing unit in real time via a USB 3.0 interface, with a transmission rate of ≥400MB / s and no data packet loss.
[0052] 3.2 Multi-feature extraction
[0053] Color feature extraction: Convert the visible light image to HSV space and extract the color feature vector C = [H_mean, S_var, V_grad] of a single grain of rice.
[0054] Qualified rice: H_mean = 30°±3° (light yellow tone), S_var = 0.08±0.01, V_grad = 0.04±0.005;
[0055] Moldy rice: H_mean=65°±5° (yellow tone), S_var=0.18±0.02, V_grad=0.09±0.01;
[0056] Texture feature extraction: Convert the near-infrared image to grayscale, and extract the texture feature vector T = [Contrast_avg, Correlation_avg, Energy_avg, Homogeneity_avg] using GLCM (distance 1 pixel, angle 0° / 45° / 90° / 135°).
[0057] Qualified rice: Contrast_avg=25±3, Correlation_avg=0.85±0.02;
[0058] Stone impurities: Contrast_avg=85±5, Correlation_avg=0.6±0.03.
[0059] 3.3 Dynamic Feature Fusion
[0060] State parameter calculation: The semantic segmentation algorithm identifies the material area and calculates the following: rice overlap O = 0.2 (slight overlap), occlusion area ratio C = 0.1 (no obvious occlusion), and shadow intensity S = 0.05 (stable ambient light).
[0061] Dynamic weight calculation: State evaluation value E = 0.4 × 0.2 + 0.3 × 0.1 + 0.3 × 0.05 = 0.125 < E1 = 0.3, Wc = 0.6 - 0.2 × 0.125 = 0.575, Wt = 0.4 + 0.2 × 0.125 = 0.425;
[0062] Attention mechanism enhancement: The self-attention network adjusts the weighted feature vector, focusing on increasing the weight of the moldy particle “[H_mean]” dimension by 1.3 times and the stone “[Contrast_avg]” dimension by 1.5 times, and outputs the fused feature vector F.
[0063] 3.4 Sorting Decision and Execution
[0064] The sorting decision module inputs F into the improved SVM model, and the classification results are: qualified rice 9520g, moldy rice 295g, and stones 185g.
[0065] Generate control signal: Send an "open 20ms" signal to the jet valve corresponding to the moldy particles and stones, and the conveyor belt maintains a speed of 1.5m / s;
[0066] Actuator action: The jet valve precisely blows away moldy grains and stones, and qualified rice falls into the finished product warehouse.
[0067] 4. Effect Verification
[0068] Sorting accuracy: 99.2% accuracy rate for qualified rice (0.8% false positive rate), 98.3% identification rate for moldy grains, and 99.1% removal rate for stones;
[0069] Processing efficiency: The actual processing speed is 1250 kg / h, which meets the industrial production demand of 1000-1500 kg / h;
[0070] Stability: After 8 hours of continuous operation, the sorting accuracy fluctuation is ≤1.2%, which is far lower than the 5%-10% of traditional methods.
[0071] Example 2: Plastic Sorting (Recycled Plastic Materials, Solving the Problem of Sorting Mixed PP and PE Plastics)
[0072] 1. Experimental Background and Objectives
[0073] Recycled plastic sorting requires distinguishing between polypropylene (PP) and polyethylene (PE). While their colors are similar, their textures differ, and traditional methods are prone to causing discrepancies due to light reflection. The objective of this embodiment is to achieve a PP / PE sorting accuracy of ≥97% and a processing speed of ≥800 kg / h, under conditions of a material crushing particle size of 5-10 mm and a conveyor belt speed of 1.2 m / s.
[0074] 2. Experimental conditions
[0075] 2.1 Hardware Configuration
[0076] Multi-band imaging components: visible light camera (Baslerac A2500-14gm, resolution 2592×1944, frame rate 14fps), near-infrared camera (FLIRA655sc, resolution 640×512, frame rate 30fps, band 800-1000nm).
[0077] Data processing unit: Industrial computer (CPU: AMD Ryzen 7 7840X, 4.2GHz; GPU: AMD Radeon Pro V620, 16GB video memory);
[0078] Actuators: Jet valve (SMCSY3120-5LZD-M5, response time ≤2.5ms), conveyor belt (width 600mm, speed 1.2m / s).
[0079] 2.2 Material and Parameter Presets
[0080] Material to be sorted: Mixed crushed plastic (60% PP, 40% PE, total mass 50kg, all white);
[0081] Threshold parameters: E1 = 0.25, E2 = 0.55; α = 0.35, β = 0.35, γ = 0.3;
[0082] SVM model parameters: C=8, γ=0.3 (8000 sets of PP / PE feature data for training samples).
[0083] 3. Implementation Steps
[0084] 3.1 Multispectral Data Acquisition
[0085] Visible light cameras capture light reflected from plastic surfaces (PP and PE are both white, and their H_mean is similar, making them difficult to distinguish);
[0086] Near-infrared cameras capture the internal texture of plastics (PP has a semi-crystalline structure and high texture contrast; PE has an amorphous structure and uniform texture).
[0087] 3.2 Multi-feature extraction
[0088] Color feature vector C: H_mean for both PP and PE is 20°±2°, and S_var is 0.1±0.01 and 0.12±0.01 respectively (small difference);
[0089] Texture feature vector T: Contrast_avg = 45±4 for PP, Contrast_avg = 20±3 for PE (significant difference).
[0090] 3.3 Dynamic Feature Fusion
[0091] State parameters: overlap of plastic fragments O = 0.3 (medium overlap), occlusion C = 0.2, shadow S = 0.1 (slight shadow due to workshop lighting);
[0092] Weight calculation: E = 0.35 × 0.3 + 0.35 × 0.2 + 0.3 × 0.1 = 0.195 < E1 = 0.25, Wc = 0.6 - 0.2 × 0.195 = 0.561, Wt = 0.4 + 0.2 × 0.195 = 0.439;
[0093] Attention Enhancement: Focus on increasing the weight of the "[Contrast_avg]" dimension (1.6 times) to amplify the texture differences between PP and PE.
[0094] 3.4 Sorting Decision and Execution
[0095] SVM model classification: PP 30.2kg, PE 19.8kg, no crossover misclassification;
[0096] The jet valve opens in 15ms, accurately separating PP and PE.
[0097] The processing speed is 820 kg / h, which meets the target requirements.
[0098] 4. Effect Verification
[0099] Sorting accuracy: PP accuracy 97.3%, PE accuracy 97.5%;
[0100] Anti-interference performance: After adjusting the workshop lighting intensity (±20%), the sorting accuracy fluctuation is ≤1.5%;
[0101] Economic efficiency: Compared with manual sorting (efficiency 200kg / h), the efficiency is increased by 4 times and the labor cost is reduced by 80%.
[0102] Example 3: Ore sorting (mineral materials, solving the problem of sorting mixed iron ore and waste rock)
[0103] 1. Experimental Background and Objectives
[0104] Iron ore sorting requires distinguishing between magnetic iron ore (Fe3O4) and waste rock (quartz, feldspar). Both are dark gray in color, but their near-infrared absorption characteristics differ. Traditional methods are prone to characteristic distortion due to the oxide layer on the ore surface. The goal of this embodiment is to achieve an iron ore recovery rate of ≥95% and a waste rock rejection rate of ≥90% under the conditions of ore particle size of 20-30mm and conveyor belt speed of 0.8m / s.
[0105] 2. Experimental conditions
[0106] 2.1 Hardware Configuration
[0107] Multi-band imaging components: visible light camera (Baumer BA640-120gc, resolution 640×480, frame rate 120fps), near-infrared camera (XenicsXeva-2.5-320, resolution 320×256, frame rate 60fps, band 900-1700nm);
[0108] Data processing unit: Industrial computer (CPU: Intel Xeon W-1370, 3.6GHz; GPU: NVIDIA Quadro P2200, 5GB video memory);
[0109] Actuators: Pneumatic actuator (FestoDGPL-32-100-PPVA-B, response time ≤5ms), conveyor belt (width 1000mm, speed 0.8m / s).
[0110] 2.2 Material and Parameter Presets
[0111] Material to be sorted: Mixed ore (40% iron ore, 60% waste rock, total mass 200 kg, with an oxide layer on the surface);
[0112] Threshold parameters: E1 = 0.4, E2 = 0.7; α = 0.5, β = 0.3, γ = 0.2 (the impact of ore overlap is greater, so the weight of α is increased);
[0113] SVM model parameters: C = 10, γ = 0.8 (training samples: 12,000 sets of ore feature data).
[0114] 3. Implementation Steps
[0115] 3.1 Multispectral Data Acquisition
[0116] Visible light camera: Iron ore and waste rock are both dark gray in color, H_mean=180°±5°, making them difficult to distinguish;
[0117] Near-infrared camera: Iron ore has a strong absorption peak in the 1200nm band, while waste rock has no absorption peak, and the texture characteristics are significantly different.
[0118] 3.2 Multi-feature extraction
[0119] Color characteristics: The difference between S_var and V_grad of the two is <5%, which has no distinguishing value;
[0120] Texture characteristics: Energy_avg of iron ore = 0.15 ± 0.02, and Energy_avg of waste rock = 0.08 ± 0.01 (significant difference).
[0121] 3.3 Dynamic Feature Fusion
[0122] State parameters: Ore overlap O = 0.6 (severe overlap), occlusion C = 0.4, shadow S = 0.15;
[0123] Weight calculation: E = 0.5 × 0.6 + 0.3 × 0.4 + 0.2 × 0.15 = 0.45 (E1 = 0.4 ≤ 0.45 ≤ E2 = 0.7), Wc = 0.4 - 0.4 × (0.45 - 0.4) / (0.7 - 0.4) = 0.333, Wt = 0.6 + 0.4 × (0.45 - 0.4) / (0.7 - 0.4) = 0.667;
[0124] Attention Enhancement: Increase the weight of the "[Energy_avg]" dimension by 1.8 times to counteract the interference of the oxide layer on features.
[0125] 3.4 Sorting Decision and Execution
[0126] SVM model classification: Iron ore 81.2 kg (recovery rate 95.1%), waste rock 118.8 kg (rejection rate 90.5%);
[0127] The pneumatic push rod has an action time of 30ms, pushing the waste rock into the waste bin;
[0128] The processing speed is 650 kg / h, which meets the needs of mine production.
[0129] 4. Effect Verification
[0130] Key indicators: Iron ore recovery rate of 95.1% and waste rock removal rate of 90.5%, far exceeding the traditional magnetic separation method (recovery rate of 85%).
[0131] Stability: After 24 hours of continuous operation, the recovery rate fluctuation is ≤2%;
[0132] Environmental friendliness: Compared to chemical sorting methods, there is no wastewater discharge and energy consumption is reduced by 60%.
[0133] Summary of Implementation Examples
[0134] The three embodiments cover three typical materials: grains, plastics, and ores, verifying the universality of this method:
[0135] Dynamic weight allocation: The feature weights are adaptively adjusted according to the material status (overlap, occlusion, shadow) to solve the information loss problem of traditional fixed weights;
[0136] Multimodal attention mechanism: Strengthens key feature dimensions and improves feature discrimination in complex scenarios;
[0137] Industrial compatibility: The hardware configuration meets industrial-grade standards, and the processing speed and stability meet production requirements, making it widely applicable to various material sorting scenarios.
[0138] The core advantage of this method lies in "adaptive fusion," which breaks through the technical bottlenecks of traditional color sorters through algorithmic innovation, providing a practical solution for high-precision and high-efficiency sorting.
[0139] The above embodiments are only used to illustrate 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of the present invention.
Claims
1. A high-efficiency color sorter based on multi-spectral feature fusion, characterized in that, The method comprises the following steps: Multi-spectral data acquisition: through the multi-band imaging assembly of the color sorter, image data of the material to be sorted under at least two different wavebands is collected, and the wavebands include a visible light waveband and a near-infrared waveband; Multi-feature extraction: based on the multi-band image data collected in the multi-spectral data acquisition step, the color feature parameters of the material are extracted through a color feature extraction submodule, and the texture feature parameters of the material are extracted through a texture feature extraction submodule; Dynamic feature fusion: the color feature parameters and the texture feature parameters extracted in the multi-feature extraction step are input into a dynamic weight distribution module, the dynamic weight distribution module calculates the dynamic weight values of different features according to real-time state parameters of the material, and combines a multi-modal attention mechanism module to strengthen key features, and outputs a fusion feature vector; Sorting decision: the fusion feature vector output in the dynamic feature fusion step is input into a sorting decision module, the sorting decision module determines the category of the material through a preset classification model, generates a control signal of an execution mechanism of the color sorter, and completes the sorting of the material.
2. The high-efficiency sorting method of a color sorter based on multi-spectral feature fusion according to claim 1, characterized in that, The multi-band imaging assembly in step 1 includes a visible light camera and a near-infrared camera, the resolution of the visible light camera is not less than 1920*1080, the frame rate is not less than 30fps, the resolution of the near-infrared camera is not less than 1280*720, the frame rate is not less than 30fps, and the shooting trigger signals of the two cameras are synchronized, and the synchronization error is less than 1ms.
3. The high-efficiency sorting method of a color sorter based on multi-spectral feature fusion according to claim 1, characterized in that, In step 2, the color feature extraction submodule extracts color feature parameters using an HSV color space model, specifically including hue mean, saturation variance and lightness gradient, wherein the value range of hue is 0°-360°, and the value range of saturation and lightness is 0-1; the texture feature extraction submodule extracts texture feature parameters using a gray level co-occurrence matrix, specifically including contrast, correlation, energy and uniformity, and the calculation distance of the gray level co-occurrence matrix is 1 pixel and the gray level is 256 levels.
4. The high-efficiency sorting method of a color sorter based on multi-spectral feature fusion according to claim 1, characterized in that, In step 3, the material real-time state parameter is obtained by: identifying the material area through an image segmentation algorithm based on semantic segmentation, calculating the material overlap degree O and the occlusion area ratio C; calculating the shadow intensity S through a shadow detection algorithm based on double-band gray difference, wherein the gray difference is the difference between the near-infrared waveband image gray value and the visible light waveband image gray value.
5. The high-efficiency sorting method of a color sorter based on multispectral feature fusion according to claim 1, characterized in that, The weight calculation process of the dynamic weight distribution module in step 3 includes: Constructing a state evaluation function: taking the material overlap degree, the occlusion area ratio and the shadow intensity as inputs, calculating the state evaluation value E, the formula is E=α×O+β×C+γ×S, wherein α=0.4, β=0.3, γ=0.3, and α+β+γ=1; Dynamic weight mapping: preset low state threshold E1=0.3, high state threshold E2=0.6, color feature weight Wc and texture feature weight Wt are obtained by segment function mapping according to state evaluation value E, when E<0.3, Wc=0.6-0.2×E, Wt=0.4+0.2×E; when 0.3≤E≤0.6, Wc=0.4-0.4×(E-0.3) / 0.3, Wt=0.6+0.4×(E-0.3) / 0.3; when E>0.6, Wc=0.2, Wt=0.
8.
6. The high-efficiency sorting method of a color sorter based on multispectral feature fusion according to claim 1, characterized in that, The multi-modal attention mechanism module in step 3 includes a self-attention network and a feature adjustment unit. The input of the self-attention network is the dynamically weighted color feature vector and texture feature vector. The network structure includes 2 fully connected layers and 1 activation layer, and the attention weight matrix is obtained by training. The feature adjustment unit adjusts the weight of each dimension of the feature vector element by element according to the attention weight matrix, and the adjustment formula is F_i=F_i×A_i.
7. The high-efficiency sorting method of a color sorter based on multispectral feature fusion according to claim 6, characterized in that, The training sample of the self-attention network includes at least 5000 groups of multi-spectral feature data of materials in different states. Each sample contains color feature vector, texture feature vector, material real-time state parameter and artificial labeled key feature label. The training process uses Adam optimizer, the initial value of learning rate is set to 0.001, and the learning rate is attenuated to 0.8 times of the previous one every 100 epochs. The number of training iterations is not less than 1000 times.
8. The high-efficiency sorting method of a color sorter based on multispectral feature fusion according to claim 1, characterized in that, The classification model of the sorting decision module in step 4 uses an improved support vector machine model. The kernel function of the model is radial basis function, and the kernel function formula is K(x, y)=exp(-γ||x-y||2), where γ is the kernel function parameter. The model parameters are optimized by 5-fold cross-validation method, and the optimization goal is to maximize the classification accuracy. The final preferred parameter range is: the penalty coefficient C value range is 1-10, and the kernel function parameter γ value range is 0.1-1.
9. The high-efficiency sorting method of a color sorter based on multispectral feature fusion according to claim 1, characterized in that, The control signal of the execution mechanism in step 4 includes air jet valve control signal and conveyor belt speed signal. The response delay of the air jet valve control signal is less than 5ms, and the air jet time can be adaptively adjusted in the range of 10-50ms according to the size of the material. The conveyor belt speed signal can adjust the speed of the conveyor belt in the range of 0.5-2m / s, and the adjustment accuracy is 0.01m / s.
10. A color sorter system implementing the method of any one of claims 1-9, characterized by It includes a multi-band imaging assembly, a data processing unit, an execution mechanism, and a human-computer interaction unit. The data processing unit uses an industrial-grade CPU and GPU to run multi-spectral feature extraction, dynamic weight distribution, multi-modal attention mechanism, and sorting decision algorithm. The human-computer interaction unit includes a touch screen and a status indicator light, which can display sorting accuracy, processing speed and other operating parameters in real time, and support manual adjustment of threshold parameters. The multi-band imaging assembly, execution mechanism, and human-computer interaction unit are communicated with the data processing unit through industrial Ethernet, and the communication rate is not less than 100Mbps.