Tomato raw material loading, unloading and sorting method and system fused with AI visual identification

Abstract

The invention discloses a tomato raw material loading, unloading and sorting method and system fused with AI visual recognition, and relates to the technical field of AI vision. A tomato raw material loading, unloading and sorting method fused with AI visual identification comprises the following steps that multiple frames of images are collected, qualified images are reserved through quality detection, the qualified frames are subjected to pixel-level segmentation, tomato segmentation masks are generated through multi-frame fusion, and blind areas are complemented; based on the complemented segmentation mask, high residual target points are judged, space attributes are extracted, the optimal moving path and action queue of the crane pipe are obtained, and a dynamic complementing and scanning queue is inserted to be issued and executed; and in the execution process, residual changes are monitored, the flushing effect is judged, dynamic recognition, complementary scanning and queue updating are conducted, parameters are adjusted, and sorting is conducted till discharging is completed. Accurate recognition, efficient path planning and dynamic supplementary scanning in the tomato loading, unloading and sorting process are achieved, and then the problems that in the prior art, due to insufficient recognition precision, path planning is unreasonable, and residual areas are missed in scanning are effectively solved.

Description

TECHNICAL FIELD

[0001] The application relates to the field of AI vision technology, and in particular to a tomato raw material loading and unloading and sorting method and system fusing AI vision recognition. BACKGROUND

[0002] Continuous breakthroughs in AI vision recognition technology provide core technical support for the automation and precision of tomato raw material sorting. Multi-spectral image recognition and three-dimensional imaging technology can capture the appearance characteristics of tomatoes in all directions. Through an image processing and recognition system, the shape and size of the tomato pile can be accurately identified, and the spatial coordinates of the tomato pile can be located to achieve accurate identification of the residual material distribution in the vehicle. The visual recognition technology can not only accurately identify the basic morphological information of the tomatoes, but also sensitively detect minor defects such as scratches, punctures, water marks, dry scars, rotting, and cracking. This provides a rich and reliable image data basis for subsequent sorting and judgment. Through the training of a large number of tomato image samples, the system has strong target detection and classification capabilities and can quickly and accurately identify the type, maturity level, and various defect information of the tomatoes. The development of image segmentation technology, especially instance segmentation technology, allows the sorting system to clearly locate the defect position and range, enabling automatic sorting actions.

[0003] For example, the Chinese invention patent with the publication number CN119228791B discloses a wood board detection method based on visual recognition, which includes: obtaining a first detection image set on a first detection platform; performing image enhancement processing on the first gray image to obtain an enhanced detection image; determining the contour image information of each wood board; performing image recognition processing on the contour image information of each wood board through a contour recognition algorithm to determine the contour feature parameters of the contour image information of each wood board; and inputting the contour image information of each wood board in the enhanced detection image into a pre-constructed board defect recognition model for recognition to obtain the infrared defect recognition result of each wood board.

[0004] In the prior art, when tomato raw materials are unloaded and sorted, single-frame image segmentation is mostly relied on. Once strong light reflection, water stains, stains, and occlusions occur, the pixel gray scale, color distribution, and texture features of the region will deviate significantly from the normal sample rules, which may lead to misjudgment during the judgment process. The tomato target may be incorrectly classified as background or impurities. In the image fusion process, some key areas are contaminated in most frames, the fusion result is missing, the information is not recoverable, and the real target cannot be identified. Since the abnormal areas cannot be accurately identified and processed for a long time, they become sorting dead angles or cleaning blind areas, affecting the overall unloading and sorting efficiency. Ultimately, the lack of recognition accuracy leads to unreasonable path planning and residual area missed scanning. SUMMARY

[0005] In order to solve the technical problem that the rule updating and node inspection are not synchronized in the data labeling updating link of data quality inspection in the prior art, the embodiment of the present application provides a tomato raw material loading and unloading and sorting method and system fusing AI visual identification.

[0006] In one aspect, a tomato raw material loading and unloading and sorting method fusing AI visual identification is provided, which comprises the following steps:

[0007] S1, a plurality of fixed cameras are used to continuously collect a plurality of frames of images, and image quality detection is performed, and unqualified images are removed, and local pollution areas in qualified images are specially marked, and local confidence information is retained, the unqualified image refers to an image whose quality is too poor (for example, a large range of blur, overexposure, and serious occlusion), so that the whole frame cannot be normally recognized, and the unqualified image is directly removed, and for the whole frame, the target main body is still clear, but there is a local pollution area (such as water stain, reflection, dust, and partial occlusion), which does not cause the whole frame to be scrapped, but affects the confidence of the local pixels, and the local pixels are retained;

[0008] S2, pixel-level segmentation is performed on each frame of image of all qualified frames to obtain the category (such as tomato, impurity, or background) and segmentation confidence matrix of each pixel, and all qualified frames are fused in the same spatial position to generate a global tomato segmentation mask (the segmentation confidence scores of all qualified frames in the same pixel position are weighted and fused) and a blind area, and the blind area is inferred and completed by a neighborhood to obtain a fusion and completion segmentation mask, the segmentation confidence matrix represents the probability score of each pixel belonging to each category for quantifying the reliability of identification, and the segmentation mask is used to clearly indicate the distribution area of each type of object in the image;

[0009] S3, based on the fusion and completion segmentation mask, the number of tomato pixels and the spatial height in each spatial grid block are counted to calculate the thickness, density, and residual area of the tomato, determine the high residual target point, and extract the spatial attributes (spatial geometric center coordinates, average thickness, residual area, etc.) of the high residual target point, and all high residual target points are collected to form a target point set that needs to be washed, combined with the kinematic model of the crane pipe and the environmental constraints, a sampling-based path planning algorithm (such as the rapid extended random tree algorithm) is used to comprehensively consider the length of the connecting rod of the mechanical arm, the rotation radius, the movement speed / acceleration, the effective coverage range of the spray head, and the obstacle avoidance requirement, generate an optimal movement path of the crane pipe covering all target points in turn, generate an action queue containing execution parameters including but not limited to movement trajectory, crane pipe spray head parameter, and blind area with confidence not meeting the standard is listed in the supplementary scanning queue and dynamically inserted into the action queue, and then sequentially issued to the specific execution component;

[0010] S4, during the action execution, continuously high-frequency acquisition and analysis of new images, real-time monitoring of all high residual target points, blind area and its neighborhood of tomato residual change, obtaining the residual area change rate of each target point or blind area to judge the flushing effect, and dynamically identifying the supplementary scanning according to the flushing effect, if new supplementary scanning demand is found in the execution process, updating the supplementary scanning queue and dynamically adjusting the flushing parameters, until the unloading is completed and the tomato raw material sorting is carried out.

[0011] In another aspect, a tomato raw material loading and unloading and sorting system with AI visual recognition fusion is provided, which comprises an image quality detection module, a multi-frame fusion module, an action queue generation module and an identification supplementary scanning module.

[0012] The image quality detection module is used for continuously acquiring multiple frames of images through multiple fixed cameras, and performing image quality detection to remove unqualified images, and performing special identification on local contaminated areas while retaining local confidence information.

[0013] The multi-frame fusion module is used for pixel-level segmentation of each frame of image of all qualified frames to obtain the region classification and segmentation confidence matrix of each pixel, and multi-frame fusion of all qualified frames at the same spatial position to generate a global tomato segmentation mask and a blind area, and neighborhood inference completion of the blind area to obtain a fusion and completion segmentation mask.

[0014] The action queue generation module is used for counting the number of tomato pixels and spatial height in each spatial grid block based on the fusion and completion segmentation mask, determining high residual target points and extracting their spatial attributes, collecting all high residual target points to form a target point set that needs to be flushed, obtaining an optimal moving path of the snorkel through a path optimization algorithm, and generating an action queue containing execution parameters.

[0015] The identification supplementary scanning module is used for continuously high-frequency acquisition and analysis of new images during the action execution, real-time monitoring of all high residual target points, blind area and its neighborhood of tomato residual change, obtaining the residual area change rate of each target point or blind area to judge the flushing effect, and dynamically identifying the supplementary scanning according to the flushing effect, if new supplementary scanning demand is found in the execution process, updating the supplementary scanning queue and dynamically adjusting the flushing parameters, until the unloading is completed and the tomato raw material sorting is carried out.

[0016] The technical scheme provided by the embodiment of the application has at least the following beneficial effects:

[0017] 1. Most existing technologies rely on single-frame, single-view segmentation or detection, which are prone to misjudgment and missed detection due to stains, reflections, and occlusion. This technology employs multi-camera, multi-frame fusion and segmentation confidence quantification, using neighborhood inference completion and dynamic thresholding to effectively cover previously easily missed blind spots and dead zones. This solves the problems of poor adaptability and missegmentation in complex scenes, contamination, and local occlusion encountered by traditional methods. Furthermore, the fixed motion path leads to repetitive, missed, or inefficient rinsing. This technology calculates the optimal movement path based on spatial attributes and generates an action queue to control the movement of the loading arm. High-frequency image acquisition and analysis during the rinsing process allows for real-time detection of residual area changes. Upon detecting re-scanning needs, it automatically prioritizes and readjusts parameters, solving the "over-rinse / under-rinse" dilemma where actual residue is difficult to detect and completely remove in a timely manner, thus avoiding ineffective repeated rinsing and missed scans.

[0018] 2. By combining image quality detection with pixel-level segmentation fusion, the reliability of the overall recognition and segmentation results is improved. High-quality images are selected frame by frame based on multiple indicators such as image clarity, reflection, contamination, and occlusion. Based on the fusion of spatially co-located multi-frame information, high-confidence category determination at the pixel level is achieved, generating a global segmentation mask. This effectively solves the problems of single-frame misjudgment, target loss caused by environmental interference, and segmentation breakage. Compared with traditional techniques that rely solely on single-frame or unfiltered fusion, selecting qualified frames before fusion improves the continuity and stability of target segmentation, reduces the risk of missegmentation and target loss caused by environmental interference, and enhances the ability to distinguish between tomato targets and impurities, laying a solid data foundation for subsequent blind zone detection and spatial completion.

[0019] 3. For pixels marked as "unidentified / segmented unsuccessfully / in blind spots," most are simply "left blank," "classified as background," or "directly interpolated for filling," which cannot guarantee the consistency of categories and spatial rationality of the filled area. It is difficult to achieve category adaptation and confidence adaptation of the filled result. Instead of simply using full-image averaging or direct interpolation, this approach uses spatial consistency inference through neighborhood inference for filling. It searches for high-confidence pixels in the surrounding area and performs local statistics for each category. Both the filled category and confidence are based on the statistics of the real neighborhood, resulting in strong spatial consistency. Neighborhood inference combined with spatial mean filling makes the segmentation result more spatially coherent and less prone to "islands" or irrelevant points. This effectively reduces segmentation breaks, edge fragmentation, and spatial blind spot residues, improves the efficiency of blind spot recognition and filling, and reduces the misjudgment rate.

[0020] 4. Rinsing / washing is mostly done in one go or on a fixed cycle, lacking quantitative criteria based on actual residue levels. Blindly adding sweeps / repeated rinsing wastes resources, and missed sweeps or under-washing are not easily detected in time, leaving some areas with long-term blind spots. However, after each rinse, AI vision automatically calculates the "tomato residue area" of each target point or blind spot. The rinsing effect can be intuitively reflected by the rate of change of residue area. If the monitoring effect is insufficient, it automatically adds sweeps at high frequency and dynamically adds to the sweeping queue, prioritizing resource allocation. Compared with traditional methods, which are difficult to intelligently allocate resources and are prone to "not adding what should be added and repeatedly sweeping ineffective points", the entire rinsing process is intelligently monitored, effectively solving the problems of rigid queue management, inefficient addition of new difficult areas, and low overall efficiency due to resource dispersion. This improves the accuracy, efficiency, and resource utilization of the entire unloading and sorting chain. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart of a tomato raw material loading, unloading and sorting method integrating AI visual recognition provided in an embodiment of the present invention;

[0023] Figure 2 This is a flowchart of neighborhood inference and completion provided in an embodiment of the present invention;

[0024] Figure 3 A flowchart of the supplementary scanning requirement processing stage provided in an embodiment of the present invention;

[0025] Figure 4 This is a schematic diagram of a tomato raw material loading, unloading and sorting system that integrates AI visual recognition, provided in an embodiment of the present invention. Detailed Implementation

[0026] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0027] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0028] To make the technical problems, technical solutions and advantages to be solved by the present application clearer, specific embodiments will be described in detail below with reference to the drawings.

[0029] As Figure 1 shown, a flow chart of a tomato raw material loading and unloading and sorting method fusing AI visual recognition provided by an embodiment of the present application, the method comprising the following steps: continuously collecting multiple frames of images through multiple fixed cameras, and performing image quality detection, rejecting unqualified images, and simultaneously performing special identification on local contaminated areas, retaining local confidence information, and unqualified images representing whole frames of images that cannot be normally identified; performing pixel-level segmentation on each frame of image of all qualified frames to obtain the classification of each pixel (such as tomato, impurities or background) and the segmentation confidence matrix, and fusing all qualified frames at the same spatial position to generate a global tomato segmentation mask (weighting and fusing the segmentation confidence scores of all qualified frames at the same pixel position) and a blind area region, and performing neighborhood inference completion on the blind area region to obtain a fusion and completion segmentation mask, the segmentation confidence matrix representing the probability score of each pixel belonging to each category for quantifying the reliability of identification, and the segmentation mask being used to clearly indicate the distribution area of each type of object in the image; based on the fusion and completion segmentation mask, the number of tomato pixels and the spatial height in each spatial grid block are counted to calculate the thickness, density and residual area of the tomato, determine the high residual target points and extract their spatial attributes (spatial geometric center coordinates, average thickness, residual area, etc.), and all high residual target points are summarized to form a target point set that needs to be flushed, an optimal moving path of the snorkel covering all target points in turn is obtained through a path optimization algorithm (such as the rapid expansion random tree algorithm), an action queue containing execution parameters is generated, the execution parameters including but not limited to motion trajectory, snorkel nozzle parameter, and the blind area region with confidence not meeting the standard is listed in the supplementary scanning queue and dynamically inserted into the action queue and sequentially issued to the specific execution component; during the action execution, new images are continuously collected and analyzed at a high frequency, the changes of tomato residues of all high residual target points, blind area regions and their neighborhoods are monitored in real time, the residual area change rate of each target point or blind area is obtained to judge the flushing effect, and the supplementary scanning is dynamically identified accordingly, if new supplementary scanning demand is found during the execution process, the supplementary scanning queue is updated and the flushing parameters are dynamically adjusted until the unloading is completed and the tomato raw material is sorted.

[0030] In this embodiment, by continuously collecting and screening high-quality images through multiple fixed cameras, the accuracy of obtaining the actual situation of the unloading site and the anti-interference ability are improved. During the unloading process of the tomato tank car, invalid images that cannot effectively distinguish the target due to factors such as blur, pollution, and obstruction can be automatically identified and removed, ensuring the reliability of subsequent analysis data. Based on the fusion and completion of the segmentation mask, the internal space of the carriage is further analyzed by grid, automatically counting the number of tomato pixels and the distribution height in each block, and accurately calculating the thickness, density, and total area of the residual tomatoes. When the thickness and area of the residual tomatoes in a certain low-lying corner of the carriage are detected to be much higher than the set threshold, the spatial geometric center and average thickness of the region are automatically extracted, and it is quickly determined that it is a high-residual target that needs to be focused on washing. High-frequency image acquisition and analysis are maintained to monitor the changes in tomato residues in the target points and blind areas in real time, and the washing effect is judged in a timely manner and new supplementary scanning requirements are dynamically identified. If it is found that the residual area change rate of a certain region after supplementary scanning is still below the threshold, the system will automatically optimize the washing parameters and update the supplementary scanning queue to ensure that the unloading process is always in a closed-loop adaptive optimal state until the residual in the carriage is completely removed and smoothly enters the raw material sorting link.

[0031] Further, image quality detection is performed to remove unqualified images. The specific steps are as follows: the clarity, reflection, stain, and obstruction of each frame of image are detected in turn; the Laplacian variance of each frame of image is calculated, if the Laplacian variance is lower than the set clarity threshold, it is determined that the frame is a blurred image and is removed, otherwise it is retained as a qualified frame and reflection image determination is performed, the set clarity threshold is represented by the result of summing and averaging the historical clarity thresholds in the historical image quality detection process in the database; the high-light pixel ratio in the image is calculated, i.e. the percentage of pixels with a gray value greater than a set gray value in the total pixels, if the high-light pixel ratio is greater than a set high-light pixel ratio, it is determined that the frame is a reflection image and is removed, otherwise it is retained as a qualified frame and stain image determination is performed, the set high-light pixel ratio is represented by the result of summing and averaging the historical high-light pixel ratios in the historical image quality detection process in the database.

[0032] The pixel ratio of abnormal color spots in the image can be detected by color clustering or template matching, if the pixel ratio of abnormal color spots is greater than a set ratio threshold, it is determined that the frame is a contaminated image and is removed, otherwise it is retained as a qualified frame and obstruction image determination is performed; the area ratio of the effective target region in the image is calculated, such as the percentage of the actually recognized target region in the total image area, if the area ratio is lower than a set area threshold, it is determined that the frame is an obstruction image and is removed, otherwise it is retained as a qualified frame; if any of the detection items is unqualified, it is determined that the frame is an unqualified frame and is removed, only frames with all indicators qualified are retained for subsequent identification.

[0033] In the embodiment, through systematic and multi-level quality detection on each frame of captured image, in the tomato tank unloading operation, the automatic discrimination of typical distortion factors such as clarity, reflection, stain and shielding is carried out on each frame of image in turn. For the frames with blur, strong reflection, water stain pollution or shielding phenomenon, accurate identification and automatic rejection can be achieved through Laplace variance, gray statistical and color clustering algorithm, and only the high-quality frames that meet all indicators such as clarity, illumination, color and effective target area ratio are reserved. Taking the actual scene as an example, during the unloading process, the car body wall reflection or local shielding by pipelines and foreign matters often occurs. If the low-quality images are not rejected, it is easy to cause target misjudgment, region omission and fuzzy segmentation boundary in subsequent segmentation and recognition. Through careful multi-index screening, the frames with information distortion are completely rejected, and high-confidence data reflecting the unloading site is ensured for subsequent AI model processing. In this way, not only the segmentation accuracy and spatial distribution restoration capability of tomatoes and impurities are greatly improved, but also the misjudgment rate of blind area, dead angle and special pollution area is significantly reduced, providing high-reliability basic data support for subsequent global mask generation, high-residual point extraction and path planning steps.

[0034] Further, the global tomato segmentation mask and the blind area region are generated. The specific steps are as follows: collecting the segmentation confidence scores of all qualified frames in the same space, weighting and fusing the segmentation confidence scores of all qualified frames in the same pixel position through image registration algorithm to obtain the fusion confidence score of the pixel point in each category; for each pixel point, comparing the fusion confidence scores of each category, selecting the category with the maximum value as the final classification label of the pixel, and obtaining the final classification label of all pixels, and combining the final classification label of each pixel to form a complete segmentation mask.

[0035] The number of frames with segmentation confidence scores lower than the set segmentation confidence score threshold in all qualified frames is counted for each pixel point, which is recorded as the low-confidence frame number of the pixel. The low-confidence frame number and the total frame number are used to obtain the low-confidence proportion of the pixel. The set segmentation confidence score threshold is represented by the sum average result of the historical segmentation confidence score threshold in the historical weighting fusion process in the database. If the low-confidence proportion is greater than the set low-confidence proportion, it indicates that the pixel is polluted, shielded or segmentation is not reliable in most frames, and the pixel is determined as a blind area region and marked as a to-be-completed object. If the low-confidence proportion is not greater than the set low-confidence proportion, the segmentation result of the pixel in most frames is reliable, and the category with the highest fusion confidence score can be directly used as the final classification label. The set low-confidence proportion is represented by the sum average result of the historical low-confidence proportion in the historical weighting fusion process in the database.

[0036] In the embodiment, for multi-camera and images collected at different times, a unified reference view is selected as the target coordinate system for registration of all images, a feature point extraction and matching algorithm (such as SIFT, ORB, etc.) is used to obtain the spatial correspondence between each frame and the reference view, a robust estimation (such as RANSAC) is used to obtain the best spatial transformation matrix, all images are re-projected under the unified view to realize pixel-level spatial alignment, and then the segmentation confidence scores of all qualified frames at the same pixel position are fused by weighting:

[0037]

[0038] wherein, Conf fused (x,y,c) represents the confidence that the pixel at (x,y) after fusion belongs to category C; Conf i (x,y,c) is the confidence of the same pixel and the same category of the i-th frame, ω i is the weight of the i-th frame, and the weight is determined by calculating the quality indicators of each frame, such as sharpness (which can be measured by calculating the gradient amplitude of the image, etc.) and noise level (such as calculating the variance of the image, etc.).

[0039] By weighting and fusing the segmentation confidence scores of multiple qualified frames at the same spatial position pixel, the multi-angle and multi-time information can be fully utilized to identify the tomato, impurities and background objects in the unloading site with high precision. In the actual tomato tank unloading process, the segmentation result of a certain pixel region is unstable in some frames due to local reflection, stains or occlusion, and the segmentation confidence of the pixel in all qualified frames is statistically fused to improve the restoration ability of the real scene, realize the highly adaptive fusion and confidence quantification of the spatial segmentation result, significantly improve the global recognition accuracy and system robustness in complex scenes, and ensure the comprehensiveness and intelligence of the unloading and residual processing process.

[0040] As Figure 2 shown, a flowchart of neighborhood inference completion provided by the embodiment of the application is provided, and the specific design logic is as follows: for each blind area region, search for high-confidence non-blind area regions in the surrounding spatial neighborhood; then calculate the mean value of the confidence score for each category; then determine whether the mean value of the confidence score of a certain category exceeds a set threshold, if yes, take the category as a completion category label, assign the mean value of the confidence score of the category to the blind area region as a completion confidence score, and integrate the completed pixels into the full image to generate a completed segmentation mask; if not, temporarily keep the region as a blind area region.

[0041] It can be further understood that the blind area region is complemented by neighborhood inference. The specific steps are as follows: for each blind area region, high-confidence non-blind area regions in the surrounding spatial neighborhood are searched, and the confidence score mean of each category is calculated respectively; if the confidence score mean of a category in the neighborhood exceeds the set confidence score mean, the category is taken as the complement category label of the current blind area region, and the mean of the category confidence is assigned to the blind area region as the preliminary completion result. A Markov random field (MRF) spatial probability model is introduced, the category label of the blind area region is regarded as a to-be-optimized variable, the mean confidence of the candidate category is combined with the neighborhood label consistency to define a global energy function or a probability target, finally, through optimization (such as maximum a posteriori estimation or energy minimization), the optimal completion category label and confidence score are determined in all blind area regions, and spatial consistency completion is realized; if the confidence score mean of all categories in the neighborhood does not reach the set threshold, the neighborhood range can be appropriately expanded for re-determination, or the pixel can be temporarily kept as a blind area region; the completed pixels are integrated into the full image to generate a completed segmentation mask, and pixels or regions with a fusion confidence lower than a set fusion confidence threshold after completion are counted and marked.

[0042] In the embodiment, in the actual tomato tank unloading process, it is assumed that the corner of the tank is blocked by residual tomatoes and pipelines, resulting in that the segmentation confidence of the region is generally low in most acquisition frames. After recognizing that the corner is a segmentation blind area, a circle of high-confidence non-blind area pixels around the region is searched, for example, other tank bottom regions close to the corner. The confidence score mean of the pixels in the neighborhood is counted for categories such as tomatoes, impurities, and backgrounds. If it is found that the confidence of the tomato category in the neighborhood is obviously higher than the threshold, the blind area is directly completed as the tomato category, and the confidence consistent with the neighborhood is assigned. If the neighborhood mean is not sufficient to support completion, the blind area label is temporarily kept, and subsequent re-scanning is waited. All completed pixels are integrated into the global segmentation mask, reducing the missed scanning caused by local pollution or obstruction.

[0043] Further, the high residual target point is determined. The specific steps are as follows: the total number of pixels belonging to the tomato category is counted, and the residual area of the block is calculated in combination with the physical area of the pixel. The calculated residual area is compared with the set high residual threshold, and the set high residual threshold is represented by the sum average result of the historical high residual threshold in the historical high residual target point determination process in the database. If the residual amount of the block is higher than the set high residual threshold, the block is determined as a high residual region, and the spatial geometric center of the tomato pixel in the block is extracted as the representative coordinate of the high residual target point.

[0044] In the embodiment, the actual residual area of the region is calculated as follows:

[0045] Stomato = N tomato x A pixel ,

[0046] wherein S tomato is the residual area of tomatoes, N tomato is the number of tomato pixels in the area, and A pizel is the actual physical area of a single pixel.

[0047] In the unloading process of the tomato tank truck, the interior space of the carriage is divided into grids based on the segmentation mask. It is assumed that during the detection process, one grid block at the lower left corner of the carriage is severely stacked with tomatoes, and the total number of tomato category pixels in the block is significantly higher than that in other areas. Combined with the image parameters and the actual physical size, it is calculated that the residual area of tomatoes in the block has exceeded the set high residual threshold. At this time, the system will automatically determine that the block is a high residual area, and further extract the spatial geometric center of all tomato pixels as the representative coordinates of the high residual target point, which is used for subsequent positioning of the crane nozzle and key flushing.

[0048] Further, the blind area region with insufficient confidence is included in the supplementary scanning queue and dynamically inserted into the action queue. The specific process is as follows: when the automatic unloading action queue is generated, the low-confidence blind area region is included in the supplementary scanning and slow scanning. If the set fusion confidence threshold cannot be reached within the set time, the region is added to the dynamic inspection queue and continuously tracked in a closed loop until the region confidence meets the set standard, and then the region is removed from the supplementary scanning queue. The set fusion confidence threshold is represented by the result of averaging the sum of historical fusion confidence thresholds in the database during the generation of the automatic unloading action queue. Whenever a new blind area is identified or the original blind area has poor supplementary scanning effect, the blind area can be inserted into the supplementary scanning queue at a suitable spatial order node (such as a node close to the main path and capable of flushing along the path) in the action queue according to the spatial position of the blind area and the main path.

[0049] In the actual tomato unloading process, in the embodiment, when the automatic unloading action queue is generated, it is identified that a certain area near the corner at the top of the carriage still has a low confidence after fusion of multiple images, which is marked as a blind area. The blind area is included in the supplementary scanning queue and dynamically inserted into the action path of the crane nozzle, for example, arranged at a node close to the main cleaning path for supplementary scanning along the path. If the confidence of the region does not meet the set standard after several rounds of supplementary scanning, the region is automatically transferred to the dynamic inspection queue, and the cleaning effect of the region is continuously tracked in the subsequent unloading and cleaning period until the confidence meets the set standard, and then the region is removed from the supplementary scanning queue. At the same time, if a new low-confidence blind area is found during unloading, the system will also timely adjust the action queue and reasonably insert the new blind area to achieve optimal path and complete cleaning.

[0050] Further, the residual change rate of each target point or blind area is obtained to determine the flushing effect. The residual change rate includes a residual area change rate and a residual thickness change rate, which are used to quantify the relative degree of reduction of tomato residues in the area after flushing. The specific steps are as follows: after collecting a new image each time, the pixel-level classification is performed on all target points or blind area regions by using a segmentation mask. For each target point or blind area, the tomato residue area is counted to calculate the residual area change rate, and the depth value distribution in the target area is extracted by using the segmentation mask to obtain the residual thickness to calculate the residual thickness change rate. If the residual area change rate is less than a set residual area change rate threshold, it is determined that the current flushing is not up to standard, the target point is inserted into the supplementary scanning queue, and if the residual area change rate is not less than the set residual area change rate threshold, it is marked that the current area reduction amount meets the standard, and the residual thickness change rate is compared.

[0051] If the residual thickness change rate is less than a set residual thickness change rate threshold, it is determined again that the current flushing is not up to standard, the target point is re-inserted into the head of the supplementary scanning queue, and the number of times of supplementary scanning that is not up to standard is recorded. In subsequent scheduling, the supplementary scanning is repeated with a higher frequency. If the residual thickness change rate is not less than the set residual thickness change rate threshold, it is determined that the current flushing meets the standard, and the target point is not inserted into the supplementary scanning queue.

[0052] In the embodiment, before each flushing action, image collection and pixel-level segmentation are performed on the target points or blind areas to obtain the tomato residue area of the current region, which is recorded as the pre-flushing residue area. After completing a flushing action, image collection and segmentation are performed again on the same region to obtain the tomato residue area at the time of the current collection, which is recorded as the post-flushing residue area. The pre-flushing residue area and the post-flushing residue area are compared by using a difference value to obtain the residual area flushing reduction amount. The residual area flushing reduction amount and the pre-flushing residue area are compared by using a ratio to finally obtain the residual area change rate.

[0053] Through dynamic judgment of the residual area change rate, the flushing effect of each target point or blind area is accurately quantified, real-time feedback and intelligent decision of the cleaning action are realized, the response speed and cleaning completeness of the supplementary scanning of the key area are improved, supplementary scanning resources can be dynamically allocated, the high-frequency focusing on the difficult-to-clean area is realized, and the overall unloading and residue processing efficiency is optimized. In the past unloading, many tomatoes were often left in the corners of the car, and it was easy to miss the dead angle even after multiple flushing. Through automatic tracking and judgment, each corner can be continuously and intensively flushed, the residue is greatly reduced, and the car is clean and tidy after unloading.

[0054] As Figure 3As shown, the flowchart of the supplement scanning demand processing stage provided by the embodiment of the application, the specific design logic is: judging whether a high residual target point appears: if not, normally performing the action queue; if yes, updating and adding it to the supplement scanning queue. Then, judging whether the change rate of the number of high residual target points is less than the set threshold: if yes, taking measures of increasing the water pressure and reducing the speed of crane pipe movement, and continuously monitoring until the change rate returns to normal; if not, removing the target point from the supplement scanning queue and reducing its priority.

[0055] Further, if a new supplement scanning demand is found during the execution process, the supplement scanning queue is updated and the flushing parameters are dynamically adjusted, and the specific steps are: if a high residual target point appears, it is determined as a new supplement scanning demand, and is updated and added to the supplement scanning queue; if no high residual target point appears, the action queue is normally performed; if the change rate of the number of high residual target points is lower than the set change rate threshold, the water pressure is increased and the speed of the crane pipe movement is reduced, and the monitoring is continuously performed within a set time until the change rate of the number of high residual target points returns to the normal range, the set change rate threshold is represented by the result of summing and averaging the historical change rate thresholds in the historical updating of the supplement scanning queue in the database; if the change rate of the number of high residual target points is not lower than the set change rate threshold, it is removed from the supplement scanning queue and its supplement scanning priority is reduced, so that resources are focused on the non-compliant area, there is no need for additional adjustment of action parameters, more resources are allocated to new or unprocessed target points, and the overall unloading efficiency is optimized.

[0056] In this embodiment, through the mechanism of dynamically adjusting the supplement scanning queue and automatically optimizing the cleaning parameters, quick response can be made to new residual difficulties that continuously appear during the unloading process, the cleaning intensity and rhythm are automatically adjusted, the thoroughness and efficiency of flushing are improved. After unloading, there are almost no residual dead angles in the car interior, manual inspection and supplement scanning demand are greatly reduced, and tomato unloading work becomes more efficient, intelligent and labor-saving, realizing dynamic response and precise coverage of supplement scanning demand, and effectively guaranteeing the cleaning quality of tomato unloading. By real-time determination of the high residual target point as a new supplement scanning demand and inclusion in the supplement scanning queue, the problem of missed scanning is effectively solved.

[0057] Further, the tomato raw materials are sorted, and the specific steps are: the unloaded tomato raw materials are conveyed to the sorting line, cameras are arranged in the sorting area, image acquisition is performed in all directions and at high frequency during the entire sorting process, pixel-level analysis is performed on the acquired images, all tomatoes, impurities, unripe fruits and other targets are identified, and feature information such as spatial coordinates, segmentation confidence and blind area state is extracted for each target; the targets and their attributes and action priority are generated into a structured control result, and the sorting mechanism dynamically adjusts the sorting action according to the received structured control data to accurately perform the corresponding sorting task.

[0058] In this embodiment, through full-process high-precision monitoring of the tomato raw material sorting process and pixel-level analysis of the collected images, mature tomatoes, impurities and unripe fruits can be accurately distinguished, and multi-dimensional features such as real-time extraction of spatial coordinates, segmentation confidence and blind area state of each target can also be obtained. Through intelligent matching of the target and its attributes and action priority, the system forms a structured control instruction and issues it in real time to the sorting mechanism, so that the sorting mechanism can dynamically adjust the grabbing sequence and sorting action, and realize the priority processing of high-confidence targets and difficult areas. This process greatly improves the automation and precision level of tomato raw material sorting, effectively guarantees the consistency of product purity and quality, significantly reduces the probability of manual intervention and misjudgment, improves the sorting efficiency and the intelligent and collaborative ability of the overall operation of the production line, and provides a solid technical foundation for high-standard and high-efficiency processing of tomato raw materials.

[0059] The automation and intelligent level of the sorting operation are improved, which not only ensures that each mature tomato can be accurately retained, but also efficiently removes impurities and unripe fruits, greatly improving the product purity and overall quality. The system dynamically allocates sorting tasks, prioritizes high-confidence targets and difficult areas, effectively avoids omissions, efficiency fluctuations and subjective misjudgments in manual sorting, and greatly improves sorting efficiency.

[0060] As shown in Figure 4 , it is a structural schematic diagram of a tomato raw material loading and unloading and sorting system integrating AI visual recognition provided by the embodiment of the application. The system comprises an image quality detection module, a multi-frame fusion module, an action queue generation module and an identification supplementary scanning module.

[0061] The image quality detection module is used to continuously collect multiple frames of images through multiple fixed cameras, and to detect the image quality, remove unqualified images, and specially mark local contaminated areas while retaining local confidence information.

[0062] The multi-frame fusion module is used to perform pixel-level segmentation on each frame of image of all qualified frames to obtain the region classification and segmentation confidence matrix of each pixel, and to perform multi-frame fusion of all qualified frames at the same spatial position to generate a global tomato segmentation mask and a blind area region, and to perform neighborhood inference completion on the blind area region to obtain a fusion and completion segmentation mask.

[0063] The action queue generation module is used to count the number of tomato pixels and the spatial height in each spatial grid block based on the fusion and completion segmentation mask, determine high-residual target points and extract their spatial attributes, collect all high-residual target points to form a target point set that needs to be flushed, obtain an optimal moving path of the snorkel through a path optimization algorithm, and generate an action queue containing execution parameters.

[0064] The identification and supplementary scanning module is used for continuously collecting and analyzing new images at a high frequency during the action execution, monitoring the residual changes of all high residual target points, blind area regions and their neighborhoods in real time, obtaining the residual area change rate of each target point or blind area to determine the flushing effect, and dynamically identifying supplementary scanning according to the residual changes. If new supplementary scanning requirements are found during the execution, the supplementary scanning queue is updated and the flushing parameters are dynamically adjusted until the unloading is completed and the tomato raw materials are sorted.

[0065] In the embodiment, the image quality detection module rejects unqualified images, identifies pollution areas and retains confidence information, avoids residual misjudgment, provides high-quality data for subsequent identification, and guarantees accurate determination of the cleaning target from the source; the multi-frame fusion module generates a global tomato segmentation mask through pixel-level segmentation and spatial fusion, fills in the neighborhood of the blind area, solves the problem of visual blind area caused by the crane pipe shielding and accumulated shadow, avoids missed scanning, and provides complete basis for positioning high residual points; the action queue generation module determines high residual points according to the fused mask, generates an optimal path of the crane pipe and an action queue by using a path optimization algorithm, reduces path redundancy and idling, avoids excessive or insufficient cleaning, and improves the cleaning efficiency; the identification and supplementary scanning module dynamically monitors the residual changes and performs supplementary scanning to ensure that there is no residual during unloading, and at the same time, connects the sorting link, reduces impurity interference, and guarantees the sorting accuracy.

[0066] The above embodiments can be implemented in whole or in part by software, hardware (such as a circuit), firmware or any other combination. When implemented by software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions according to the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another, for example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center by wired (such as infrared, wireless, microwave, etc.) mode. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. containing a set of one or more available media. The available media can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD) or a semiconductor medium. The semiconductor medium can be a solid state disk.

[0067] It should be understood that the term "and / or" as used herein merely describes association between associated objects, and can represent three cases, for example, A and / or B can represent three cases of A alone, A and B together, and B alone, where A and B can be singular or plural. In addition, the character " / " herein generally represents an "or" relationship between the associated objects before and after it, but can also represent an "and / or" relationship, which can be understood in combination with the context before and after it.

[0068] In various embodiments of the present application, the size of the sequence number of each process described above does not mean the order of execution, and the execution order of each process should be determined by its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0069] Those skilled in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0070] Those skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working process of the above-described device, apparatus and unit can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0071] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any skilled person in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for loading and unloading and sorting tomato raw materials with fusion AI visual recognition, characterized in that, The method comprises the following steps: S1, continuously collecting multiple frames of images through multiple fixed cameras, and performing image quality detection to remove unqualified images, and performing special identification on local contaminated areas in qualified images, and retaining local confidence information, wherein the unqualified images represent images that are damaged and cannot provide effective identification information; S2, performing pixel-level segmentation on all qualified frame images to obtain the classification and segmentation confidence matrix of each pixel, and performing multi-frame fusion on all qualified frames at the same spatial position to generate a global tomato segmentation mask and a blind area region, and performing neighborhood inference completion on the blind area region to obtain a fusion and completion segmentation mask, wherein the segmentation confidence matrix represents the probability score of each pixel belonging to each category for quantifying the reliability of identification, and the segmentation mask is used to clearly indicate the distribution area of each type of object in the image; S3, based on the fusion and completion segmentation mask, the number of tomato pixels and the spatial height in each spatial grid block are counted to determine high residual target points and extract their spatial attributes, all high residual target points are summarized to form a target point set that needs to be washed, an optimal moving path of the snorkel is obtained through sampling base path planning, an action queue containing execution parameters is generated, the execution parameters include but are not limited to motion trajectory, snorkel nozzle parameters, and blind area regions with insufficient confidence are listed in the supplementary scanning queue and dynamically inserted into the action queue and sequentially issued to specific execution components; S4, during the action execution, new images are continuously collected and analyzed at a high frequency, and the changes in tomato residues of all high residual target points, blind area regions and their neighborhoods are monitored in real time to obtain the residual area change rate of each target point or blind area to judge the washing effect, and the supplementary scanning is dynamically identified accordingly, if new supplementary scanning requirements are found during the execution process, the supplementary scanning queue is updated and the washing parameters are dynamically adjusted until the unloading is completed and the tomato raw materials are sorted.

2. The method for tomato raw material loading, unloading and sorting with fusion of AI visual recognition according to claim 1, characterized in that, The image quality detection and removal of unqualified images are performed in the following steps: The clarity, reflection, stain and occlusion of each frame of image are detected in sequence; The Laplacian variance of each frame of image is calculated, if the Laplacian variance is lower than the set clarity threshold, the frame is determined as a blurred image and removed, otherwise, the frame is determined as a qualified frame and subjected to reflection image determination; The proportion of highlight pixels in the image, i.e. the percentage of pixels with a gray value greater than a set gray value in the total pixels, is calculated, if the proportion of highlight pixels is greater than a set highlight pixel proportion, the frame is determined as a reflection image and removed, otherwise, the frame is determined as a qualified frame and subjected to pollution image determination; The pixel proportion of abnormal color spots in the image is analyzed, which can be detected by color clustering or template matching, if the pixel proportion of abnormal color spots is greater than a set proportion threshold, the frame is determined as a pollution image and removed, otherwise, the frame is determined as a qualified frame and subjected to occlusion image determination; The area proportion of the effective target region in the image, i.e. the percentage of the actually recognized target region in the total image area, is calculated, if the area proportion is lower than a set area threshold, the frame is determined as an occlusion image and removed, otherwise, the frame is determined as a qualified frame.

3. The method of claim 1, wherein the tomato raw material loading and unloading and sorting method using fusion AI vision recognition is characterized by, The global tomato segmentation mask and the blind area region are generated in the following steps: The segmentation confidence scores of all qualified frames in the same space are collected, the segmentation confidence scores of all qualified frames in the same pixel position are fused by an image registration algorithm, and the fusion confidence score of the pixel in each category is obtained. For each pixel, the fusion confidence scores of each category are compared, the category with the largest value is selected as the final classification label of the pixel, and the final classification label of each pixel is obtained. The final classification label of each pixel is combined to form a complete segmentation mask. The number of frames with segmentation confidence scores lower than the set segmentation confidence score threshold is counted for each pixel, and the low confidence frame number is recorded as the low confidence frame number of the pixel. The ratio of the low confidence frame number to the total frame number is used to obtain the low confidence proportion of the pixel. If the low confidence proportion is greater than the set low confidence proportion, it indicates that the pixel is contaminated, occluded or segmentation is not reliable in most frames, and the pixel is determined as a blind area region and marked as a to-be-completed object. If the low confidence proportion is not greater than the set low confidence proportion, the segmentation result of the pixel in most frames is reliable, and the category with the highest fusion confidence score can be directly used as the final classification label.

4. The tomato raw material loading, unloading and sorting method with fusion AI visual recognition according to claim 3, characterized in that, The blind area region is inferred and completed in the neighborhood, and the specific steps are as follows: For each blind area region, search for high-confidence non-blind area regions in the neighborhood, and calculate the average confidence score of each category. If the average confidence score of a category in the neighborhood exceeds the set average confidence score, the category is used as the completion category label of the current blind area region, and the average confidence score of the category is assigned to the blind area region as the completion confidence score, achieving spatial consistency completion. If the average confidence scores of all categories in the neighborhood do not reach the set threshold, the pixel is temporarily kept as a blind area region. The completed pixels are integrated into the full image to generate a completed segmentation mask, and the pixels or regions with fusion confidence scores lower than the set fusion confidence threshold after completion are counted and marked.

5. The method for tomato raw material loading, unloading and sorting with fusion of AI visual recognition according to claim 1, characterized in that, The high residual target point is determined, and the specific steps are as follows: The total number of pixels belonging to the tomato category is counted, and the residual area of the block is calculated by combining the physical area of the pixel. The calculated residual area is compared with the set high residual threshold. If the residual quantity of the block is higher than the threshold, the block is determined as a high residual region, and the spatial geometric center of the tomato pixel in the block is extracted as the representative coordinates of the high residual target point.

6. The method for tomato raw material loading, unloading and sorting with fusion of AI visual recognition according to claim 1, characterized in that, The blind area region with confidence not meeting the standard is listed in the supplementary scanning queue and dynamically inserted into the action queue, and the specific process is as follows: During the generation of the automatic unloading action queue, the blind area region with low confidence is listed in the supplementary scanning and slow scanning. If the set fusion confidence threshold cannot be reached within the set time, the region is added to the dynamic inspection queue, and the region confidence is continuously closed-loop tracked until it meets the set standard and then exits the supplementary scanning queue. When a new blind area is identified or the supplementary scanning effect of the original blind area is poor, the blind area can be inserted into the supplementary scanning queue according to the spatial position of the blind area and the main path at a suitable spatial order node in the action queue (such as a point close to the main path and washable along the path).

7. The method for tomato raw material loading, unloading and sorting with fusion of AI visual recognition according to claim 1, characterized in that, The residual area change rate of each target point or blind area is acquired to determine the flushing effect, and the specific steps are as follows: After each new image is collected, all target points or blind area regions are classified at the pixel level by using a segmentation mask, the number of pixels in each target point or blind area region that are classified as tomatoes is counted, and the residual area of the region is calculated by combining the actual physical area of the pixels; If the residual area change rate is less than the set residual area change rate threshold, it is determined that the current flushing is not up to standard, the target point is reinserted into the head of the supplementary scanning queue, and the number of times of supplementary scanning that does not meet the standard is recorded, and the supplementary scanning is repeated with a higher frequency in subsequent scheduling; If the residual area change rate is greater than the set residual area change rate threshold, it is determined that the current flushing meets the standard, and the target point is not inserted into the supplementary scanning queue.

8. The method for tomato raw material loading, unloading and sorting with fusion of AI visual recognition according to claim 1, characterized in that, If new supplementary scanning requirements are found during execution, the supplementary scanning queue is updated and the flushing parameters are dynamically adjusted, and the specific steps are as follows: If a high residual target point appears, it is determined that there is a new supplementary scanning requirement, and the target point is added to the supplementary scanning queue. If no high residual target point appears, the action queue is normally performed. If the change rate of the number of high residual target points is less than the set change rate threshold, the water pressure is increased and the speed of the crane pipe movement is reduced, and the change rate of the number of high residual target points is continuously monitored within a set time until it returns to the normal range. If the change rate of the number of high residual target points is not less than the set change rate threshold, the target point is removed from the supplementary scanning queue, the priority of supplementary scanning is reduced, resources are focused on the non-standard area, action parameters do not need to be adjusted, and more resources are allocated to new or unprocessed target points.

9. The tomato raw material loading, unloading and sorting method with fusion AI visual recognition according to claim 8, characterized in that, The specific steps for sorting tomato raw materials are as follows: After the tomatoes are unloaded, the tomato raw materials are transported to the sorting line, a camera is arranged in the sorting area, and image acquisition is performed in all directions and at a high frequency during the entire sorting process. The collected images are analyzed at the pixel level to identify all targets such as tomatoes, impurities, and unripe fruits, and feature information such as spatial coordinates, segmentation confidence, and blind area state is extracted for each target. The target and its attributes and the action priority are used to generate a structured control result, and the sorting mechanism dynamically adjusts the sorting action according to the received structured control data to accurately perform the corresponding sorting task.

10. The system for loading, unloading and sorting tomato raw materials using fusion AI visual recognition according to any one of claims 1-9, characterized in that, It includes: An image quality detection module, a multi-frame fusion module, an action queue generation module, and an identification and supplementary scanning module; The image quality detection module is used to continuously collect multiple images through multiple fixed cameras, detect the image quality, and remove unqualified images. At the same time, local contaminated areas are specially marked, and local confidence information is retained. The multi-frame fusion module is used to perform pixel-level segmentation on each image of all qualified frames to obtain the region classification and segmentation confidence matrix of each pixel, and perform multi-frame fusion on all qualified frames at the same spatial position to generate a global tomato segmentation mask and a blind area region. The blind area region is inferred and completed to obtain a fusion and completion segmentation mask. The action queue generation module is configured to count the number of tomato pixels and the spatial height in each spatial grid block based on the fused and completed segmentation mask, determine high residual target points and extract their spatial attributes, collect all the high residual target points to form a target point set that needs to be washed, obtain an optimal moving path of the snorkel by using a path optimization algorithm, and generate an action queue containing execution parameters; The identification and supplementary scanning module is configured to continuously collect and analyze new images at a high frequency during the execution of the action, monitor the changes in tomato residues of all high residual target points, blind area regions and their neighborhoods in real time, obtain the residual area change rate of each target point or blind area to determine the washing effect, dynamically identify supplementary scanning based on the washing effect, update the supplementary scanning queue and dynamically adjust the washing parameters if new supplementary scanning requirements are found during the execution, and perform tomato raw material sorting until the unloading is completed.

Images