A rice panicle whole plant multi-parameter yield estimation method based on image recognition

Abstract

This invention discloses a method for multi-parameter yield prediction of rice panicles based on image recognition, belonging to the field of image recognition technology. The method includes: aligning a candidate set of rice panicle instances within a viewpoint to a unified coordinate system, establishing cross-viewpoint association links, performing deduplication, constraining the consistency of panicle attribution throughout the plant, and generating a list of deduplicated rice panicle instances and a panicle region index table; based on the panicle region index table, dividing the region into distinguishable grain areas and strongly occluded areas, and performing cross-viewpoint consistency verification and weighted fusion to generate a fused record of grain count per panicle and an occlusion compensation record; combining the fused record of grain count per panicle, the occlusion compensation record, and the list of deduplicated rice panicle instances to count the total number of grains and the number of effective panicles throughout the plant, performing yield conversion and encapsulation, writing the results into a confidence boundary, and generating a rice panicle yield prediction result set. This invention achieves unique attribution of rice panicle instances throughout the plant, avoids duplicate counting, and improves the accuracy of yield prediction based on image recognition.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a method for predicting the yield of a whole rice panicle based on image recognition and multiple parameters. Background Technology

[0002] In recent years, with the continuous evolution of deep learning and multi-view visual perception technologies, image recognition-based crop phenotypic analysis methods have made progress in the field of smart agriculture. Particularly in rice yield prediction, researchers have gradually built an intelligent perception framework that integrates object detection, instance segmentation, 3D reconstruction, and multimodal information alignment. Image recognition technology, with its advantages of being non-contact, high-throughput, and automated, provides a new technical path for analyzing rice panicle morphology and extracting key agronomic parameters. Current mainstream research focuses on using convolutional neural networks or Transformer architectures to perform pixel-level semantic understanding of rice panicle regions and combining this with scale references to estimate grain count.

[0003] Existing image recognition-based rice panicle yield prediction technologies have two main shortcomings: First, most methods have not established an effective cross-view correlation mechanism, which leads to the same panicle being counted repeatedly or missed under different views, seriously affecting the accuracy of the effective panicle count of the whole plant; Second, in areas with strong shading (such as densely overlapping parts of the panicle), there is a lack of reliable grain number compensation strategies, and relying solely on a single view image is prone to underestimating the grain number, thereby weakening the reliability of yield prediction. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a multi-parameter yield prediction method for rice panicles based on image recognition to solve the problems of repeated omissions in cross-view counting and underestimation of grain number in strongly occluded areas.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a method for multi-parameter yield prediction of rice panicles based on image recognition. The method includes: acquiring multi-view images of rice plants and writing unified timestamps, viewpoint markers, and scale reference information; performing sharpness filtering, exposure consistency normalization, and distortion correction to generate a multi-view calibration image frame set; locating candidate regions of rice panicles from the multi-view calibration image frame set, performing panicle instance segmentation, and extracting the panicle axis skeleton line and panicle base key points, while simultaneously writing appearance features and confidence levels to generate a candidate set of rice panicle instances within the viewpoints; and aligning the candidate set of rice panicle instances within the viewpoints to a unified... A coordinate system is used to establish cross-view correlation links and perform deduplication to constrain the consistency of the entire panicle attribution, generating a list of deduplicated rice panicle instances and a panicle region index table. Based on the panicle region index table, distinguishable grain regions and strongly occluded regions are divided, and cross-view consistency verification and weighted fusion are performed to generate a grain count fusion record and occlusion compensation record per panicle. The grain count fusion record per panicle, occlusion compensation record and deduplicated rice panicle instance list are combined to count the total number of grains and the number of effective panicles per plant, and the yield is converted and encapsulated, written into the confidence boundary, and a rice panicle yield prediction result set is generated.

[0008] As a preferred embodiment of the image recognition-based multi-parameter yield prediction method for rice panicles described in this invention, the steps include: acquiring multi-view images of rice plants, writing unified timestamps, viewpoint markers, and scale reference information, performing sharpness filtering, exposure consistency normalization, and distortion correction, and generating a multi-view calibrated image frame set.

[0009] Collect multi-view images of rice plants and write them with a unified timestamp, viewpoint marker and scale reference information to generate a multi-view original image set;

[0010] Sharpness filtering is performed on the original image set from multiple perspectives, and exposure consistency normalization is performed with a unified reference frame to generate a sharp and normalized image set.

[0011] By combining a clear, normalized image set with scale reference information, locating the reference geometry, performing distortion correction, and encapsulating it with a unified timestamp, a multi-view calibrated image frame set is generated.

[0012] As a preferred embodiment of the image recognition-based multi-parameter yield prediction method for rice panicles according to the present invention, the steps of locating candidate regions of rice panicles from a set of multi-view calibration image frames and performing rice panicle instance segmentation are as follows:

[0013] Based on a multi-view calibration image frame set, the rice canopy region is located, and the rice panicle response map is extracted by the slender texture of the panicle and the edge direction of the awn. At the same time, the connected regions are solidified to generate a set of candidate regions for rice panicles.

[0014] The image blocks of the panicle are cropped according to the candidate region set of the panicle, and the panicle instance segmentation is performed. The overlapping areas of the leaves are separated by the consistency constraint of the panicle axis direction, and a mask set of panicle instances is generated.

[0015] As a preferred embodiment of the image recognition-based multi-parameter yield prediction method for rice panicles according to the present invention, the steps for generating a candidate set of rice panicle instances within the viewpoint are as follows:

[0016] Morphological refinement is performed on the mask set of rice panicle instances, the panicle axis skeleton line is extracted, and the panicle base key points are anchored back along the endpoints of the skeleton to generate a skeleton key point record set.

[0017] Extract the outline and texture appearance features of rice ears from the skeleton key point record set, write them into the segmentation confidence, encapsulate them into candidate entries, aggregate and organize them according to the viewpoint label, and generate a candidate set of rice ear instances within the viewpoint.

[0018] As a preferred embodiment of the image recognition-based multi-parameter yield prediction method for rice panicles described in this invention, the steps of aligning the candidate set of rice panicle instances within a viewpoint to a unified coordinate system and establishing cross-viewpoint association links are as follows:

[0019] By using scale reference information and viewpoint markers, the key points of the panicle base and the panicle axis skeleton line in the candidate set of rice panicle instances within the viewpoint are converted to a unified coordinate system and their spatial positions are registered to generate an aligned candidate record set.

[0020] Based on the aligned candidate record set, the panicle base key point is used as the anchor point. The appearance feature similarity check and the panicle axis skeleton line morphology similarity check are performed on the rice panicle instances from different perspectives to generate a cross-view association link set.

[0021] As a preferred embodiment of the image recognition-based multi-parameter yield prediction method for rice panicles according to the present invention, the steps for generating a list of duplicate rice panicle instances and a panicle region index table are as follows:

[0022] By combining the cross-view association link set, segmentation confidence and rice ear instance mask set, representative instances of the main view and redundant view instances are obtained, and deduplication and merging are performed to generate a list of deduplicated rice ear instances.

[0023] Based on the list of deduplicated rice panicle instances, the consistency constraint of panicle affiliation for the whole plant is executed through the spatial distribution of panicle base and the direction of tiller connectivity, and panicle trimming boundary pointers are written from each perspective to generate a panicle region index table.

[0024] As a preferred embodiment of the image recognition-based multi-parameter yield prediction method for rice panicles according to the present invention, the step of dividing the discernible grain region and the strongly occluded region according to the panicle region index table is as follows:

[0025] The ear region index table is combined with the multi-view calibration image frame set, and ear-by-ear positioning and cropping are performed to generate a multi-view ear image block set.

[0026] In the vicinity of the spikelet image block set from multiple perspectives, the spikelet axis direction is cropped to construct candidate search bands for spikelets, and the discernible grain region and the strongly occluded region are divided to generate a layered spikelet region slice set.

[0027] As a preferred embodiment of the image recognition-based multi-parameter yield prediction method for rice panicles according to the present invention, the steps for generating the fusion record of grain count per panicle and the occlusion compensation record are as follows:

[0028] Grain candidate profiles are extracted from the resolvable grain regions in the layered spikelet region slice set, and the grain candidate points are fixed to generate resolvable grain count records.

[0029] In the strongly occluded region of the layered spikelet area slice set, the particulate response is extracted and aggregated into occlusion compensation count. At the same time, combined with the resolvable grain count record, a candidate record set of grain count within the view is generated.

[0030] Align the candidate record set of grain count within the view with the deduplicated rice panicle instance, and perform cross-view consistency mutual verification weighted fusion to generate a fused record of grain count per panicle and an occlusion compensation record.

[0031] As a preferred embodiment of the image recognition-based multi-parameter yield prediction method for rice panicles described in this invention, the steps of combining the fused record of grain count per panicle, the occlusion compensation record, and the list of duplicate rice panicle instances to count the total number of grains per plant and the number of effective panicles per plant are as follows:

[0032] Align the grain count fusion record with the shading compensation record, and encapsulate the grain count accounts at the ear level to generate a double-account converged table for the grain count at the ear level.

[0033] The double-account table of grain count at the panicle level is checked against the list of deduplicated rice panicle examples one by one. The number of effective panicles per plant is counted, and the total number of grains per plant and the total amount of shading compensation per plant are aggregated to generate a statistical package of total plant yield.

[0034] As a preferred embodiment of the image recognition-based multi-parameter yield prediction method for rice panicles according to the present invention, the steps for generating the rice panicle yield prediction result set are as follows:

[0035] Based on the total number of grains and the number of effective ears per plant in the whole plant yield composition statistical package, yield conversion is performed through the yield conversion parameter table to generate a yield conversion package.

[0036] The credibility boundary is obtained by combining the total amount of whole-plant shading compensation with the cross-view consistency verification and weight reduction results, and then attached to the yield conversion package for encapsulation and processing to generate a set of rice panicle yield prediction results.

[0037] The beneficial effects of this invention are as follows: through multi-view image acquisition and calibration steps, high-quality and standardized image recognition input is achieved, providing a reliable data foundation for accurate segmentation and grain number estimation, and improving the accuracy of yield prediction; through cross-view association and deduplication steps, the unique attribution of whole rice panicle instances is achieved, avoiding duplicate counting and improving the accuracy of yield prediction based on image recognition. Attached Figure Description

[0038] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the 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.

[0039] Figure 1 This is a flowchart of a multi-parameter yield prediction method for rice panicles based on image recognition.

[0040] Figure 2 This is a flowchart for multi-view image acquisition and calibration.

[0041] Figure 3 The flowchart shows the segmentation and feature extraction process for rice ear instances.

[0042] Figure 4 A flowchart for cross-perspective correlation and output forecasting. Detailed Implementation

[0043] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0044] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0045] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0046] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for predicting the yield of a single rice panicle based on image recognition, comprising the following steps:

[0047] S1: Acquire multi-view images of rice plants, write unified timestamps, viewpoint markers and scale reference information, perform sharpness filtering, exposure consistency normalization and distortion correction, and generate a multi-view calibration image frame set.

[0048] S1.1: Collect multi-view images of rice plants and write them with a unified timestamp, viewpoint marker and scale reference information to generate a multi-view original image set;

[0049] Furthermore, using the same rice plant as the data collection object, shooting was arranged by using surround angles and pitch angles. The shooting distance was kept within the reproducible range at each shooting angle, and the focal length was kept consistent to reduce scale drift. Scale reference information was placed in the shooting frame and ensured to be visible in the same frame as the rice ear area for subsequent pixel scale conversion. A unified timestamp was written to each collected multi-view image of the rice plant, and the viewpoint marker corresponding to the shooting orientation was bound to the scale reference information and written into the metadata area. The unified timestamp was generated using the same clock source to ensure time consistency across multiple views. The viewpoint marker was assigned according to the shooting arrangement order to ensure traceability across views. The scale reference information was written according to the same annotation rules to ensure consistency across batches, generating a multi-view original image set.

[0050] S1.2: Perform sharpness filtering on the original multi-view image set and perform exposure consistency normalization processing with a unified reference frame to generate a sharpness normalized image set;

[0051] Furthermore, the original multi-view image set is traversed frame by frame according to the viewpoint marker, and sharpness and motion blur metrics are extracted. Images with sharpness metrics below the sharpness threshold or motion blur metrics above the blur threshold are removed, and the removal ratio is recorded for quality backtracking. The remaining images are judged for exposure status based on brightness distribution and highlight overflow ratio, and the image with the closest exposure status to neutral and the most complete texture details is selected as the unified reference frame. The unified reference frame is bound to the corresponding viewpoint marker to ensure cross-viewpoint consistency. Using the brightness distribution and color response of the unified reference frame as the normalization benchmark, exposure consistency normalization processing is performed on each of the remaining images. The normalization processing uses the same normalization parameter set and the normalization result is limited to the effective grayscale range. Edge detail preservation constraints are performed on the normalized images to avoid texture distortion caused by excessive stretching. The normalized images are reordered and packaged according to the viewpoint marker to generate a sharp normalized image set.

[0052] It should be noted that the sharpness threshold (example range: 0.30–0.60) is set with the lower limit based on the minimum high-frequency energy level required to identify the slender texture of the rice spike and the edge direction of the awn in the original multi-view image set, and the upper limit based on the critical high-frequency energy level at which edge sharpening noise begins to dominate and causes instability in the extraction of the spike skeleton line within the sharp normalized image set.

[0053] Blur threshold (example range: 0.10–0.25), where the lower limit is set according to the minimum sharpness level corresponding to "subsequent rice panicle candidate region set extraction and rice panicle instance segmentation can still be stably completed" in the calibration sample, and the upper limit is set according to the degree of ghosting corresponding to the time when the grain candidate outlines begin to stick together and cause distortion in the division between distinguishable grain regions and strongly occluded regions.

[0054] S1.3: Combine the clear normalized image set with scale reference information, locate the reference geometry, perform distortion correction, encapsulate with a unified timestamp, and generate a multi-view calibrated image frame set.

[0055] Furthermore, the sharp, normalized image set is processed frame by frame according to viewpoint markers, and reference geometry is located by combining scale reference information. Reference geometry location is accomplished by extracting the geometric shape of the reference object boundary or reference line corresponding to the scale reference information and solidifying the set of corner points and line segments. At the same time, the visibility of the reference geometry is verified to exclude occluded or missing reference geometry. Distortion feature entries are generated based on the curvature and straightness preservation deviation of the reference geometry in the image, and a distortion correction parameter set is generated. The distortion correction parameter set is used to correct the reference geometry to a straight line shape and maintain the consistency of the geometric relationship between reference lines. Distortion correction is performed frame by frame on the sharp, normalized image set, and the straightness consistency of the distortion-corrected reference geometry is verified again to ensure that the distortion correction parameter set is effective. The distortion-corrected image is encapsulated and organized with a unified timestamp, viewpoint markers, and scale reference information, and a multi-view frame index sequence is generated based on the unified timestamp and viewpoint markers. Under the same timestamp, the viewpoint markers are organized into viewpoint sets, generating a multi-view calibration image frame set.

[0056] It should be noted that distortion feature entries refer to structured descriptive records used to characterize the degree of bending offset and straight-line preservation deviation of the reference geometry relative to the ideal straight-line shape in a clear normalized image set.

[0057] S2: Locate candidate regions of rice ears from the multi-view calibration image frame set, perform rice ear instance segmentation, extract the skeleton line of the panicle axis and key points of the panicle base, and write the appearance features and confidence scores to generate a candidate set of rice ear instances within the view.

[0058] S2.1: Based on the multi-view calibration image frame set, locate the rice canopy region, extract the rice panicle response map through the slender texture of the panicle and the edge direction of the awn, and solidify the connected regions to generate a set of candidate regions for rice panicles;

[0059] Furthermore, based on a multi-view calibrated image frame set, the rice canopy region is located frame by frame according to the viewpoint marker. Canopy region localization is performed by segmenting the region using vegetation color distribution and canopy texture density, and the canopy boundary is solidified with the maximum connected coverage. Within the rice canopy region, multi-directional fringe response extraction is performed around the slender texture of the panicle, and directional gradient aggregation is performed around the edge of the awn. The fringe response and directional gradient aggregation results are consistently superimposed to form a panicle response map. In the panicle response map, pixels with response values ​​reaching the response intensity threshold are marked as candidate pixels, while pixels that do not reach the threshold are set as non-candidate pixels. Candidate pixels are aggregated according to pixel adjacency relationships. Adjacent and continuously distributed candidate pixels are merged into the same connected region, and the boundary of the connected region is solidified to form the candidate region outline. The area of ​​each connected region is calculated, and connected regions with an area less than the region area threshold are judged as scattered noise regions and discarded. Further check the boundaries of the retained connected regions to see if there are a large number of breaks or slender bifurcations, and delete abnormal bifurcation boundary segments. Retain the boundaries of connected regions with continuous boundaries and slender band-like shapes as entries for the rice ear candidate region set, and generate the rice ear candidate region set.

[0060] It should be noted that the response intensity threshold (example range: 0.45–0.70) is set with the lower limit based on the minimum response level at which the slender texture of the panicle and the edge direction of the awn can still form a continuous and stable connected region within the rice plant canopy area, and the upper limit based on the critical response level at which the high-gloss texture of the leaves and the background stripes begin to be greatly suppressed, resulting in the missed detection of the candidate region set of the rice panicle.

[0061] The area threshold (example range: 80px–260px) is set based on the area of ​​the connected region that can still be repeatedly solidified in the multi-view calibration image frame set, where the minimum stable connected segment formed by the edge of the panicle is the largest. The upper limit is set based on the area of ​​the critical connected region that begins to cover multiple panicles or the overlapping zone of leaves and causes the boundary of the candidate region of the panicle to expand outward.

[0062] S2.2: Crop the panicle image block according to the candidate region set of rice panicles, and perform rice panicle instance segmentation. Separate the overlapping area of ​​leaves by the consistency constraint of panicle axis direction to generate a mask set of rice panicle instances;

[0063] Furthermore, the panicle image blocks are obtained by cropping the candidate region boundaries of the candidate region set in the multi-view calibration image frame set at the corresponding viewpoint. When cropping the panicle image blocks, the candidate region boundaries are expanded to cover the connection between the awn end and the panicle base. The panicle image blocks are bound to the viewpoint marker and the candidate region number to maintain traceability. For each panicle image block, panicle instance segmentation is performed to output the main pixel set of the panicle. The main direction distribution is extracted in the main pixel set of the panicle to determine the panicle axis direction. The panicle axis direction consistency constraint is used to screen out slender structures that deviate significantly from the panicle axis direction and merge the deviated structures into the leaf overlapping region. The leaf overlapping region is peeled out from the main pixel set of the panicle and the remaining region is processed to retain a single main connected region. At the same time, the boundary burrs are smoothed to ensure the stability of subsequent morphological refinement. The processed main region of the panicle is output as a panicle instance mask and is aggregated and processed according to the candidate region number to generate a set of panicle instance masks.

[0064] It should be noted that the panicle axis direction consistency constraint is obtained by statistically analyzing the main direction of the elongated texture of the main pixel set of the main panicle in the panicle image block to obtain the panicle axis direction reference, and the elongated structure that deviates from the panicle axis direction reference by more than the allowable deviation range is identified as the leaf overlapping area to achieve separation.

[0065] S2.3: Perform morphological refinement on the mask set of rice panicle instances, extract the panicle axis skeleton line, and backtrack along the endpoints of the skeleton to anchor the key points of the panicle base, generating a skeleton key point record set;

[0066] Furthermore, morphological refinement is performed on each rice ear instance mask set to obtain a connected skeleton structure with a single pixel width. Before morphological refinement, hole filling and small branch suppression are performed on the rice ear instance masks to reduce skeleton burrs and maintain the continuity of the main axis. The connected main path is extracted from the skeleton structure obtained by morphological refinement and solidified into the ear axis skeleton line. The longest connected path is selected from the ear axis skeleton line by accumulating the path length to exclude lateral short branches. The skeleton endpoint set is located in the ear axis skeleton line and backtracks inward along the skeleton endpoints to the stable intersection position of the skeleton line and the outer boundary of the rice ear instance mask to determine the ear end anchor point pair. The endpoint of the ear end anchor point pair that is closer to the stem connected boundary in the main direction of the ear axis skeleton line is anchored as the ear base key point. The pixel coordinates of the ear base key point and the ear axis skeleton line are bound to the candidate region number and written into the record entry. The skeleton key point record set is generated by aggregating and organizing according to the candidate region number.

[0067] S2.4: Extract the outline and texture appearance features of rice ears from the skeleton key point record set, write them into the segmentation confidence, encapsulate them into candidate entries, aggregate and organize them according to the viewpoint label, and generate a candidate set of rice ear instances within the viewpoint.

[0068] Furthermore, the spikelet skeleton line and spikelet base key points are located one by one from the skeleton key point record set according to the candidate region number, and the spikelet contour and texture appearance features are extracted from the corresponding spikelet image blocks. The spikelet contour is obtained by tracking the boundary of the spikelet instance mask to obtain the contour point sequence and solidify the contour shape description. The texture appearance features are formed by dividing multiple equidistant sampling bands along the direction of the spikelet skeleton line and statistically analyzing the texture undulation and granular texture density within the sampling bands. At the same time, local texture fragments are extracted in the neighborhood of the spikelet base key points to form appearance fragments that can be compared across viewpoints. The segmentation confidence is calculated under the conditions of boundary consistency and connectivity integrity of the spikelet instance segmentation results and written into candidate entries. The candidate entries are encapsulated with candidate region number, viewpoint label, spikelet skeleton line, spikelet base key points, spikelet contour and texture appearance features and segmentation confidence as content. The candidate entries are merged according to viewpoint label and sorted according to candidate region number to generate a candidate set of spikelet instances within the viewpoint.

[0069] The formula for calculating the segmentation confidence score is:

[0070] ;

[0071] in, Indicates the split confidence level. Represents the boundary consistency coefficient. This represents the average value of the boundary consistency deviation. Indicates the distance normalization reference scale, Represents the connectivity integrity coefficient. This represents the total area of ​​the mask for each rice ear instance. This represents the area of ​​the maximum connected region in the mask for the rice ear instance. This represents the total area of ​​the holes inside the mask of the rice ear instance.

[0072] It should be noted that the distance normalization reference scale converts the pixel distance into a uniform length scale based on the scale reference information, selects the upper limit of the allowable boundary positioning error under the verification of the effectiveness of the reference geometric distortion correction, and converts it back to the pixel distance for determination.

[0073] Boundary consistency coefficient (example range: 0.30–0.70) is determined based on the clarity of the rice ear outline edge and the stability of the edge point set in the ear image block to avoid edge noise dominating the score;

[0074] The connectivity integrity coefficient (example range: 0.30–0.70) is determined based on the degree of interference of the proportion of main connected regions and the proportion of holes in the rice ear instance mask set on the extraction of grain candidate contours, so as to prioritize the constraint of counting bias caused by mask breakage and omission.

[0075] S3: Align the candidate set of rice panicle instances within the viewpoint to a unified coordinate system, establish cross-viewpoint association links, perform deduplication, constrain the consistency of the entire panicle's affiliation, and generate a list of deduplicated rice panicle instances and an index table of panicle regions.

[0076] S3.1: Using scale reference information and viewpoint markers, the key points of the panicle base and the panicle axis skeleton line in the candidate set of rice panicle instances within the viewpoint are converted to a unified coordinate system, and their spatial positions are registered to generate an aligned candidate record set.

[0077] Furthermore, the candidate set of rice panicle instances within the viewpoint is processed one by one using scale reference information and viewpoint markers. Scale reference information is used to convert the pixel coordinates of the panicle base key point and the panicle axis skeleton line into a unified length scale coordinate and eliminate the pixel scale difference between different viewpoints. Viewpoint markers are used to determine the viewpoint order and viewpoint orientation of the same rice plant to establish an axial convention for a unified coordinate system. For each candidate entry, the panicle base key point is used as the position anchor point in the unified coordinate system. The point sequence of the panicle axis skeleton line is resampled according to the unified length scale coordinate and the relative displacement relationship with the panicle base key point is maintained. At the same time, the direction vector of the main direction of the panicle axis skeleton line is registered in the unified coordinate system to maintain cross-viewpoint comparability. Spatial position registration writes the coordinates of the panicle base key point, the point sequence coordinates of the panicle axis skeleton line, the viewpoint marker, and the candidate region number under the unified length scale coordinate into the aligned candidate record entry and groups them according to the rice plant identity to generate an aligned candidate record set.

[0078] It should be noted that the unified coordinate system refers to determining the unified length scale and the conversion relationship from pixel to length using scale reference information, determining the coordinate axis orientation convention using viewpoint markers, and using the same coordinate origin and the same axis to define the coordinate expression space for multi-view results under the same rice plant identity. The coordinate origin can be taken as the center position of the set of panicle base key points corresponding to the rice plant identity under the unified length scale coordinate. The coordinate axis direction is fixed by the viewpoint orientation and viewpoint sequence corresponding to the viewpoint markers to a same-direction rotation convention, so that the coordinates of the panicle base key points and the panicle axis skeleton line point sequence coordinates fall into the same scale, the same direction, and the same origin coordinate framework after conversion under different views, thereby supporting the spatial location registration and cross-view comparison in the alignment candidate record set.

[0079] S3.2: Based on the aligned candidate record set, take the panicle base key point as the anchor point, perform appearance feature similarity check and panicle axis skeleton line morphology similarity check on rice panicle instances from different perspectives, and generate a cross-view association link set;

[0080] Furthermore, based on the aligned candidate record set, a cross-view pairing candidate set is established using the panicle base key point as the anchor point in a unified coordinate system. The cross-view pairing candidate set is formed by nearest neighbor retrieval of the panicle base key point neighborhood and the nearest neighbor radius is limited. The appearance feature similarity check is performed on the cross-view pairing candidate set, converting the panicle outline and texture appearance features into similarity scores and comparing them with the similarity threshold to filter out mismatched candidates. For the candidate pairings that pass the appearance feature similarity check, the panicle axis skeleton line morphology similarity check is performed, performing length normalization and direction alignment on the panicle axis skeleton line point sequence to obtain morphological difference scores, and comparing them with the morphological difference threshold to filter out candidates with inconsistent morphology. The candidate pairings that pass both the appearance feature similarity check and the panicle axis skeleton line morphology similarity check are concatenated in the order of view markers and written into the link identifier and member list to generate a cross-view associated link set.

[0081] It should be noted that the nearest neighbor radius refers to the radius of the spatial retrieval range established with the spikelet key point as the center in a unified coordinate system. It is used to limit the cross-view pairing candidate set to include only spikelet key point pairing candidates that fall within the retrieval range, thereby excluding non-same spikelet instance candidates that are too far apart in spatial location.

[0082] Similarity threshold (example range: 0.60–0.80), where the lower limit is set according to the lowest similarity score corresponding to the stable matching of the outline and texture appearance features of the same rice ear under different views, and the upper limit is set according to the critical similarity score when the similarity between different rice ears begins to be over-distinguished, resulting in broken links and missing matches in the cross-view association link set.

[0083] The morphological difference threshold (example range: 0.15–0.35) is set based on the minimum morphological deviation that is still allowed to exist in the sequence of spikelet skeleton lines and points under multiple views of the same rice panicle after length normalization and direction alignment, and the upper limit is set based on the critical morphological deviation at which the bending morphology of the spikelet between different rice panicles begins to be misjudged as consistent and causes mismatch in the cross-view association link set.

[0084] S3.3: Combine the cross-view association link set, segmentation confidence and rice ear instance mask set to obtain the representative instance of the main view and redundant view instances, and perform deduplication and merging to generate a deduplicated list of rice ear instances.

[0085] Furthermore, the cross-view association link set is expanded one by one according to the link identifier, and the candidate region sequence number and view mark are located in the list of each link member. The candidate region sequence number and view mark are mapped to the candidate set of rice ear instances within the view to extract the segmentation confidence and mapped to the rice ear instance mask set to extract the rice ear instance mask. The main view representative instance is selected from the link members according to the joint sorting result of the segmentation confidence and the sorting result of the main connected region integrity of the rice ear instance mask. The candidate entry with the highest joint sorting result of the segmentation confidence and the sorting result of the main connected region integrity of the rice ear instance mask is selected. The candidate entries in the link member list that are not selected as the main view representative instance are registered as redundant view instances. Deduplication and merging are carried out with the main view representative instance as the merging target. The panicle skeleton line, panicle base key point, rice ear outline and texture appearance features corresponding to the redundant view instances are retained as association fields according to the view mark and written into the main view representative instance entries. At the same time, the mapping pointer from the link member to the main view representative instance is written to maintain cross-view traceability. The main view representative instance entries are aggregated and output according to the link identifier to generate a deduplicated rice ear instance list.

[0086] S3.4: Based on the list of deduplicated rice panicle instances, perform consistency constraints on the entire panicle affiliation by spatial distribution of panicle base and tiller connectivity, and write panicle trimming boundary pointers from each perspective to generate a panicle region index table.

[0087] Furthermore, based on the list of deduplicated rice panicle instances, the coordinates of key points at the panicle base are unified into a set of spatial distribution points at the panicle base. These points are used to establish candidate tiller connectivity chains based on spatial proximity and to fit tiller connectivity directions. The tiller connectivity directions are checked against the main direction of the panicle axis skeleton line to eliminate cross-tillering errors and maintain the continuity of the connectivity chain direction. For each deduplicated rice panicle instance, key points at the panicle base are bound to the tiller connectivity direction, and a whole-plant panicle attribution consistency constraint is applied. This constraint is checked against the spatial distribution order of the panicle base along the same tiller connectivity direction and the pointing relationship of the panicle axis skeleton line. Conflicting deduplicated rice panicle instance entries are rebound according to spatial distance and directional consistency. For deduplicated rice panicle instance entries that pass the whole-plant panicle attribution consistency constraint, panicle trimming boundaries are determined by viewpoint markings and encapsulated as panicle trimming boundary pointers for each viewpoint. These pointers are then bound to the deduplicated rice panicle instance identifiers, aggregated, and organized to generate a panicle region index table.

[0088] It should be noted that the tiller candidate connectivity chain is a connectivity sequence formed by connecting key points of the panicle base that are spatially adjacent and whose main directions of the panicle axis skeleton line are similar, in order of distance. It is used to characterize the arrangement relationship of rice panicles on the same tiller connectivity direction.

[0089] S4: Based on the ear region index table, divide the distinguishable grain region and the strong shading region, and perform cross-view consistency mutual verification weighted fusion to generate a fusion record of the number of grains per ear and a shading compensation record.

[0090] S4.1: Combine the ear region index table with the multi-view calibration image frame set, and perform ear-by-ear positioning and cropping to generate a multi-view ear image block set;

[0091] Furthermore, the panicle region index table is expanded line by line according to the deduplicated panicle instance identifier, and the viewpoint markers and panicle cropping boundary pointers of each viewpoint are located in the entries. The multi-view calibration image frame set is indexed and located according to the viewpoint markers to obtain the corresponding viewpoint image frames. The corresponding viewpoint image frames are located and cropped one by one according to the panicle cropping boundary pointers of each viewpoint to obtain panicle image blocks. When cropping the panicle image blocks, the cropping boundary is expanded outward to cover the end of the awn and the outer edge of the strongly shaded area and to ensure that the subsequent region division is not truncated. At the same time as generating the panicle image blocks, the deduplicated panicle instance identifier, viewpoint marker, and cropping boundary pointer are written into the association field of the panicle image block entry and an entry index field is generated to complete the one-to-one binding. Multiple viewpoint panicle image blocks under the same deduplicated panicle instance identifier are sorted and aggregated according to the viewpoint marker and the same cropping scale is kept consistent. The aggregation result is archived and encapsulated according to the deduplicated panicle instance identifier to generate a multi-view panicle image block set.

[0092] S4.2: For the multi-view ear image block set, crop the ear axis direction in the neighborhood of the ear axis skeleton line to construct the ear grain candidate search band, and divide the distinguishable grain region and the strong occlusion region to generate a layered ear region slice set.

[0093] Furthermore, the multi-view panicle image block set is processed block by block according to the deduplicated panicle instance identifier and viewpoint label. In each panicle image block, the panicle axis skeleton line is located, and a panicle axis direction clipping coordinate axis is established in the main direction of the panicle axis skeleton line. The panicle axis direction clipping retains the entire panicle range along the length direction of the panicle axis skeleton line and expands outward in the normal direction of the panicle axis skeleton line with a fixed bandwidth to form a panicle grain candidate search band. The panicle grain candidate search band is used to limit the spatial range of grain candidate extraction and exclude the leaf background at the outer edge of the panicle image block. Grain edge response and local texture undulation are extracted in the panicle grain candidate search band, and the edge closure degree and texture grain separation degree are jointly encoded into a region distinguishable metric. The region distinguishable metric is compared with the distinguishable judgment threshold to divide the panicle grain candidate search band into distinguishable grain regions and strongly occluded regions. The region boundaries of the distinguishable grain regions and strongly occluded regions are solidified and cut into multiple continuous slices according to the panicle axis direction sequence to maintain a longitudinal segmented structure that can be aligned across views. The slice boundary, region type and viewpoint label are written into the slice index field and merged and organized according to the deduplicated panicle instance identifier to generate a hierarchical panicle region slice set.

[0094] It should be noted that the edge closure degree refers to the degree to which edge segments formed by grain edge response within the candidate grain search zone can form a continuous closed contour, used to characterize whether a single grain contour is obscured or cut off or adhered to adjacent contours; the texture grain separation degree refers to the degree to which grain peaks and valleys formed by local texture undulations within the candidate grain search zone are clearly separated, used to characterize whether grain texture can be distinguished as multiple independent grains rather than being obscured or blurred into a continuous texture block.

[0095] The distinguishable determination threshold (example range: 0.50–0.70) is set based on the lowest region distinguishable measure value corresponding to the degree of edge closure and texture particle separation within the distinguishable grain region, which can still stably extract the candidate grain contour. The upper limit is set based on the critical region distinguishable measure value of the region where the scattered grain edges that may remain in the strongly occluded region are excessively suppressed, causing the distinguishable grain region to be mistakenly classified as a strongly occluded region.

[0096] S4.3: Extract candidate grain contours from the resolvable grain regions in the layered spikelet region slice set, solidify the candidate grain points, and generate a resolvable grain count record.

[0097] Furthermore, within the resolvable grain region of the layered panicle area slice set, grain edge responses are extracted slice by slice according to the slice index field, and closed contour tracing is performed on the edge responses to form grain candidate contours. The grain candidate contours are validated according to the contour shape compactness and contour area range to eliminate false contours formed by the awn lines and leaf textures. The center position of the verified grain candidate contours is marked and solidified as grain candidate points. Nearest neighbor merging is performed on the grain candidate points within the same slice to handle duplicate points caused by the overlap of adjacent contours, and trajectory continuity is checked on the grain candidate points across adjacent slices to merge the repeated occurrence of the same grain in the longitudinal slice. The number of merged grain candidate points is written as the resolvable grain count into the counting entry and bound and encapsulated with the deduplicated panicle instance identifier, view mark and slice index field, and aggregated to generate a resolvable grain count record.

[0098] S4.4: Extract the graininess response in the strongly occluded region of the layered spikelet slice set and aggregate it into occlusion compensation count. At the same time, combine it with the resolvable grain count record to generate a candidate record set of grain count within the view.

[0099] Furthermore, in the layered panicle region slice set, the boundary of the strongly occluded region is located slice by slice according to the slice index field. The strongly occluded region is divided into regular grid blocks. In each grid block, the frequency of alternation between bright and dark and the local gradient fluctuation are extracted to form a granular response value. The granular response value is compared with the granular response threshold to screen out low-response grid blocks. The granular response values ​​of the remaining grid blocks are aggregated to form a slice-level granular response summary. The slice-level granular response summary is converted with the response-count comparison relationship formed by the resolvable grain count records in adjacent resolvable grain regions to generate a slice-level occlusion compensation count. The slice-level occlusion compensation count is accumulated within the same deduplicated panicle instance identifier and viewpoint mark range to form a viewpoint-level occlusion compensation count. The viewpoint-level occlusion compensation count and the viewpoint-level resolvable grain count in the resolvable grain count record are merged and packaged into a grain count candidate entry. The grain count candidate entries are aggregated and organized according to the deduplicated panicle instance identifier and viewpoint mark to generate a viewpoint-level grain count candidate record set.

[0100] It should be noted that the graininess response threshold (example range: 0.50–0.75) is set based on the minimum response level that can still reflect the grainy bright and dark alternation and stably form the total amount of slice-level graininess response in the area of ​​strong occlusion, and the upper limit is set based on the critical response level where leaf shadow texture and background noise are largely eliminated, while the real occlusion grain response is excessively eliminated, resulting in a small occlusion compensation count.

[0101] S4.5: Align the candidate record set of grain count within the view with the deduplicated rice panicle instance, and perform cross-view consistency mutual verification weighted fusion to generate a fused record of grain count per panicle and an occlusion compensation record.

[0102] It should be noted that the candidate record set of grain count within the viewpoint is merged according to the deduplicated rice ear instance identifier, and aligned with the deduplicated rice ear instance identifier in the deduplicated rice ear instance list. For each deduplicated rice ear instance identifier, the candidate entries of grain count under the corresponding viewpoint mark are aggregated, and the viewpoint-level resolvable grain count and viewpoint-level occlusion compensation count in the candidate entries of grain count are sorted with the same caliber. According to the viewpoint mark, cross-viewpoint consistency verification is performed on the candidate entries of grain count to obtain the deviation between viewpoints. The viewpoint mark with the large deviation is registered as a deweighted member, and the grain count candidate entries of the deweighted member are downweighted. The viewpoint mark with the small deviation is sorted according to the segmentation confidence and the rice ear instance mask connectivity integrity to obtain the fusion weight, and the fusion weight is bound to the candidate entries of grain count. Under the constraint of the fusion weight, the viewpoint-level resolvable grain count is weighted and aggregated, and the viewpoint-level occlusion compensation count is weighted and aggregated, respectively encapsulated into a grain count fusion record per ear and an occlusion compensation record, and bound to the deduplicated rice ear instance identifier and written into a traceable field.

[0103] S5: Combine the grain count fusion record, occlusion compensation record and deduplicated rice panicle instance list to count the total grain count and the effective panicle count of the whole plant, perform yield conversion and encapsulation, write the confidence boundary, and generate a rice panicle yield prediction result set.

[0104] S5.1: Align the grain count fusion record with the shading compensation record, and encapsulate the grain count accounts at the ear level to generate a double-account converged table for the grain count at the ear level;

[0105] Furthermore, a one-to-one correspondence is established by matching the grain count fusion record and the shading compensation record one by one according to the deduplicated rice ear instance identifier. For each pair of matched records, the consistency of the deduplicated rice ear instance identifier is checked and the record pairing pointer is fixed. The grain count fusion value of each ear in the grain count fusion record and the shading compensation convergence value of each ear in the shading compensation record are written into the ear-level grain count account entry according to the same field caliber. The ear-level grain count account is packaged and encapsulated, and the ear-level grain count account entries are organized into a dual-account structure with the deduplicated rice ear instance identifier as the primary key. The dual-account structure retains the grain count fusion value and the shading compensation convergence value of each ear respectively and writes them into the traceable field formed by cross-view consistency mutual verification and weight reduction fusion to support the subsequent generation of credibility boundaries. For the deduplicated rice ear instance identifier of missing paired records, gap registration is performed and null values ​​are filled in according to the gap type. The dual-account structure is sorted and aggregated according to the deduplicated rice ear instance identifier and encapsulated into an ear-level grain count dual-account aggregation table.

[0106] S5.2: Check the double-account aggregation table of grain count at the panicle level against the list of duplicate rice panicle examples one by one, count the effective panicle count of the whole plant, and aggregate the total grain count of the whole plant and the total amount of shading compensation of the whole plant to generate a statistical package of the whole plant yield.

[0107] Furthermore, the double-ledger aggregation table for ear-level grain count is traversed line by line according to the deduplicated ear instance identifier, and then compared one by one with the deduplicated ear instance list according to the deduplicated ear instance identifier. During the comparison process, entries that exist in the deduplicated ear instance list but are not found in the double-ledger aggregation table for ear-level grain count are registered as gap types. Entries that exist in the double-ledger aggregation table for ear-level grain count but are not found in the deduplicated ear instance list are registered as redundant types. Gap types and redundant types are marked according to the consistency judgment rules and used for subsequent analysis. Reliability boundary generation; the verified duplicate rice panicle instances are included in the total effective panicle count and accumulated to form the total effective panicle count. The grain count of each panicle in the panicle-level grain count double-account aggregation table is accumulated one by one to form the total grain count of the whole plant. The shading compensation aggregation value of each panicle in the panicle-level grain count double-account aggregation table is accumulated one by one to form the total shading compensation amount of the whole plant. The total effective panicle count, total grain count of the whole plant and total shading compensation amount of the whole plant are packaged and organized according to the rice plant identity and supplemented with gap type and redundancy type summary entries to generate the total yield composition statistical package.

[0108] It should be noted that the consistency judgment rule is used to mark the acceptable range of gap and redundancy types in the double-account aggregation table of ear-level grain count and the list of deduplicated rice ear instances. The consistency judgment rule is obtained by merging the view coverage of the multi-view calibration image frame set under the same acquisition batch, the statistical results of the connectivity integrity of the rice ear candidate region set, and the link breakage ratio of the cross-view related link set.

[0109] S5.3: Based on the total number of grains per plant and the number of effective ears per plant in the whole plant yield composition statistical package, the yield is converted through the yield conversion parameter table, and a yield conversion package is generated.

[0110] Furthermore, based on the overall yield composition statistical package, the total number of grains and the number of effective panicles per plant are extracted and the numerical units are verified. The total number of grains and the number of effective panicles per plant are associated with the yield conversion parameter table according to the rice plant identity to locate the corresponding conversion parameter set. The conversion parameter set covers the grain weight conversion caliber and the statistical caliber correction caliber, and parameter gap registration is performed for missing items. Based on the conversion parameter set, grain weight conversion is performed on the total number of grains per plant and panicle caliber correction is performed on the number of effective panicles per plant. Grain weight conversion is completed by the grain weight estimate corresponding to the total number of grains per plant, and out-of-bounds flags are registered for values ​​that exceed the applicable range of the conversion parameter set. Panicle caliber correction is completed by the caliber consistency value corresponding to the number of effective panicles per plant, and consistency verification failure flags are registered for abnormally high values. The grain weight estimate and caliber consistency value are merged according to the same encapsulated field structure and written into the yield conversion source pointer and conversion parameter set index to form traceable conversion entries, which are then merged and organized according to the rice plant identity to generate the yield conversion encapsulated package.

[0111] It should be noted that the yield conversion parameter table is used to provide a unified yield conversion caliber and applicable scope boundary for the total number of grains per plant and the effective number of panicles per plant. The yield conversion parameter table is formed by aggregating the measured grain weight and measured yield records of the calibration samples by using the variety type and cultivation season of the same rice-growing area as grouping keys, and solidifying the statistics of the stable interval within the group into a conversion parameter set, and registering the applicable scope and out-of-bounds judgment conditions for each conversion parameter set.

[0112] S5.4: Obtain the confidence boundary based on the total amount of whole plant shading compensation and the cross-view consistency verification and weight reduction results, and attach it to the yield conversion package for packaging and processing to generate a set of rice panicle yield prediction results.

[0113] Furthermore, based on the total amount of whole-plant shading compensation and the total number of grains per plant in the whole-plant yield composition statistical package, the shading ratio is generated and mapped to a shading uncertainty metric. At the same time, the list of deweighted members and the contribution weight of each perspective retained in the grains per panicle fusion record are aggregated and organized according to the deduplicated panicle instance identifier, and mapped to a consistency uncertainty metric based on the dispersion of the contribution weight of each perspective. The shading uncertainty metric and the consistency uncertainty metric are combined into a confidence boundary according to a unified standard and upper and lower limit pruning is performed to limit the boundary amplification caused by extreme shading or single-perspective dominance. The confidence boundary and the yield conversion package are aligned one by one according to the rice plant identity and written into the confidence boundary field of the yield conversion package. At the same time, the total amount of whole-plant shading compensation and the summary of the deweighted member list are written into the traceability field of the yield conversion package to maintain the interpretability of the results. The packaged and organized items are merged and output according to the rice plant identity to generate a rice panicle yield prediction result set.

[0114] It should be noted that the confidence boundary pointer provides a boundary description of the confidence interval for the yield conversion result corresponding to each rice plant identification in the yield conversion package. The confidence boundary is jointly determined by the shading uncertainty measure represented by the proportion of the total amount of shading compensation to the total number of grains in the whole plant and the consistency uncertainty measure represented by the dispersion of the contribution weight of each perspective in the cross-view consistency mutual verification weight reduction fusion process. It is used to characterize the strength of the constraint of the degree of shading compensation dependence and the reliability of multi-view consistency on the range of result fluctuation, and uses preset upper and lower limits to avoid abnormal amplification of the boundary due to local extreme shading or single perspective dominance.

[0115] In summary, this invention achieves high-quality, standardized image recognition input through multi-view image acquisition and calibration steps, providing a reliable data foundation for accurate segmentation and grain count estimation, and improving yield prediction accuracy; through cross-view association and deduplication steps, it achieves unique attribution of whole rice panicle instances, avoids duplicate counting, and improves the accuracy of yield prediction based on image recognition.

[0116] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the yield of a single rice panicle based on image recognition and multiple parameters, characterized in that: include, Collect multi-view images of rice plants and write them with a unified timestamp, viewpoint marker and scale reference information. Perform sharpness filtering, exposure consistency normalization and distortion correction to generate a multi-view calibrated image frame set. The candidate regions of rice ears are located from the multi-view calibration image frame set, rice ear instance segmentation is performed, and the skeleton line of the panicle axis and key points of the panicle base are extracted. At the same time, appearance features and confidence are written to generate a candidate set of rice ear instances within the view. Align the candidate set of rice panicle instances within the viewpoint to a unified coordinate system, establish cross-viewpoint association links, perform deduplication, constrain the consistency of the entire panicle's affiliation, and generate a list of deduplicated rice panicle instances and an index table of panicle regions. Based on the ear region index table, the distinguishable grain region and the strong shading region are divided, and cross-view consistency mutual verification weighted fusion is performed to generate a fusion record of the number of grains per ear and a shading compensation record. By combining the grain count fusion record, occlusion compensation record, and deduplicated rice panicle instance list, the total number of grains per plant and the number of effective panicles per plant are counted, and the yield is converted and encapsulated, written into the confidence boundary, and a rice panicle yield prediction result set is generated.

2. The image recognition based multi-parameter yield estimation method of rice panicle and plant according to claim 1, characterized in that: The process involves acquiring multi-view images of rice plants, writing unified timestamps, viewpoint markers, and scale reference information, performing sharpness filtering, exposure consistency normalization, and distortion correction, and generating a multi-view calibrated image frame set. The steps are as follows: Collect multi-view images of rice plants and write them with a unified timestamp, viewpoint marker and scale reference information to generate a multi-view original image set; Sharpness filtering is performed on the original image set from multiple perspectives, and exposure consistency normalization is performed with a unified reference frame to generate a sharp and normalized image set. By combining a clear, normalized image set with scale reference information, locating the reference geometry, performing distortion correction, and encapsulating it with a unified timestamp, a multi-view calibrated image frame set is generated.

3. The image recognition based multi-parameter yield estimation method of rice panicle and plant according to claim 2, characterized in that: The steps for locating candidate rice ear regions from a set of multi-view calibration image frames and performing rice ear instance segmentation are as follows: Based on a multi-view calibration image frame set, the rice canopy region is located, and the rice panicle response map is extracted by the slender texture of the panicle and the edge direction of the awn. At the same time, the connected regions are solidified to generate a set of candidate regions for rice panicles. The image blocks of the panicle are cropped according to the candidate region set of the panicle, and the panicle instance segmentation is performed. The overlapping areas of the leaves are separated by the consistency constraint of the panicle axis direction, and a mask set of panicle instances is generated.

4. The image recognition based multi-parameter yield estimation method of rice panicle and plant according to claim 3, characterized in that: The steps for generating a candidate set of rice ear instances within the specified viewpoint are as follows: Morphological refinement is performed on the mask set of rice panicle instances, the panicle axis skeleton line is extracted, and the panicle base key points are anchored back along the endpoints of the skeleton to generate a skeleton key point record set. Extract the outline and texture appearance features of rice ears from the skeleton key point record set, write them into the segmentation confidence, encapsulate them into candidate entries, aggregate and organize them according to the viewpoint label, and generate a candidate set of rice ear instances within the viewpoint.

5. The method for predicting the yield of a whole rice panicle based on image recognition as described in claim 4, characterized in that: The steps for aligning the candidate set of rice ear instances within a single viewpoint to a unified coordinate system and establishing cross-viewpoint association links are as follows: By using scale reference information and viewpoint markers, the key points of the panicle base and the panicle axis skeleton line in the candidate set of rice panicle instances within the viewpoint are converted to a unified coordinate system and their spatial positions are registered to generate an aligned candidate record set. Based on the aligned candidate record set, the panicle base key point is used as the anchor point. The appearance feature similarity check and the panicle axis skeleton line morphology similarity check are performed on the rice panicle instances from different perspectives to generate a cross-view association link set.

6. The method for predicting the yield of a whole rice panicle based on image recognition as described in claim 5, characterized in that: The steps for generating the list of duplicate rice panicle instances and the panicle region index table are as follows: By combining the cross-view association link set, segmentation confidence and rice ear instance mask set, representative instances of the main view and redundant view instances are obtained, and deduplication and merging are performed to generate a list of deduplicated rice ear instances. Based on the list of deduplicated rice panicle instances, the consistency constraint of panicle affiliation for the whole plant is executed through the spatial distribution of panicle base and the direction of tiller connectivity, and panicle trimming boundary pointers are written from each perspective to generate a panicle region index table.

7. The method for predicting the yield of a whole rice panicle based on image recognition as described in claim 6, characterized in that: The steps for dividing the discernible grain region and the strongly shaded region according to the ear region index table are as follows: The ear region index table is combined with the multi-view calibration image frame set, and ear-by-ear positioning and cropping are performed to generate a multi-view ear image block set. In the vicinity of the spikelet image block set from multiple perspectives, the spikelet axis direction is cropped to construct candidate search bands for spikelets, and the discernible grain region and the strongly occluded region are divided to generate a layered spikelet region slice set.

8. The method for predicting the yield of a whole rice panicle based on image recognition as described in claim 7, characterized in that: The steps for generating the fused record of grains per ear and the shading compensation record are as follows: Grain candidate profiles are extracted from the resolvable grain regions in the layered spikelet region slice set, and the grain candidate points are fixed to generate resolvable grain count records. In the strongly occluded region of the layered spikelet area slice set, the particulate response is extracted and aggregated into occlusion compensation count. At the same time, combined with the resolvable grain count record, a candidate record set of grain count within the view is generated. Align the candidate record set of grain count within the view with the deduplicated rice panicle instance, and perform cross-view consistency mutual verification weighted fusion to generate a fused record of grain count per panicle and an occlusion compensation record.

9. The method for predicting the yield of a whole rice panicle based on image recognition as described in claim 8, characterized in that: The steps for combining the grain count fusion record, shading compensation record, and deduplicated rice panicle example list to calculate the total grain count and effective panicle count of the whole plant are as follows: Align the grain count fusion record with the shading compensation record, and encapsulate the grain count accounts at the ear level to generate a double-account converged table for the grain count at the ear level. The double-account table of grain count at the panicle level is checked against the list of deduplicated rice panicle examples one by one. The number of effective panicles per plant is counted, and the total number of grains per plant and the total amount of shading compensation per plant are aggregated to generate a statistical package of total plant yield.

10. The method for predicting the yield of a whole rice panicle based on image recognition as described in claim 9, characterized in that: The steps for generating the rice panicle yield prediction result set are as follows: Based on the total number of grains and the number of effective ears per plant in the whole plant yield composition statistical package, yield conversion is performed through the yield conversion parameter table to generate a yield conversion package. The credibility boundary is obtained by combining the total amount of whole-plant shading compensation with the cross-view consistency verification and weight reduction results, and then attached to the yield conversion package for encapsulation and processing to generate a set of rice panicle yield prediction results.

Images