Bimodal intelligent cotton foreign fiber sorting system and method

Abstract

The invention discloses a bimodal fusion intelligent cotton foreign fiber sorting system and method. The sorting system comprises a shell, an image collecting module, an image processing module, a conveying module and a removing module. The system specifically comprises an LED light source, a polarization camera, an RGB camera, an industrial personal computer, a camera support, a cotton flow channel, a spray valve control panel and a high-speed spray valve. Cotton flow enters the cotton flow channel through the cotton inlet pipeline, image data are transmitted to the industrial personal computer through the image acquisition module, a detection result is quickly obtained through algorithm processing of the image processing module, and the spray valve control panel drives the high-speed spray valve to remove foreign fibers according to the detection result. According to the invention, real-time detection and elimination of foreign fibers in cotton can be realized, and the detection speed and precision are greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of cotton processing, foreign fiber identification and removal, and in particular to a dual-mode intelligent cotton foreign fiber sorting system and method. Background Art

[0002] Cotton is an important raw material in my country's textile industry. In the process of cotton collection, foreign fibers such as polypropylene yarn, ground film, ground irrigation tape, hair, etc. will inevitably be mixed in. These foreign fibers are mixed in cotton and will be broken into countless small fiber defects during the processing of cotton. These defects are not only difficult to remove during the textile processing, but also easily broken or combed into shorter and thinner fibers during the impurity removal process, forming a large number of small fiber defects. When spinning, these small defects are easy to cause the yarn to break, thereby reducing production efficiency; when weaving, it will affect the surface quality of the cloth; when dyeing, it may cause uneven dyeing of fibers and cloth, affecting the appearance of the finished product.

[0003] Traditional cotton foreign fiber detection methods mostly rely on manual visual inspection and simple physical sorting, which have problems such as slow detection speed, low efficiency, and poor accuracy. It is also difficult to cope with the complex types of foreign fibers and the diversity of cotton quality. In recent years, with the development of image processing and deep learning technology, foreign fiber detection methods based on image recognition and intelligent algorithms have gradually become a research hotspot. However, the existing technology still has certain challenges in the accuracy and real-time performance of foreign fiber identification. Therefore, a cotton foreign fiber detection and sorting system that can combine image processing and deep learning technology is needed to improve detection accuracy and processing efficiency.

[0004] Chinese patent document CN115201219A discloses a cotton foreign fiber sorting machine, which aims to solve the problems of low efficiency, high labor intensity and unsatisfactory effect of traditional manual foreign fiber picking through automation technology. The sorting machine includes a device body, a conveyor belt mechanism, a dust blowing mechanism, a dust suction mechanism, a winding mechanism and a controller. The conveyor belt is used to transmit cotton at a constant speed. The dust blowing mechanism and the dust suction mechanism cooperate with the foreign matter detected by the CCD camera to automatically blow up and suck in foreign matter to ensure that foreign matter is accurately collected and avoid scattering. The winding mechanism organizes the cotton and stores it in the collection box after sorting. When the above scheme processes a large amount of cotton, the sorting speed may become a bottleneck, and the processing of different types of foreign matter (such as light ground film and heavier drip irrigation tape) may not be flexible enough. Summary of the invention

[0005] The purpose of the present invention is to provide a cotton foreign fiber sorting system and a sorting method thereof by fusing RGB and polarization image information in view of the deficiencies in the prior art.

[0006] Technical solution: The technical solution adopted by the present invention to solve the problem is: a dual-modal fusion intelligent cotton foreign fiber sorting system, including a shell, an image acquisition module, an image processing module, a transmission module, and a rejection module; the image acquisition module includes: a first RGB camera fixed to a first camera bracket, a second RGB camera fixed to a second camera bracket, a first polarization camera fixed to a third camera bracket, and a second polarization camera fixed to a fourth camera bracket; the transmission module includes a cotton flow channel, and a plurality of LED light sources are alternately installed on both sides; the image information processing module includes an industrial computer installed inside the box; the rejection module includes a spray valve control panel located on the side of the cotton flow channel, which is wired to the industrial computer through electrical signals, and is respectively connected to the gas tank and the high-speed spray valve through air pipes.

[0007] Preferably, the shell is made of opaque aluminum alloy and the interior is painted with high-absorption black paint. The LED light source includes a white light source and an ultraviolet light source, which are respectively fixed on both sides of the cotton flow channel. With each camera field of view as the center, light sources are installed up and down on the same side of each camera field of view. The angle of the light source is installed 45 degrees clockwise or counterclockwise according to the direction of the cotton flow transport pipeline.

[0008] The present invention also provides a dual-mode fusion intelligent cotton foreign fiber sorting method, comprising the following steps:

[0009] S1: read the RGB image and polarization image into the image processing module on the industrial computer;

[0010] S2: Reconstruct and fuse the RGB image and the polarization image using an autoencoder to generate a bimodal fusion image;

[0011] S3: Use deep learning detection algorithm to detect foreign fibers in dual-modal fusion images, and use brilliant color detection algorithm to detect foreign fibers in RGB images simultaneously;

[0012] S4: The spray valve control board combines the dual-modal fusion image and the RGB image detection results, and controls the high-speed spray valve to open when the foreign fiber reaches the rejection position to remove the foreign fiber.

[0013] Preferably, the step of reconstructing and fusing the RGB image and the polarization image by using an autoencoder to generate a dual-modal fusion image in S2 comprises: selecting a variational autoencoder model including a multi-branch attention mechanism, and performing a multi-branch attention on the input RGB image I rgb and polarization image I pol Encoded as latent vectors Z rgb and Z pol , and the loss function shown in the following formula takes into account both the reconstruction accuracy and the potential space distribution constraints during the model training process, so as to obtain a fused image that retains the key features of visible light and polarization information and suppresses noise:

[0014]

[0015] In the formula, is the reconstructed output image after fusion, represents the reconstruction error (i is the sample number), D KL is the relative entropy, which is used to measure the difference between the potential distribution q(z|I) obtained by model encoding and the prior distribution p(z), so as to make the potential vector distribution more stable and improve the generalization ability of the model. α and β are the weighting coefficients of the loss function. By achieving a balance between reconstruction accuracy and distribution constraints, the robustness of bimodal feature fusion and the ability to capture foreign fiber details can be effectively enhanced.

[0016] Preferably, the method of using a deep learning detection algorithm to detect foreign fibers in the dual-modal fusion image and simultaneously using a brilliant color detection algorithm to detect foreign fibers in the RGB image in S3 includes: using the global characteristics of the dual-modal fusion image and the prior characteristics of the brilliant color detection algorithm to design a guided deep instance segmentation network, wherein the network first generates a foreign fiber saliency map M using the brilliant color detection algorithm. color The feature extraction backbone network of the deep learning detection algorithm is embedded, and a saliency weight adjustment mechanism is added to each layer of the feature map. The feature weight is calculated by the following formula:

[0017] F guided =F fusion ⊙Upsample(M color ),

[0018] In the formula, F guided is the target layer feature weight, F fusion is the fusion feature map, ⊙ represents the pixel weighted operation, and Upsample() is the multi-resolution interpolation of the saliency map.

[0019] Preferably, the spray valve control board in S4 combines the detection results of the dual-modal fusion image and the RGB image, and controls the high-speed spray valve to open when the foreign fiber reaches the rejection position to reject the foreign fiber, which specifically includes the following steps:

[0020] S41: Use multi-frame image capture and target similarity analysis to predict the movement speed of defective targets, including continuously shooting all objects in the cotton flow channel with a high-speed camera, recording the image frame number and time stamp, and ensuring that each impurity target is photographed at least twice;

[0021] S42: feature matching is performed on the impurity targets in two consecutive frames of images, and the similarity S is calculated using the shape contour, color information and feature point distribution of the impurity targets, where the similarity calculation formula is:

[0022]

[0023] Among them, f 1,i and f 2,i Respectively represent the i-th impurity target feature of the target in frame 1 and frame 2, Sim() is the feature similarity measurement function, and N is the total number of feature points; when the similarity S exceeds the set threshold, the impurity targets in the two frames are determined to be the same impurity target. Subsequently, the position change of the impurity target in the two frames is used to calculate its motion speed vector

[0024] S43: After determining that the target is the same defective object, the displacement and rotation characteristics of the target during the movement are further combined to improve the rejection accuracy by optimizing the spray valve control signal. By analyzing the position and contour features of the target in two consecutive frames of images, the center offset of the target ΔC = (x2-x1, y2-y1) and the shape rotation angle change θ = arctan ((y2-y1) / (x2-x1)) are calculated. Based on the displacement and rotation information, the movement path of the target is corrected, and the simplified corrected velocity vector is:

[0025]

[0026] In the formula, It is the corrected velocity vector, and dynamically adjusts the triggering time and angle of the spray valve action.

[0027] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0028] (1) The present invention combines the data acquisition technology of RGB camera and polarization camera, uses autoencoder to reconstruct and fuse images, generates dual-modal fusion images, and combines deep learning with bright color detection algorithm to achieve high-precision detection of foreign fibers in cotton. Polarization images effectively reduce the interference of ambient light sources, especially for objects with high reflected light intensity (such as plastic films and artificial fibers), making the detection system more robust and adaptable under complex lighting conditions.

[0029] (2) The present invention adopts multi-frame image capture technology, predicts the movement speed through impurity target similarity analysis, and calculates the target's corrected speed vector in combination with the offset and rotation characteristics to achieve dynamic spray valve control. This method effectively solves the problem of rejection deviation caused by the difference in flying speed of cotton and foreign fibers in the channel and the complexity of the movement trajectory, greatly improves the accuracy of the rejection action, and meets the needs of large-scale high-speed sorting.

[0030] (3) Through the collaborative work of the industrial computer and FPGA, the present invention significantly enhances the ability to process large-scale image data, while ensuring the high stability of the system during long-term operation. With the high-quality light source layout and shell design, it reduces noise interference and image deviation, is suitable for industrial application scenarios of large-scale cotton processing, and greatly improves production efficiency and sorting reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a schematic diagram of the structure of the present invention;

[0032] Figure 2 The figure is a flow chart of the foreign fiber sorting method. DETAILED DESCRIPTION

[0033] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. The embodiments are implemented based on the technical solution of the present invention. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0034] like Figure 1-2As shown, a cotton foreign fiber sorting system comprises a housing 1, an image acquisition module, an image processing module, a transmission module, and a rejection module; a first LED light source 51, a second LED light source 52, a third LED light source 53, a fourth LED light source 54, a fifth LED light source 55, a sixth LED light source 56, a seventh LED light source 57, and an eighth LED light source 58 are installed on both sides of the cotton flow channel inside the housing 1; the image acquisition module and the transmission module are located inside the housing 1, and the image processing module is installed on the side of the housing; the image acquisition module comprises: an RGB camera 31 fixed to a first camera bracket, an RGB camera 31 fixed to a second camera bracket camera 32, a polarization camera 33 fixed on a third camera bracket, and a polarization camera 34 fixed on a fourth camera bracket; the transmission module includes a cotton flow channel 4, a waste cotton pipe 10, and a cotton outlet pipe 11; the image information processing module includes an industrial computer 6 and a spray valve control board 7, and the industrial computer is connected to the first RGB camera 31, the second RGB camera 32, the first polarization camera 33, the second polarization camera 34, the first LED light source 51, the second LED light source 52, the third LED light source 53, the fourth LED light source 54, the fifth LED light source 55, the sixth LED light source 56, the seventh LED light source 57, and the eighth LED light source 58. The shell 1 is made of opaque aluminum alloy, and the interior is painted with high-absorption black paint. The first LED light source 51, the second LED light source 52, the third LED light source 53, the fourth LED light source 54, the fifth LED light source 55, the sixth LED light source 56, the seventh LED light source 57, and the eighth LED light source 58 are installed in an alternating order up and down on both sides of the internal cotton flow channel. The first camera bracket 21 is located in the middle of the first LED light source 51 and the second LED light source 52 in the vertical direction, the second camera bracket 22 is located in the middle of the fourth LED light source 54 and the fifth LED light source 55 in the vertical direction, and the third camera bracket 23 is located in the middle of the third LED light source and the fourth LED light source 54 in the vertical direction; the fourth camera bracket 24 is located in the seventh LED light source 57 In the middle of the vertical direction with the eighth LED light source 58; the first LED light source 51 and the second LED light source 52 are installed with the line scanning field of view of the first RGB camera 31 as the center, and are installed at 45 degrees in opposite directions perpendicular to the direction of the cotton flow channel 4; the third LED light source 53 and the fourth LED light source 54 are installed with the line scanning field of view of the second RGB camera 32 as the center, and are installed at 45 degrees in opposite directions perpendicular to the direction of the cotton flow channel 4; the fifth LED light source 55 and the sixth LED light source 56 are installed with the line scanning field of view of the first polarization camera 33 as the center, and are installed at 45 degrees in opposite directions perpendicular to the direction of the cotton flow channel 4; the seventh LED light source 57 and the eighth LED light source 58 are installed with the line scanning field of view of the second polarization camera 34 as the center, and are installed at 45 degrees in opposite directions perpendicular to the direction of the cotton flow channel 4

[0035] The shell 1, the first camera bracket 21, the second camera bracket 22, the third camera bracket 23, and the fourth camera bracket 24 are all independent of the cotton flow channel 4, thereby preventing the vibration generated by the fan during the transportation of the cotton flow from causing vibration interference to the first RGB camera 31, the second RGB camera 32, the first polarization camera 33, and the second polarization camera 34 installed on the camera bracket.

[0036] The shell 1 has an opening on the side to facilitate the installation, debugging and maintenance of the camera, and is made of aluminum as a whole, which is convenient for the internal camera equipment to dissipate heat while isolating external light and interference signals. The horizontal length of the cotton flow channel 4 is 1.3 meters, and the channel width is 20 centimeters. The frequency converter is used to control the speed of the cotton flow channel fan. The wind generated by the fan provides the movement speed for the cotton entering the cotton flow channel, and the movement speed of the cotton flow is adjusted. The overall closure of both sides of the channel can effectively reduce the interference of internal and external airflows, and has good wear resistance and corrosion resistance.

[0037] The cotton raw material is input from the top of the top cotton flow channel 4, and a square wave signal is sent to the spray valve control board 7 through the encoder at the motor shaft. The square wave signal can be used to track the change of the motor speed and thus track the movement of the object. The spray valve control board 7 divides the square wave signal and outputs the frequency that triggers the camera to shoot. The program written using marix-SDK enables two RGB cameras and two polarization cameras to simultaneously capture images of the cotton flow. Several LED light sources are installed on both sides of the cotton flow channel 4, with the camera line scan field of view as the center, perpendicular to the cotton flow transport direction and rotated 45 degrees in opposite directions. This can effectively supplement the light source, allowing several cameras to accurately shoot the cotton flow raw materials.

[0038] The first RGB camera 31, the second RGB camera 32, the first polarization camera 33, and the second polarization camera 34 all use line scanning industrial cameras. The four cameras are connected to the industrial computer through the CameraLink interface to ensure the transmission of image data. The industrial computer is equipped with an image acquisition card and a GPU. The acquisition card is responsible for reading image data, and the GPU processes a large amount of image information quickly to ensure that the detection algorithm in the industrial computer can run efficiently. The industrial computer and the spray valve control board are connected through a network cable to transmit the data of the image processing results. The spray valve control board is equipped with an onboard FPGA and an ARM core. The ARM core is used to communicate with the industrial computer to facilitate the acquisition of image processing results and control the operation of the corresponding FPGA module; the FPGA is responsible for dividing the encoder square wave signal and outputting a stable square wave frequency to trigger the camera and store different image processing results in FIFO by line, and merge different image processing results, thereby controlling the high-speed spray valve to remove foreign fibers in cotton.

[0039] like Figure 2As shown, an algorithm flow chart of a cotton foreign fiber sorting method; write a camera acquisition program, and use an industrial computer and a spray valve control board to simultaneously control the first RGB camera 31, the second RGB camera 32, the first polarization camera 33, and the second polarization camera 34 to stably acquire image data using an external trigger mode. The RGB camera is used to obtain cotton color image data. The polarization characteristic image of foreign fibers, especially transparent or reflective foreign fibers, is captured by the polarization camera.

[0040] For RGB data, the OpenCV library is used to preprocess the RGB image, including denoising and color enhancement operations, to optimize the image quality. Subsequently, the RGB image is converted into LAB color space and HSV color space respectively, and the information of different color channels is extracted to determine whether it meets the corresponding target color, and the saturation channel threshold is extracted to determine whether the saturation reaches the bright color standard. Through the preset threshold conditions, the bright color detection algorithm is applied to identify brightly colored foreign fibers, and the corresponding mask is generated as the result of the bright color detection algorithm and the foreign fiber significance map M is generated. color .

[0041] The RGB image and the polarization image are reconstructed and fused using an autoencoder. A variational autoencoder model with a multi-branch attention mechanism is used to encode the RGB image and the polarization image into latent vectors respectively. The loss function shown in the following formula takes into account both the reconstruction accuracy and the latent space distribution constraints during the model training process, thereby obtaining a fused image that retains the key features of visible light and polarization information and suppresses noise:

[0042]

[0043] In the formula, is the reconstructed output image after fusion, represents the reconstruction error (i is the sample number), D KL is the relative entropy, which is used to measure the difference between the potential distribution q(z|I) obtained by model encoding and the prior distribution p(z), so as to make the potential vector distribution more stable and improve the generalization ability of the model. α and β are the weighting coefficients of the loss function, which can enhance the robustness of the bimodal feature fusion and the ability to capture foreign fiber details by balancing the reconstruction accuracy and distribution constraints.

[0044] The label-studio annotation tool is used to annotate the preprocessed cotton images. The annotation types include polypropylene yarn, ground film, ground irrigation tape, hair and other foreign fibers.

[0045] The image containing polarization and RGB fusion is paired with the RGB image. The label-studio annotation tool is used to annotate the foreign fibers in the RGB image with rectangular frames. The types of foreign fibers include but are not limited to: polypropylene fibers, ground films, ground irrigation tapes, hair, etc., and a TXT annotation file corresponding to the image including the foreign fiber category and rectangular frame coordinates is obtained. The RGB image is paired with the image with fusion information, and the brilliant color detection algorithm is used to generate a foreign fiber saliency map M. color The feature extraction backbone network of the deep learning detection algorithm YOLO is embedded, and a saliency weight adjustment mechanism is added to each layer of the feature map. The feature weight calculation is shown in the following formula.

[0046] F guided =F fusion ⊙Upsample(M color )

[0047] In the formula, F guided is the target layer feature weight, F fusion is the fusion feature map, ⊙ represents the pixel weighted operation, and Upsample() is the multi-resolution interpolation of the saliency map.

[0048] Capture multiple frames of images of the cotton flow and analyze target similarity to estimate the movement speed of defective targets. Trigger each camera for continuous high-frequency shooting to ensure that each impurity target is photographed at least twice. The time interval between image shooting can be calculated based on the triggering frequency, and the position coordinates of the detected impurities in different images can be calculated to predict the movement trajectory of the object.

[0049] The similarity S of the impurity target is calculated using its shape contour, color information and feature point distribution, where the similarity calculation formula is:

[0050]

[0051] Among them, f 1,i and f 2,i Represents the i-th impurity target feature of the target in frame 1 and frame 2 respectively, Sim() is the feature similarity measurement function, and N is the total number of feature points; when the similarity S exceeds the set threshold, the impurity targets in the two frames are determined to be the same impurity target. The motion speed vector of the impurity target is calculated using the position change of the impurity target in the two frames.

[0052] After determining that the target is the same defective object, the offset and rotation characteristics of the target during the movement are further combined to improve the rejection accuracy by optimizing the spray valve control signal. By analyzing the position and contour features of the target in two consecutive frames of images, the center offset of the target ΔC = (x2-x1, y2-y1) and the shape rotation angle change θ = arctan ((y2-y1) / (x2-x1)) are calculated. Based on the offset and rotation information, the target's motion path is corrected, and the simplified corrected velocity vector is:

[0053]

[0054] In the formula, It is the corrected velocity vector, and dynamically adjusts the triggering time and angle of the spray valve action.

[0055] After calculating the motion trajectory of the impurity target, the processing result is generated as the corresponding mask. The area with 0 in the mask is normal cotton, and the area with 1 is the foreign fiber that needs to be removed. Before sending the results to the spray valve control board, the detection results of the images of the cameras on both sides need to be gradually merged. The RGB camera and polarization camera on the right are generally higher than the left. In order to ensure the integrity of the final data, the processing results of the images on both sides need to be merged and then sent to the lower computer. The complete detection results of the right camera are fused with the partial detection results of the left camera. The unmerged part will continue to be processed in the next round of fusion to ensure the integrity of the final data. The final result of the image mask after each round of merging is sent to the spray valve control board.

[0056] The spray valve control board stores the results of image detection in the FIFO queue and outputs it to the spray valve control system to control the opening and closing of the spray valve. The time when the foreign fiber reaches the spray valve nozzle is calculated based on the detection results, and the image delay time is accurately determined by considering the distance between the camera and the nozzle and the movement speed of the cotton flow. Then, this delay time is matched with the dequeue frequency of the FIFO queue to ensure that the spray valve can be accurately started when the foreign fiber reaches the rejection position. When the test result is "1", the spray valve opens to reject the foreign fiber; when the result is "0", the spray valve closes. Efficient detection and accurate rejection of foreign fibers in cotton are achieved, and the automation level and detection accuracy of cotton sorting are improved.

[0057] The above specific implementation is only a preferred embodiment of the present invention and is not intended to limit the implementation of the present invention and the scope of the claims. All equivalent changes and modifications made according to the content of the patent protection scope of the present invention should be included in the scope of the patent application of the present invention.

Claims

1. A dual-mode fusion intelligent cotton foreign fiber sorting system, characterized by: It comprises a housing (1), an image acquisition module, an image processing module, a transmission module, and a rejection module; The image acquisition module comprises: a first RGB camera (31) fixed to a first camera support (21), a second RGB camera (32) fixed to a second camera support (22), a first polarization camera (33) fixed to a third camera support (23), and a second polarization camera (34) fixed to a fourth camera support (24); the transmission module comprises a cotton flow channel (4), with a plurality of LED light sources alternately installed on both sides; the image information processing module comprises an industrial computer (6) installed inside a box body; the rejection module comprises a spray valve control panel (7) located on the side of the industrial computer (6), connected to the industrial computer (6) by wire through an electrical signal, and respectively connected to a gas tank (8) and a high-speed spray valve (9) through an air pipe.

2. The dual-mode fusion intelligent cotton foreign fiber sorting system according to claim 1 is characterized by: The shell (1) is made of an opaque aluminum alloy material, and the interior is painted with a high-absorption black paint. The LED light source includes a white light source and an ultraviolet light source, which are respectively fixed on both sides of the cotton flow channel (4). With each camera field of view as the center, the light source is installed up and down on the same side of each camera field of view, and the angle of the light source is rotated 45 degrees clockwise or counterclockwise according to the direction of the cotton flow transport pipeline.

3. A dual-mode fusion intelligent cotton foreign fiber sorting method, characterized by: The following steps are involved: S1: Read the RGB image and polarization image into the image processing module on the industrial computer; S2: Reconstruct and fuse the RGB image and the polarization image using an autoencoder to generate a dual-modal fusion image; S3: Using the deep learning detection algorithm to detect foreign fibers in the dual-modal fusion image, and simultaneously using the brilliant color detection algorithm to detect foreign fibers in the RGB image; S4: The spray valve control board combines the dual-modal fusion image and the RGB image detection results, and controls the high-speed spray valve to open when the foreign fiber reaches the rejection position to reject the foreign fiber.

4. A dual-mode intelligent cotton foreign fiber sorting method according to claim 3, characterized in that: In S2, the RGB image and the polarization image are reconstructed and fused by using an autoencoder to generate a dual-modal fusion image, which includes: selecting a variational autoencoder model including a multi-branch attention mechanism, and performing a multi-branch attention on the input RGB image I rgb and polarization image I pol Encoded as latent vectors Z rgb and Z pol , and the loss function shown in the following formula takes into account both the reconstruction accuracy and the potential space distribution constraints during the model training process, so as to obtain a fused image that retains the key features of visible light and polarization information and suppresses noise: In the formula, is the reconstructed output image after fusion; represents the reconstruction error (i is the sample number); D KL is the relative entropy, which is used to measure the difference between the potential distribution q(z|I) encoded by the model and the prior distribution p(z), so as to make the potential vector distribution more stable and improve the generalization ability of the model. α and β are the weighting coefficients of the loss function, which enhance the robustness of the bimodal feature fusion and the ability to capture foreign fiber details by striking a balance between reconstruction accuracy and distribution constraints.

5. A cotton foreign fiber sorting method according to claim 3, characterized in that: In S3, the foreign fiber detection in the dual-modal fusion image is detected by using a deep learning detection algorithm, and the foreign fiber detection in the RGB image is detected by using a brilliant color detection algorithm simultaneously, which includes: using the global characteristics of the dual-modal fusion image and the prior characteristics of the brilliant color detection algorithm, a guided deep instance segmentation network is designed, and the network first generates a foreign fiber saliency map M by using the brilliant color detection algorithm. color The feature extraction backbone network of the deep learning detection algorithm is embedded, and a saliency weight adjustment mechanism is added to each layer of the feature map. The feature weight is calculated by the following formula: F guided =F fusion ⊙Upsample(M color ), In the formula, F guided is the target layer feature weight, F fusion is the fusion feature map, ⊙ represents the pixel weighted operation, and Upsample() is the multi-resolution interpolation of the saliency map.

6. A dual-mode intelligent cotton foreign fiber sorting method according to claim 3, characterized in that: The spray valve control board in S4 merges the dual-modal fusion image and the RGB image detection results, and controls the high-speed spray valve to open when the foreign fiber reaches the rejection position to remove the foreign fiber, including: using multi-frame image capture and target similarity analysis to predict the movement speed of the defective target, including continuously shooting all objects in the cotton flow channel with a high-speed camera, recording the image frame number and time stamp, and ensuring that each impurity target is shot at least twice; then, feature matching is performed on the impurity targets in two consecutive frames of images, and the shape contour, color information and feature point distribution of the impurity target are used to calculate its similarity S, where the similarity calculation formula is: Among them, f 1,i and f 2,i Respectively represent the i-th impurity target feature of the target in frame 1 and frame 2, Sim() is the feature similarity measurement function, and N is the total number of feature points; when the similarity S exceeds the set threshold, the impurity targets in the two frames are determined to be the same impurity target, and then the position change of the impurity target in the two frames is used to calculate its motion speed vector 7. A dual-mode intelligent cotton foreign fiber sorting method according to claim 6, characterized in that: In S4, the position change of the impurity target in the two frames of images is used to calculate its movement speed v. After determining that the target is the same defective object, the displacement and rotation characteristics of the target during the movement are further combined to improve the rejection accuracy by optimizing the spray valve control signal. By analyzing the position and contour features of the target in two consecutive frames of images, the center offset ΔC = (x2-x1, y2-y1) and the shape rotation angle change θ = arctan ((y2-y1) / (x2-x1)) of the target are calculated. Based on the offset and rotation information, the movement path of the target is corrected. The simplified corrected speed vector is: In the formula, It is the corrected velocity vector, and dynamically adjusts the triggering time and position of the spray valve action.

Images