A bamboo joint recognition device and system based on AI visual detection

Abstract

This invention relates to the field of image processing and measurement technology, specifically disclosing a bamboo joint recognition device and system based on AI visual detection. The device includes an outer frame that docks with a bamboo drawing machine. One side of the outer frame has a bamboo strip inlet flush with the output port of the bamboo drawing machine. A driven wheel is slidably mounted on the inner wall of the outer frame near the bamboo drawing machine. An active feeder is mounted on the side of the outer frame away from the bamboo drawing machine. An encoder connected to the outer surface of the outer frame is located below the driven wheel. This invention establishes a sub-pixel-level synchronous mapping relationship between image frames and displacement sequences during bamboo strip transportation through a hybrid mechanism of encoder hardware triggering and image optical flow dynamic compensation. This effectively eliminates the problem of asynchronous movement caused by factors such as mechanical vibration, transportation speed fluctuations, and installation errors, ensuring that the pixel position of the bamboo joint in the image accurately corresponds to its actual physical coordinates, significantly improving the positioning accuracy of bamboo joint detection.

Description

Technical Field

[0001] This invention relates to the field of image processing and measurement technology, and in particular to a bamboo joint recognition device and system based on AI visual detection. Background Technology

[0002] As a core raw material for the "bamboo-for-plastic" strategy, bamboo has characteristics such as protruding nodes, dense fibers, uneven mechanical properties, and susceptibility to cracking and mold, which significantly increase the processing difficulty and cost. At the same time, it affects the quality stability of bamboo products. Traditional manual inspection methods are inefficient and have poor accuracy, making it difficult to meet the demand for accurate identification of bamboo node positions in large-scale production. With the continuous expansion of bamboo application fields, higher requirements are being placed on the processing efficiency, material uniformity, and surface quality of node-free bamboo products.

[0003] In existing technologies, there is a problem that the image acquisition and encoder distance signal are difficult to synchronize during the movement of bamboo strips, which leads to a deviation in the correspondence between bamboo joint position and distance information, affecting the subsequent positioning accuracy. Moreover, the repetitive texture of the bamboo strip surface and the changes in lighting cause feature matching mismatch when stitching multiple frames of images, further interfering with the accurate identification and tracking of bamboo joints in continuous image sequences. At the same time, due to the diverse shapes and blurred edges of bamboo joints, it is difficult to achieve high-precision measurement and classification of bamboo joint spacing, which restricts the processing and adaptation capabilities of bamboo joint recognition results. To address these issues, we propose a bamboo joint recognition device and system based on AI visual detection. Summary of the Invention

[0004] To overcome the shortcomings of the prior art, the present invention provides a bamboo joint recognition device and system based on AI visual detection, which can effectively solve the problems involved in the prior art.

[0005] The objective of this invention can be achieved through the following technical solution: Firstly, this invention provides a bamboo joint recognition device based on AI visual detection, comprising an outer frame that docks with a bamboo drawing machine. One side of the outer frame has a bamboo strip inlet flush with the output port of the bamboo drawing machine, ensuring a smooth transition of the bamboo strip from the drawing machine into the recognition system. A driven wheel is slidably installed on the inner wall of the outer frame near the bamboo drawing machine to achieve adaptive guidance when the bamboo strip enters. An active feeder is installed on the side of the outer frame away from the bamboo drawing machine to provide active traction for bamboo strip transport. An encoder connected to the outer surface of the outer frame is located below the driven wheel to collect bamboo strip displacement signals in real time. A camera is fixedly installed on the top of the outer frame to achieve continuous imaging of the bamboo strip surface. A photoelectric sensor located above the driven wheel is opened on the top of the outer frame to detect the bamboo strip's position and trigger the coordinated operation of the encoder and camera. A light source fixed to the top of the outer frame is located on one side of the camera, and the light source is tilted to provide uniform and stable supplementary lighting conditions.

[0006] Preferably, the active feeder includes a drive motor fixedly installed at the bottom of the outer frame to ensure stable and reliable feeding power. The output end of the drive motor is connected to a first feeding roller via belt drive to achieve smooth power transmission. A second feeding roller is arranged above the first feeding roller to form a roller clamping structure, ensuring the stability of the bamboo strip's posture during the conveying process. The two ends of the first feeding roller rotate with the inner wall of the outer frame to provide stable rotational support. The two ends of the second feeding roller slide with the inner wall of the outer frame to adapt to the passage requirements of bamboo strips of different thicknesses.

[0007] Secondly, an AI-based visual detection-based bamboo joint recognition system, mounted on the aforementioned bamboo joint recognition device, includes the following modules: The data acquisition module is used to acquire data by using an encoder and a camera to obtain the distance information of the bamboo strip movement and image data containing bamboo node information. It also uses image optical flow method to dynamically compensate for motion offset, realizes sub-pixel level synchronous correction of bamboo strip displacement and image frame, ensures that the bamboo node position and distance information correspond accurately, and provides a reliable benchmark for subsequent positioning. The image recognition module is used to process images captured by the camera using a pre-trained large visual model, and to construct a robust stitching model for bamboo strip image sequences by combining SIFT feature point matching and depth optical flow field estimation. This eliminates matching misalignment caused by texture repetition and illumination fluctuations, and then analyzes the acquired bamboo strip images to identify the position of bamboo nodes in the image, achieving misalignment-free stitching of panoramic bamboo strip images and high-precision semantic recognition of bamboo node regions. The distance calculation module is used to integrate YOLOv8 object detection and Transformer regression network to locate the bamboo joint region. Based on the camera's imaging principle and the synchronization relationship between the encoder and the camera, it calculates the actual distance between bamboo joints, and completes the sub-pixel level localization of the bamboo joint center point and the high-precision generation of the bamboo joint spacing sequence. The logic implementation module is used to integrate various modules in the LabVIEW programming environment, build a complete logic flow, and construct a graphical control flow. It realizes the coordinated scheduling of data acquisition, image stitching, bamboo joint detection, and joint distance calculation to form a closed-loop control system, output bamboo joint recognition results and bamboo joint distance information, realizes multi-task coordinated scheduling and adaptive optimization of acquisition strategy, and ensures stable and efficient operation of the system.

[0008] Preferably, the data acquisition module specifically includes: The encoder is coaxially mounted on the drive shaft end of the conveying roller below the driven wheel to collect pulse signals during the bamboo strip conveying process. After calibration, the signals are converted into bamboo strip displacement and a distance sequence synchronized with the bamboo strip movement is generated. This enables high-precision continuous recording of bamboo strip displacement and provides a reliable distance reference for subsequent positioning. The camera is set to external hardware trigger mode, receives trigger pulses output by the encoder at set displacement intervals, and controls the camera to expose and capture images when the bamboo strip moves to the preset position. The original image frames are aligned with the distance sequence time, ensuring that each image frame corresponds precisely to the actual displacement position of the bamboo strip, thus eliminating time deviation from the source. A subpixel-level displacement field estimation of the surface texture motion of bamboo strips between consecutive image frames is performed using the Farneback optical flow method. The estimated displacement is then fused and corrected with the encoder displacement to compensate for synchronization deviations caused by mechanical vibration and conveying speed fluctuations, thereby achieving subpixel-level synchronization correction and eliminating the impact of asynchronous motion on the positioning accuracy of bamboo joints.

[0009] Preferably, the data acquisition module further includes: While the encoder triggers the camera to acquire images, the encoder's cumulative pulse value corresponding to each frame of the image is recorded. This constructs an original mapping table between the image frame and the actual displacement of the bamboo strip, serving as the reference data for subsequent synchronous correction and laying the data foundation for synchronous correction. The optical flow motion vector of the natural texture on the surface of bamboo strips is extracted from continuous image frames to establish the continuous motion trajectory of bamboo strips in the image coordinate system. The trajectory is then aligned with the encoder displacement curve on the time axis to identify the local offset between the two, thereby achieving accurate identification and quantification of synchronization deviation caused by mechanical vibration and speed fluctuation. By interpolating and correcting the original mapping table using local offsets, a subpixel-level precision image frame-displacement synchronization mapping relationship is generated. This enables precise binding of each micro-segment displacement of the bamboo strip with the corresponding image region, ensuring a high-precision correspondence between the bamboo joint image position and the physical displacement coordinates, and eliminating positioning deviations.

[0010] Preferably, the image recognition module specifically includes: The scale space of the acquired bamboo strip image sequence is constructed, SIFT feature points are extracted and their descriptors are calculated. The preliminary registration relationship between adjacent image frames is established by feature point matching, eliminating mismatches caused by repeated textures on the bamboo strip surface, effectively suppressing texture repetition interference, and improving registration stability. By combining a deep optical flow field estimation network, pixel-level motion in image sequences is densely modeled. The continuity constraint of the optical flow field is used to optimize the feature point matching results, and a robust splicing model of bamboo strip images under continuous motion is constructed to achieve pixel-level motion compensation and enhance the robustness of the splicing model. The stitched long bamboo strip image is input into a pre-trained large visual model. The semantic features of the bamboo joint region are extracted through the attention mechanism in the model, and the pixel-level positioning results of the bamboo joint in the global coordinate system of the image are output. The semantic features of the bamboo joint are accurately extracted, ensuring the reliability of the positioning results.

[0011] Preferably, the image recognition module further includes: In the SIFT feature point matching process, a matching filtering mechanism based on the nearest neighbor distance ratio is introduced to eliminate low-confidence matching point pairs caused by illumination fluctuations and reflections on the bamboo strip surface, and retain stable feature points for spatial alignment of image sequences, effectively improving the purity and registration stability of feature point matching. By utilizing the dense optical flow vector output by the deep optical flow field estimation network, local deformation compensation is performed on the seam area in the robust stitching model, eliminating stitching distortion caused by bamboo strip bending or camera angle changes, generating continuous and misaligned panoramic images of bamboo strips, and significantly enhancing the continuity and geometric consistency of the stitched images. Based on the recognition of bamboo joint positions using a large visual model, and combined with the motion boundary information provided by the optical flow field, the edges of the bamboo joint region are refined, and a sequence of bamboo joint boundary coordinates with sub-pixel accuracy is output, which greatly improves the precision and measurement accuracy of bamboo joint boundary positioning.

[0012] Preferably, the distance calculation module specifically includes: The panoramic image of bamboo strips is divided into equally spaced image blocks. The YOLOv8 object detection network is used to perform preliminary detection of bamboo joint regions in each image block. The bounding boxes and confidence scores of bamboo joint candidate regions are output, realizing rapid localization and confidence evaluation of bamboo joint regions, and improving detection efficiency and reliability. The detected bamboo node candidate regions are input into the Transformer regression network, which uses its self-attention mechanism to extract global contextual features of the bamboo node regions, refines the position of the bounding boxes, and outputs the image coordinates of the bamboo node center point, thereby enhancing the bamboo node localization accuracy, effectively distinguishing targets with similar shapes, and reducing localization errors. Based on the synchronous mapping relationship between the encoder and the camera, the image coordinates of the center point of the bamboo node are back-projected to the actual displacement coordinate system of the bamboo strip. Combined with the camera imaging geometric parameters, the physical distance between adjacent bamboo nodes is calculated, generating a bamboo node spacing sequence. This enables accurate calculation of the physical spacing of bamboo nodes, providing accurate distance data for subsequent processing.

[0013] Preferably, the distance calculation module further includes: When detecting bamboo node candidate regions in YOLOv8, a multi-scale feature pyramid structure is used to fuse image features under different receptive fields, which enhances the robustness of detection for bamboo node targets with blurred edges and varied shapes, outputs a complete set of bamboo node candidate regions, effectively reduces the false detection and false detection rates of bamboo nodes, and improves the completeness of detection. A positional encoding mechanism is introduced into the Transformer regression network to embed the relative positional information of bamboo joint candidate regions in the panoramic image into the feature vector, enabling the network to distinguish bamboo joint targets with similar shapes but different positions, thereby improving the localization accuracy and enhancing the network's ability to distinguish similar bamboo joint targets. The calculated bamboo joint spacing sequence is compared with the original encoder displacement sequence in a closed-loop verification to identify abnormal spacing values ​​caused by image stitching errors or detection omissions. The average spacing of adjacent bamboo joints is then used for compensation and correction to ensure the continuity of the bamboo joint spacing sequence and the reliability of the measurement, and to eliminate interference from outliers.

[0014] Preferably, the logic implementation module specifically includes: In the LabVIEW programming environment, an event-driven state machine architecture is built, which encapsulates data acquisition, image stitching, bamboo joint detection and joint calculation as independent execution units. Data transmission and state synchronization between units are realized through a queue message mechanism to ensure the timing consistency and resource coordination when multiple tasks are processed in parallel, and to avoid data blocking and resource conflicts. The system sets up a global closed-loop control logic to dynamically adjust the displacement interval of the encoder-triggered camera based on the bamboo joint spacing sequence output by the distance calculation module, optimizes the image acquisition density, achieves an adaptive balance between bamboo joint detection accuracy and data processing load, realizes dynamic closed-loop optimization of acquisition strategy and detection results, and improves the system's adaptability to different bamboo specifications. The graphical front panel displays a real-time panoramic image of bamboo strips, bamboo node identification marks, and bamboo node spacing distribution map. The identification results are stored in the database according to bamboo strip batches, forming a traceable bamboo node detection record. This facilitates real-time monitoring of the detection status by operators and enables traceable management and quality control of the detection results.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. This bamboo joint recognition device and system based on AI visual detection establishes a sub-pixel-level synchronous mapping relationship between image frames and displacement sequences during the bamboo strip conveying process through a hybrid mechanism of encoder hardware triggering and image optical flow dynamic compensation. This effectively eliminates the problem of asynchronous movement caused by factors such as mechanical vibration, conveying speed fluctuations and installation errors, ensuring that the pixel position of the bamboo joint in the image corresponds precisely to its actual physical coordinates, and significantly improving the positioning accuracy of bamboo joint detection.

[0016] 2. This bamboo joint recognition device and system based on AI visual detection combines scale-invariant feature transformation feature point matching and depth optical flow field estimation to construct a robust stitching model for bamboo strip image sequences. By combining sparse feature point registration with dense optical flow compensation, it effectively overcomes matching misalignment and stitching distortion caused by factors such as repeated surface textures of bamboo strips, changes in illumination, and bending deformation of bamboo strips, generating continuous and misaligned panoramic images of bamboo strips, providing a complete and clear image foundation for bamboo joint recognition.

[0017] 3. This is a bamboo joint recognition device and system based on AI visual detection. It uses a pre-trained large visual model to perform semantic segmentation on panoramic images of bamboo strips, and combines motion boundary information provided by optical flow field to refine the edges of bamboo joint regions, achieving sub-pixel level positioning of bamboo joint boundaries. By integrating the semantic features of bamboo joints and motion field gradient features, it can accurately identify bamboo joint targets with diverse shapes and blurred edges, effectively suppress false detections caused by the similarity between bamboo joints and bamboo body textures, and output bamboo joint positioning results with high confidence. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the external structure of a bamboo joint recognition device based on AI visual detection according to the present invention. Figure 1 ; Figure 2 This is a schematic diagram of the external structure of a bamboo joint recognition device based on AI visual detection according to the present invention. Figure 2 ; Figure 3 This is a partial structural schematic diagram of a bamboo joint recognition device based on AI visual detection according to the present invention; Figure 4 This is a three-dimensional structural diagram of a bamboo joint recognition device based on AI visual detection according to the present invention; Figure 5 This is a schematic diagram illustrating the workflow of a bamboo joint recognition system based on AI visual detection according to the present invention.

[0019] In the diagram: 1. Outer frame; 2. Thread inlet; 3. Driven wheel; 4. Active feeder; 41. Drive motor; 42. First feed roller; 43. Second feed roller; 5. Encoder; 6. Camera; 7. Light source; 8. Photoelectric sensor. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0021] Example 1, please refer to Figures 1 to 4This invention provides a technical solution: a bamboo joint recognition device based on AI visual detection, including an outer frame 1 that connects to a bamboo drawing machine. One side of the outer frame 1 has a bamboo strip inlet 2 flush with the output port of the bamboo drawing machine, ensuring a smooth transition of the bamboo strip from the drawing machine into the recognition system and preventing jamming or misalignment. A driven wheel 3 is slidably installed on the inner wall of the outer frame 1 near the bamboo drawing machine to achieve adaptive guidance when the bamboo strip enters, ensuring conveying stability. An active feeder 4 is installed on the side of the outer frame 1 away from the bamboo drawing machine, providing active traction for bamboo strip conveying and ensuring uniform and controllable throughput. An encoder 5 connected to the outer surface of the outer frame 1 is installed below the wheel 3 to collect bamboo strip displacement signals in real time and provide a synchronous reference for distance measurement. A camera 6 is fixedly installed on the top of the outer frame 1 to realize continuous imaging of the bamboo strip surface and obtain complete bamboo joint image information. A photoelectric sensor port 8 located above the driven wheel 3 is opened on the top of the outer frame 1 to detect the bamboo strip's position and trigger the encoder and camera to work together. A light source 7 fixed to the top of the outer frame 1 is set on one side of the camera 6, and the light source 7 is tilted to provide uniform and stable supplementary lighting conditions and enhance the recognizability of bamboo joint texture. Furthermore, the active feeder 4 includes a drive motor 41 fixedly installed at the bottom of the outer frame 1 to ensure stable and reliable feeding power and improve the continuity and controllability of bamboo strip conveying. The output end of the drive motor 41 is connected to the first feeding roller 42 via belt drive to achieve smooth power transmission and reduce transmission impact and mechanical vibration. A second feeding roller 43 is set above the first feeding roller 42 to form a roller clamping structure to ensure the stability of the bamboo strip's posture during conveying. The two ends of the first feeding roller 42 rotate with the inner wall of the outer frame 1 to provide stable rotational support and ensure the reliability of the feeding roller's long-term operation. The two ends of the second feeding roller 43 slide with the inner wall of the outer frame 1 to adapt to the passage requirements of bamboo strips of different thicknesses and enhance the universal adaptability of the device.

[0022] Example 2, as Figures 1 to 5 As shown, based on Embodiment 1, the present invention also provides a bamboo joint recognition system based on AI visual detection, mounted on the above-mentioned bamboo joint recognition device, including the following modules: The data acquisition module uses encoder 5 and camera 6 to acquire data, obtaining distance information of bamboo strip movement and image data including bamboo node information. It then uses image optical flow to dynamically compensate for motion offset, achieving sub-pixel-level synchronization correction between bamboo strip displacement and image frames. This ensures precise correspondence between bamboo node position and distance information, providing a reliable benchmark for subsequent positioning. Encoder 5 is coaxially mounted on the drive shaft end of the conveyor roller below driven wheel 3, acquiring pulse signals during bamboo strip conveying. These signals are calibrated and converted into bamboo strip displacement, generating a distance sequence synchronized with the bamboo strip movement. This achieves high-precision continuous recording of bamboo strip displacement, providing a reliable distance benchmark for subsequent positioning. Camera 6 is set to external hardware trigger mode, receiving trigger pulses output by encoder 5 at set displacement intervals, and controlling camera 6 to expose and capture images when the bamboo strip moves to the preset position, acquiring original image frames aligned with the distance sequence time, ensuring that each image frame corresponds precisely to the actual displacement position of the bamboo strip, eliminating time deviation from the source, and using Farneback optical flow method to perform sub-pixel level displacement field estimation of the bamboo strip surface texture movement between consecutive image frames, fusing and correcting the estimated displacement with encoder displacement, compensating for synchronization deviation caused by mechanical vibration and conveying speed fluctuations, achieving sub-pixel level synchronization correction, and eliminating the impact of asynchronous movement on bamboo joint positioning accuracy; It should be noted that encoder 5 is coaxially mounted on the drive shaft end of the conveyor roller below the driven wheel, ensuring that the encoder 5 shaft rotates synchronously with the conveyor roller. During the conveying process, the bamboo strip drives the conveyor roller to rotate, and encoder 5 generates pulse signals linearly related to the bamboo strip displacement. By pre-calibrating the encoder 5 pulse equivalent, i.e., the actual bamboo strip movement distance corresponding to each pulse, the real-time acquired pulse sequence is converted into a bamboo strip displacement sequence. This displacement sequence continuously records the movement trajectory of the bamboo strip in the conveying direction at high resolution. Camera 6 operates in an external hardware trigger mode, with its trigger signal input connected to the pulse output terminal of encoder 5. During the bamboo strip conveying process, encoder 5 generates a trigger signal for every preset number of pulses accumulated, i.e., every set displacement interval the bamboo strip moves. The pulse is sent to camera 6. Upon receiving the trigger pulse, camera 6 immediately performs an exposure operation to acquire an image frame of the bamboo strip surface at the current moment. Through this hardware-level triggering mechanism, each image frame corresponds to a known displacement position of the bamboo strip, thereby achieving time alignment between the image frame and the displacement sequence during the acquisition stage. The optical flow motion vector of the natural texture of the bamboo strip surface is extracted from adjacent image frames, the actual displacement of the bamboo strip in the image coordinate system is calculated, and this displacement is compared segment by segment with the corresponding displacement recorded by encoder 5. Based on the local offset between the two, the mapping relationship between the original image frame and the displacement sequence is interpolated and corrected to generate a synchronously corrected mapping table, realizing the precise binding of each micro-segment displacement of the bamboo strip with the corresponding image region, and eliminating the influence of asynchronous motion on the positioning accuracy of the bamboo joint. Furthermore, the data acquisition module also includes: while the encoder 5 triggers the camera 6 to acquire images, recording the cumulative pulse value of the encoder 5 corresponding to each frame of the image, constructing an original mapping table between the image frame and the actual displacement of the bamboo strip, serving as the reference data for subsequent synchronous correction, laying the data foundation for synchronous correction, extracting the optical flow motion vector of the natural texture on the surface of the bamboo strip in continuous image frames, establishing the continuous motion trajectory of the bamboo strip in the image coordinate system, aligning the trajectory with the displacement curve of the encoder 5 on the time axis, identifying the local offset between the two, realizing the accurate identification and quantification of the synchronous deviation caused by mechanical vibration and speed fluctuation, using the local offset to interpolate and correct the original mapping table, generating a sub-pixel level precision image frame-displacement synchronous mapping relationship, realizing the precise binding of each micro-segment displacement of the bamboo strip with the corresponding image area, ensuring a high-precision correspondence between the bamboo joint image position and the physical displacement coordinates, and eliminating positioning deviation; It should be noted that encoder 5 is installed at the end of the conveyor roller drive shaft. As the bamboo strip is conveyed, it generates continuous pulse signals, synchronously reads the current pulse count value, and binds it to the exposure time of camera 6, forming a raw mapping table between the image frame index and the actual displacement of the bamboo strip. This mapping table is stored in the system cache as an array, recording the starting displacement coordinates corresponding to each image frame, providing a reference for subsequent synchronous correction. The accuracy of the mapping table construction directly depends on the calibration accuracy of the encoder 5 pulse equivalent. Therefore, before actual deployment, encoder 5 needs to be calibrated at multiple points to ensure that the linear relationship error between the pulse value and the actual displacement of the bamboo strip is controlled within the allowable range. A pyramid-based Farneback optical flow algorithm is used to perform pixel-level motion estimation on adjacent image frames, outputting the surface of the bamboo strip... The displacement field of feature points is then used to obtain the actual displacement curve of the bamboo strip in the image plane through global motion vector statistics. This curve is compared point by point with the displacement curve of encoder 5 on the time axis. The dynamic time warping algorithm is used to identify the local offset between the two. This offset reflects the synchronization deviation caused by factors such as mechanical vibration, conveying speed fluctuation, and encoder 5 installation error. Using the encoder 5 displacement sequence as a reference, the optical flow offset is used as a correction term. The cubic spline interpolation method is used to compensate the displacement coordinates in the original mapping table frame by frame, so that each frame of the image corresponds to an actual bamboo strip displacement value that has been corrected at the sub-pixel level. The corrected mapping table realizes the precise binding of each micro-segment displacement of the bamboo strip with the corresponding image region, ensuring that the pixel position of the bamboo node in the image corresponds to its physical position. The image recognition module processes images captured by camera 6 using a pre-trained large-scale visual model. It combines SIFT feature point matching and depth optical flow estimation to construct a robust stitching model for bamboo strip image sequences, eliminating matching misalignments caused by texture repetition and illumination fluctuations. Furthermore, it analyzes the acquired bamboo strip images, identifying the positions of bamboo nodes within the images. This enables misaligned stitching of panoramic bamboo strip images and high-precision semantic recognition of bamboo node regions. The module also constructs scale space for the acquired bamboo strip image sequences, extracts SIFT feature points and calculates their descriptors, and establishes preliminary registration relationships between adjacent image frames through feature point matching, eliminating bamboo strip surface defects. To address mismatches caused by repeated surface textures, this paper effectively suppresses texture repetition interference and improves registration stability. By combining a deep optical flow field estimation network, pixel-level motion in the image sequence is densely modeled. The continuity constraint of the optical flow field is used to optimize the feature point matching results. A robust splicing model for bamboo strip images under continuous motion is constructed to achieve pixel-level motion compensation and enhance the robustness of the splicing model. The spliced ​​long bamboo strip image is input into a pre-trained large visual model. The semantic features of the bamboo joint region are extracted through the attention mechanism in the model, and the pixel-level positioning results of the bamboo joint in the global coordinate system of the image are output. The semantic features of the bamboo joint are accurately extracted, ensuring the reliability of the positioning results. It should be noted that in actual operation, a Gaussian difference scale space is first constructed for the bamboo strip image sequence acquired by camera 6. SIFT feature points with scale invariance are extracted through extremum detection, and a 128-dimensional descriptor vector is calculated for each feature point. For adjacent image frames, the nearest neighbor distance ratio method is used for initial feature point matching. A distance ratio threshold is set to eliminate mismatched point pairs caused by the periodic repetition of bamboo strip surface texture. For the retained stable matching point pairs, the homography transformation matrix between adjacent frames is estimated using the random sampling consensus algorithm to establish a preliminary image registration relationship. Through multi-layer analysis of the scale space, feature points can be stably detected at different scales, effectively overcoming feature confusion caused by the similarity between bamboo nodes and bamboo body textures. Based on the preliminary registration, a deep optical flow field estimation model based on a convolutional neural network is introduced to perform pixel-level dense motion modeling on the image sequence. This model takes two adjacent frames as input and outputs the optical flow vector of each pixel through an encoder-decoder structure, forming a complete motion field description. The spatial continuity and smoothness constraints of the optical flow field are used to refine the SIFT feature point matching results. Global optimization corrects local registration biases, especially in areas where bamboo strips are curved or reflective. Optical flow estimation effectively compensates for the lack of sparse feature points. Subsequently, based on the optimized transform parameters, a multi-band fusion algorithm is used to seamlessly stitch the image sequence, constructing a continuous representation of the bamboo strips in the panoramic image coordinate system. This ensures that the stitched image is natural and coherent in the texture transition area, without misalignment or ghosting. The stitched panoramic image of the bamboo strips is then input into a pre-trained large-scale visual model. This model uses a Transformer architecture and extracts multi-scale semantic features from the image through a multi-head attention mechanism. During the training phase, the model has learned the morphological features and contextual information of the bamboo joint region, enabling it to accurately activate the feature response region corresponding to the bamboo joint in the panoramic image. During inference, the model outputs a heatmap of the bamboo joint region. After post-processing non-maximum suppression, pixel-level boundary coordinates of the bamboo joint in the global image coordinate system are obtained. This localization result is output in sequence form, with each bamboo joint corresponding to a set of precise start and end coordinates, along with a confidence score. By associating the bamboo joint coordinates with the aforementioned synchronous correction mapping table, the precise conversion of the bamboo joint image position to the physical displacement coordinate is achieved. Furthermore, the image recognition module also includes: in the SIFT feature point matching process, a matching and filtering mechanism based on the nearest neighbor distance ratio is introduced to eliminate low-confidence matching point pairs caused by illumination fluctuations and reflections on the bamboo strip surface, and retain stable feature points for spatial alignment of image sequences, effectively improving the purity and registration stability of feature point matching; using the dense optical flow vector output by the deep optical flow field estimation network, local deformation compensation is performed on the seam area in the robust splicing model to eliminate splicing distortion caused by bamboo strip bending or changes in camera angle, generating continuous and misaligned panoramic images of bamboo strips, significantly enhancing the continuity and geometric consistency of the spliced ​​images; based on the recognition of bamboo joint positions by the large visual model, the edge refinement of the bamboo joint area is performed by combining the motion boundary information provided by the optical flow field, and outputting a bamboo joint boundary coordinate sequence with sub-pixel accuracy, greatly improving the precision and measurement accuracy of bamboo joint boundary positioning; It should be noted that, firstly, the 128-dimensional descriptor vectors extracted from adjacent image frames are subjected to nearest neighbor distance ratio matching and filtering. A distance ratio threshold is set, and matching point pairs with ratios higher than the threshold are judged as low-confidence matches and are removed. This effectively suppresses mismatches caused by illumination fluctuations, bamboo strip surface reflections, and periodic repetition of textures. The remaining stable matching point pairs constitute a reliable set of feature points, which are used to estimate the homography transformation matrix between adjacent frames. The optimal transformation parameters are solved iteratively through a random sampling consensus algorithm to achieve preliminary spatial alignment of the image sequence. This filtering mechanism ensures that the feature points participating in the registration have high consistency and representativeness, providing a stable spatial transformation benchmark for image stitching and significantly improving the robustness and accuracy of the registration process. Based on the preliminary registration, a deep optical flow field estimation model based on a convolutional neural network is introduced. Taking two adjacent frames as input, the dense optical flow vector of each pixel is output through an encoder-decoder structure to form a complete motion field description. Utilizing the spatial continuity and smoothness constraints of the optical flow field, local deformation compensation is performed on the seam area in the stitching model, taking into account the deformation of the bamboo strip during transport. The bending deformation and perspective distortion caused by the camera's 6-angle changes during the process are corrected point by point by pixel-level motion vectors to fix the splicing boundary. Combined with a multi-band fusion algorithm, the corrected image sequence is seamlessly fused into a continuous and misaligned panoramic image of bamboo strips, ensuring natural and coherent texture transition areas and eliminating ghosting and splicing marks. The large visual model outputs a heat map of the bamboo joint area and performs non-maximum suppression processing. Then, it further combines the motion boundary information provided by the depth optical flow field to refine the edge of the bamboo joint area. The optical flow field shows a significant change in motion vector gradient at the bamboo joint position. This feature can be used as an auxiliary positioning basis for the bamboo joint boundary. By fusing semantic recognition results and motion field gradient information, the bamboo joint boundary is corrected to sub-pixel accuracy, and the precise start and end coordinate sequence corresponding to each bamboo joint is output. Subsequently, the corrected bamboo joint image coordinates are associated with the synchronous correction mapping table generated by the data acquisition module. Using the sub-pixel correspondence between the encoder 5 displacement and the image frame in the mapping table, the position of the bamboo joint in the image coordinate system is accurately converted into physical displacement coordinates, realizing the accurate mapping of the bamboo joint image positioning to the actual spatial position. The distance calculation module is used to fuse YOLOv8 target detection and Transformer regression network to locate the bamboo joint region. Based on the imaging principle of camera 6 and the synchronization relationship between encoder 5 and camera 6, it calculates the actual distance between bamboo joints, and completes the sub-pixel level localization of the bamboo joint center point and the high-precision generation of bamboo joint spacing sequence. The logic implementation module is used to integrate various modules in the LabVIEW programming environment, build a complete logic flow, and construct a graphical control flow. It realizes the coordinated scheduling of data acquisition, image stitching, bamboo joint detection, and joint distance calculation to form a closed-loop control system, output bamboo joint recognition results and bamboo joint distance information, realizes multi-task coordinated scheduling and adaptive optimization of acquisition strategy, and ensures stable and efficient operation of the system.

[0023] Example 3, as Figures 1 to 5 As shown, based on Embodiments 1 and 2, the present invention provides a technical solution: the distance calculation module specifically includes: dividing the panoramic image of bamboo strips into equally spaced image blocks, using the YOLOv8 target detection network to perform preliminary detection of bamboo joint regions in each image block, outputting the bounding boxes and confidence scores of bamboo joint candidate regions, realizing rapid localization and confidence evaluation of bamboo joint regions, improving detection efficiency and reliability, inputting the detected bamboo joint candidate regions into the Transformer regression network, using its self-attention mechanism to extract global contextual features of bamboo joint regions, refining the position of the bounding boxes, outputting the image coordinates of the bamboo joint center point, enhancing the bamboo joint localization accuracy, effectively distinguishing similar targets, reducing localization deviation, and based on the synchronous mapping relationship between encoder 5 and camera 6, back-projecting the image coordinates of the bamboo joint center point to the actual displacement coordinate system of the bamboo strip, combining the camera imaging geometric parameters, calculating the physical distance between adjacent bamboo joints, generating a bamboo joint spacing sequence, realizing accurate calculation of the physical spacing of bamboo joints, and providing accurate distance data for subsequent processing; It should be noted that after the panoramic image of the bamboo strip is constructed, the panoramic image is first divided into equally spaced image blocks. Each image block is input into the YOLOv8 target detection network as an independent processing unit. The YOLOv8 network adopts a multi-scale feature pyramid structure, which integrates image features from different receptive fields to perform preliminary detection of bamboo joint regions in each image block, and outputs the bounding boxes and confidence scores of bamboo joint candidate regions. To improve detection accuracy, the network has learned the feature representations of bamboo joints under different shapes and lighting conditions during the training phase, which can effectively identify bamboo joint targets with blurred edges and varied shapes. Subsequently, the detected bamboo joint candidate regions are input into the Transformer regression network. This network uses a self-attention mechanism to extract global contextual features of the bamboo joint region, refines the position of the bounding box, and outputs the image coordinates of the bamboo joint center point. The Transformer regression network introduces a position encoding mechanism to embed the relative position information of the candidate region in the panoramic image into the feature vector, enabling the network to distinguish bamboo joint targets with similar shapes but different positions. After obtaining the image coordinates of the bamboo joint center point, based on the synchronous mapping relationship between encoder 5 and camera 6 established by the data acquisition module, the image coordinates of the bamboo joint center point are back-projected to the actual position of the bamboo strip. A coordinate system is used, and this synchronous mapping relationship records the binding information between each micro-segment displacement and the corresponding image region with sub-pixel precision, ensuring that the conversion error from image coordinates to physical coordinates is controlled within the allowable range. During the back projection process, the projection results are precisely corrected by combining the imaging geometric parameters of camera 6, including lens distortion coefficient, imaging magnification, and calibration parameters of camera 6 installation angle. Based on this, the distance difference between the center points of adjacent bamboo nodes in the actual displacement coordinate system of the bamboo strip is calculated sequentially to generate a bamboo node spacing sequence. This sequence records the actual physical distance between each bamboo node on the bamboo strip at high resolution. To ensure the bamboo node spacing sequence... To ensure measurement reliability, the calculated bamboo joint spacing sequence is compared with the original displacement sequence of encoder 5 in a closed-loop verification. During the verification process, the bamboo joint spacing value is compared with the corresponding interval length in the displacement record of encoder 5. Abnormal spacing values ​​caused by factors such as image splicing error, bamboo joint detection omission, or image distortion are identified. For the identified abnormal values, the average spacing of adjacent bamboo joints is called for compensation and correction. Linear interpolation or cubic spline interpolation methods are used to generate reasonable replacement values ​​to maintain the continuity and integrity of the spacing sequence. The corrected bamboo joint spacing sequence and the synchronous correction mapping table together constitute complete bamboo joint positioning data. Furthermore, the distance calculation module also includes: when YOLOv8 detects bamboo joint candidate regions, it adopts a multi-scale feature pyramid structure to fuse image features under different receptive fields, enhances the robustness of detection for bamboo joint targets with blurred edges and varied shapes, outputs a complete set of bamboo joint candidate regions, effectively reduces the false detection rate of bamboo joints, and improves the detection completeness. It introduces a position encoding mechanism in the Transformer regression network, embeds the relative position information of bamboo joint candidate regions in the panoramic image into the feature vector, enables the network to distinguish bamboo joint targets with similar shapes but different positions, improves the positioning accuracy, enhances the network's ability to distinguish similar bamboo joint targets, and performs closed-loop verification between the calculated bamboo joint spacing sequence and the original displacement sequence of encoder 5 to identify abnormal spacing values ​​caused by image stitching errors or detection omissions, and uses the average spacing of adjacent bamboo joints for compensation and correction, ensuring the continuity of the bamboo joint spacing sequence and the reliability of measurement, and eliminating outlier interference. It should be noted that the constructed panoramic image of bamboo strips is divided into several image blocks at equal intervals. Each image block is sequentially input into the YOLOv8 object detection network. This network adopts a multi-scale feature pyramid structure, which enhances the robustness of detecting bamboo joints with blurred edges and varied shapes by fusing image features from different receptive fields. During the training phase, the network has learned the feature representations of bamboo joints under different shapes and lighting conditions. During inference, it outputs the bounding boxes and confidence scores of bamboo joint candidate regions, forming a complete set of bamboo joint candidate regions. During the detection process, the multi-layer structure of the feature pyramid effectively suppresses false detections caused by the similarity between bamboo joint and bamboo body textures, ensuring the integrity and accuracy of the candidate region set. The detected bamboo joint candidate regions are then input into the Transformer regression network. This network uses a self-attention mechanism to extract global contextual features of the bamboo joint region, refines the position of the bounding boxes, and outputs the image coordinates of the bamboo joint center point. The Transformer regression network introduces a position encoding mechanism to embed the relative position information of the candidate region in the panoramic image into the feature pyramid. The eigenvectors enable the network to distinguish bamboo joint targets with similar shapes but different positions, significantly improving positioning accuracy. Subsequently, based on the synchronous mapping relationship between encoder 5 and camera 6 established by the data acquisition module, and combined with the imaging geometric parameters of camera 6, the image coordinates of the bamboo joint center point are back-projected to the actual displacement coordinate system of the bamboo strip. The distance difference between the center points of adjacent bamboo joints is calculated sequentially to generate a bamboo joint spacing sequence. The calculated bamboo joint spacing sequence is compared item by item with the original displacement sequence of encoder 5. A closed-loop verification mechanism is used to identify abnormal spacing values ​​caused by factors such as image stitching errors, missed bamboo joint detection, or image distortion. During the verification process, the bamboo joint spacing value is compared with the corresponding interval length in the displacement record of encoder 5. Abnormal values ​​with deviations exceeding the set threshold are marked. For the identified abnormal spacing values, the average spacing of the adjacent bamboo joints is called for compensation and correction. Linear interpolation or cubic spline interpolation methods are used to generate reasonable replacement values ​​to maintain the continuity and integrity of the spacing sequence. The corrected bamboo joint spacing sequence and the synchronous correction mapping table together constitute complete bamboo joint positioning data. The logic implementation module specifically includes: building an event-driven state machine architecture in the LabVIEW programming environment, encapsulating data acquisition, image stitching, bamboo node detection, and node spacing calculation into independent execution units, realizing data transmission and state synchronization between units through a queue message mechanism, ensuring timing consistency and resource coordination during multi-task parallel processing, avoiding data blocking and resource conflicts, setting up global closed-loop control logic, dynamically adjusting the displacement interval of encoder 5 triggering camera 6 according to the bamboo node spacing sequence output by the distance calculation module, optimizing image acquisition density, achieving an adaptive balance between bamboo node detection accuracy and data processing load, realizing dynamic closed-loop optimization of acquisition strategy and detection results, improving the system's adaptability to different bamboo specifications, displaying a panoramic image of bamboo strips, bamboo node identification marks, and bamboo node spacing distribution map in real time through a graphical front panel, and storing the identification results in the database according to bamboo strip batches to form traceable bamboo node detection records, facilitating real-time monitoring of the detection status by operators, and realizing traceable management and quality control of detection results; It should be noted that data transmission and state synchronization between execution units are achieved through a queue message mechanism, ensuring timing consistency and resource coordination during multi-task parallel processing. The state machine architecture uses the encoder's 5-pulse trigger event during the bamboo strip conveying process as the driving source for system operation, dynamically switching working states according to the current processing stage to achieve fully automated scheduling from hardware trigger acquisition to result output. This state machine architecture effectively avoids resource conflicts and data blocking problems in multi-threaded concurrent operations, providing reliable software support for the stable operation of the system under high-speed bamboo strip conveying conditions. When a dense distribution of bamboo node spacing is detected, the trigger displacement interval is automatically shortened to increase the image acquisition density and ensure complete acquisition of image information in the bamboo node area. When the bamboo node spacing is sparse, the system appropriately increases the trigger interval to reduce the amount of image data, achieving a balance between detection accuracy and data processing burden. The adaptive balancing mechanism uses the bamboo node spacing sequence as feedback. The control unit in the state machine architecture calculates the optimal trigger parameters in real time and sends the adjustment command to the encoder 5 trigger module, forming a dynamic closed-loop optimization of the acquisition strategy and detection results, which significantly improves the system's adaptability to different bamboo specifications. Operators can intuitively monitor the bamboo node detection status and measurement results during equipment operation. The front panel integrates image display controls and waveform control. The bamboo node identification position is marked with a highlight in the panoramic image, and the spacing distribution map presents the actual physical distance between each bamboo node in the form of a curve, which makes it easy for operators to quickly identify abnormal areas. At the same time, the identification results are stored in the database according to bamboo strip batches. Each record contains key information such as bamboo strip number, bamboo node spacing sequence, detection time and equipment parameters, forming a complete and traceable bamboo node detection record.

[0024] The following is a detailed description of the workflow of this AI-based visual detection bamboo joint recognition device and system.

[0025] Bamboo strips are fed from the output port of the drawing machine into the filament inlet 2 of the outer frame 1. Driven by the active feeder 4, the bamboo strips pass sequentially through the driven wheel 3 and the active feed roller group. The encoder 5 is coaxially mounted on the end of the conveyor roller drive shaft, generating pulse signals that are linearly related to the displacement as the bamboo strips are conveyed. The camera 6 adopts an external hardware trigger mode, receiving the trigger pulses output by the encoder 5 at set displacement intervals. When the bamboo strip moves to the preset position, the camera exposes and captures an image, achieving hardware-level time alignment between the image frame and the displacement sequence. During the acquisition process, the system synchronously records the cumulative pulse value of the encoder 5 corresponding to each frame of the image. A raw mapping table between image frames and the actual displacement of bamboo strips is constructed. The data acquisition module further adopts the Farneback optical flow algorithm based on pyramid layering to perform pixel-level motion estimation on adjacent image frames, extract the optical flow motion vector of the natural texture on the surface of bamboo strips, establish the continuous motion trajectory of bamboo strips in the image coordinate system, and compare the trajectory with the encoder displacement curve point by point. The dynamic time warping algorithm is used to identify local offsets. Based on the encoder displacement sequence, the cubic spline interpolation method is used to compensate the raw mapping table frame by frame to generate a synchronous correction mapping table with sub-pixel accuracy. The image recognition module constructs a Gaussian difference scale space for the acquired image sequence, extracts SIFT feature points and calculates 128-dimensional descriptor vectors. It performs initial feature point matching using the nearest neighbor distance ratio method, eliminates mismatched point pairs caused by texture repetition, estimates the homography transformation matrix using the random sampling consistency algorithm, and establishes a preliminary registration relationship. Based on this, a deep optical flow field estimation network is introduced to output a dense optical flow vector for each pixel. Local deformation compensation is performed on the seam area in the stitching model. Combined with a multi-band fusion algorithm, a continuous and misaligned panoramic image of bamboo strips is generated. The panoramic image is input into a pre-trained large visual model, and semantic features of the bamboo joint region are extracted through a multi-head attention mechanism. The heat map of the bamboo joint region is output. After non-maximum suppression, the pixel-level boundary coordinates of the bamboo joint in the global coordinate system of the image are obtained. The edge is refined by combining the optical flow field motion boundary information, and a sub-pixel precision bamboo joint boundary coordinate sequence is output. The distance calculation module divides the panoramic image into equally spaced image blocks, which are then sequentially input into the YOLOv8 target detection network. A multi-scale feature pyramid structure is used to detect bamboo node candidate regions, outputting bounding boxes and their confidence scores. The candidate regions are then input into a Transformer regression network, where a self-attention mechanism and positional encoding are used to refine the image coordinates of the bamboo node center points. Based on a synchronous correction mapping table and camera imaging geometric parameters, the bamboo node center point coordinates are back-projected onto the actual bamboo strip displacement coordinate system. The distance difference between adjacent bamboo node center points is calculated sequentially, generating a bamboo node spacing sequence. A closed-loop verification mechanism is used to compare the sequence with the encoder's original displacement sequence, identifying and compensating for abnormal spacing values. The logic implementation module constructs an event-driven state machine architecture in the LabVIEW environment, encapsulating each functional module as an independent execution unit. A queue message mechanism enables data transmission and state synchronization. The encoder's 5-pulse trigger event serves as the system's driving source, dynamically switching working states to achieve fully automated scheduling from hardware-triggered acquisition to result output. The bamboo node recognition results and spacing information are displayed in real-time on the graphical front panel and stored in batches in the database, forming a complete and traceable detection record.

[0026] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A bamboo joint recognition device based on AI visual detection, characterized in that, The device includes an outer frame (1) that connects to the wire drawing machine. One side of the outer frame (1) has a wire inlet (2) that is flush with the output port of the wire drawing machine. A driven wheel (3) is slidably installed on the inner wall of the outer frame (1) near the wire drawing machine. An active feeder (4) is installed on the side of the outer frame (1) away from the wire drawing machine. An encoder (5) connected to the outer surface of the outer frame (1) is provided below the driven wheel (3). A camera (6) is fixedly installed on the top of the outer frame (1). A photoelectric sensor (8) located above the driven wheel (3) is provided on the top of the outer frame (1). A light source (7) fixed to the top of the outer frame (1) is provided on one side of the camera (6), and the light source (7) is tilted.

2. The bamboo joint recognition device based on AI visual detection according to claim 1, characterized in that: The active feeder (4) includes a drive motor (41) fixedly installed at the bottom of the outer frame (1). The output end of the drive motor (41) is connected to a first feed roller (42) via belt drive. A second feed roller (43) is provided above the first feed roller (42) to form a roller clamping structure. The two ends of the first feed roller (42) rotate with the inner wall of the outer frame (1), and the two ends of the second feed roller (43) slide with the inner wall of the outer frame (1).

3. A bamboo joint recognition system based on AI visual detection, mounted on the bamboo joint recognition device according to any one of claims 1 or 2, characterized in that, Includes the following modules: The data acquisition module is used to acquire data by combining the encoder (5) and the camera (6), obtain the distance information of the bamboo strip movement and the image data containing the bamboo joint information, and eliminate the motion offset by dynamic compensation through the image optical flow method. The image recognition module is used to process the images captured by the camera (6) using a pre-trained large visual model, and to construct a robust splicing model of bamboo strip image sequence by combining SIFT feature point matching and depth optical flow field estimation. The module analyzes the collected bamboo strip images and identifies the position of the bamboo nodes in the image. The distance calculation module is used to fuse YOLOv8 target detection and Transformer regression network to locate the bamboo joint area, and calculate the actual distance between bamboo joints based on the imaging principle of camera (6) and the synchronization relationship between encoder (5) and camera (6). The logic implementation module is used to integrate various modules in the LabVIEW programming environment, build a complete logic flow, construct a graphical control flow, form a closed-loop control system, and output bamboo joint recognition results and bamboo joint distance information.

4. The bamboo joint recognition system based on AI visual detection according to claim 3, characterized in that: The data acquisition module specifically includes: The encoder (5) is coaxially mounted on the drive shaft end of the conveying roller below the driven wheel (3) to collect the pulse signal during the bamboo strip conveying process. After calibration, it is converted into the bamboo strip displacement and a distance sequence synchronized with the bamboo strip movement is generated. Set the camera (6) to external hardware trigger mode, receive the trigger pulse output by the encoder (5) at the set displacement interval, control the camera (6) to expose and collect images when the bamboo strip moves to the preset position, and obtain the original image frame aligned with the distance sequence time. A subpixel-level displacement field estimation of the surface texture motion of bamboo strips between consecutive image frames is performed using the Farneback optical flow method. The estimated displacement is then fused and corrected with the encoder displacement to compensate for synchronization deviations caused by mechanical vibration and conveying speed fluctuations.

5. The bamboo joint recognition system based on AI visual detection according to claim 4, characterized in that: The data acquisition module also includes: While the encoder (5) triggers the camera (6) to acquire images, the cumulative pulse value of the encoder (5) corresponding to each frame of image is recorded, and the original mapping table between the image frame and the actual displacement of the bamboo strip is constructed. Extract the optical flow motion vector of the natural texture of the bamboo strip surface in the continuous image frame, establish the continuous motion trajectory of the bamboo strip in the image coordinate system, and align the trajectory with the displacement curve of the encoder (5) on the time axis to identify the local offset between the two. By interpolating and correcting the original mapping table using local offsets, a subpixel-level precision image frame-displacement synchronization mapping relationship is generated, enabling precise binding of each micro-segment displacement of the bamboo strip to the corresponding image region.

6. The bamboo joint recognition system based on AI visual detection according to claim 3, characterized in that: The image recognition module specifically includes: The scale space of the acquired bamboo strip image sequence is constructed, SIFT feature points are extracted and their descriptors are calculated, and a preliminary registration relationship between adjacent image frames is established by feature point matching to eliminate mismatches caused by repeated surface textures of bamboo strips. By combining a deep optical flow field estimation network, pixel-level motion in image sequences is densely modeled, and feature point matching results are optimized using the continuity constraint of the optical flow field, thus constructing a robust splicing model for bamboo strip images under continuous motion. The stitched long bamboo strip image is input into a pre-trained large visual model. The semantic features of the bamboo joint region are extracted through the attention mechanism in the model, and the pixel-level localization result of the bamboo joint in the global coordinate system of the image is output.

7. The bamboo joint recognition system based on AI visual detection according to claim 6, characterized in that: The image recognition module further includes: In the SIFT feature point matching process, a matching filtering mechanism based on the nearest neighbor distance ratio is introduced to eliminate low-confidence matching point pairs caused by illumination fluctuations and reflections on the bamboo strip surface, and retain stable feature points for spatial alignment of image sequences. By using the dense optical flow vector output by the depth optical flow field estimation network, local deformation compensation is performed on the seam area in the robust splicing model to eliminate splicing distortion caused by bamboo strip bending or camera angle changes, and a continuous, misaligned panoramic image of bamboo strips is generated. Based on the identification of bamboo joint positions using a large visual model, and combined with the motion boundary information provided by the optical flow field, the edges of the bamboo joint region are refined, and a sequence of bamboo joint boundary coordinates with sub-pixel accuracy is output.

8. The bamboo joint recognition system based on AI visual detection according to claim 3, characterized in that: The distance calculation module specifically includes: The panoramic image of bamboo strips is divided into equally spaced image blocks. The YOLOv8 object detection network is used to perform preliminary detection of bamboo joint regions in each image block, and the bounding boxes and confidence scores of bamboo joint candidate regions are output. The detected bamboo node candidate regions are input into the Transformer regression network, which uses its self-attention mechanism to extract global context features of the bamboo node regions, refines the bounding boxes, and outputs the image coordinates of the bamboo node center point. Based on the synchronous mapping relationship between the encoder (5) and the camera (6), the image coordinates of the center point of the bamboo node are back-projected to the actual displacement coordinate system of the bamboo strip. Combined with the camera imaging geometric parameters, the physical distance between adjacent bamboo nodes is calculated to generate a bamboo node spacing sequence.

9. A bamboo joint recognition system based on AI visual detection according to claim 8, characterized in that: The distance calculation module also includes: When detecting bamboo joint candidate regions in YOLOv8, a multi-scale feature pyramid structure is used to fuse image features under different receptive fields, which enhances the robustness of detection for bamboo joint targets with blurred edges and varied shapes, and outputs a complete set of bamboo joint candidate regions. A positional encoding mechanism is introduced into the Transformer regression network to embed the relative positional information of bamboo joint candidate regions in the panoramic image into the feature vector, enabling the network to distinguish bamboo joint targets that are similar in shape but different in location. The calculated bamboo joint spacing sequence is compared with the original displacement sequence of the encoder (5) to perform closed-loop verification, identify abnormal spacing values ​​caused by image splicing errors or detection omissions, and use the average spacing of adjacent bamboo joints to compensate and correct them.

10. A bamboo joint recognition system based on AI visual detection according to claim 3, characterized in that: The logic implementation module specifically includes: In the LabVIEW programming environment, an event-driven state machine architecture is built, and data acquisition, image stitching, bamboo joint detection and joint calculation are encapsulated as independent execution units. Data transmission and state synchronization between units are realized through a queue message mechanism. Set up a global closed-loop control logic, dynamically adjust the displacement interval of the encoder (5) triggering the camera (6) according to the bamboo joint spacing sequence output by the distance calculation module, optimize the image acquisition density, and achieve an adaptive balance between bamboo joint detection accuracy and data processing load. The graphical front panel displays a real-time panoramic image of bamboo strips, bamboo node identification marks, and bamboo node spacing distribution map. The identification results are stored in the database according to bamboo strip batches to form a traceable bamboo node detection record.

Images