Intelligent Rice Variety Classification System and Method Based on Image Recognition

Abstract

This invention discloses an intelligent rice variety classification system and method based on image recognition, aiming to address the technical pain points of existing rice variety classification methods, such as reliance on manual experience, low efficiency, large errors, incomplete feature extraction, insufficient classification accuracy, and poor adaptability. The system includes a dual-optical-path controllable image acquisition module, a rice-specific preprocessing module, a dual-branch feature decoupling and fusion extraction module, a few-sample classification inference module, and a result output and verification module. The classification method sequentially performs image acquisition, preprocessing, feature extraction, variety identification, and result verification. Dual-path images of rice varieties are acquired through dual optical paths, and after preprocessing, the dual-branch module extracts fused features, which are then identified by the few-sample classification module. This invention achieves non-destructive, rapid, and accurate intelligent classification of rice varieties, improves the accuracy of distinguishing highly similar rice varieties, reduces usage and maintenance costs, adapts to multiple scenarios, and is applicable to fields such as agricultural seed testing and breeding research.

Description

Technical Field

[0001] This invention relates to the field of rice seed classification, specifically to an intelligent rice seed classification system and method based on image recognition. Background Technology

[0002] Accurate classification of rice varieties is a crucial step in agricultural seed testing, breeding research, and grain production, directly impacting the effectiveness of rice seed selection, variety development, and yield improvement. Currently, existing rice variety classification methods primarily rely on manual observation and experience. Operators must use their accumulated experience to distinguish different rice varieties by observing characteristics such as grain shape, husk texture, and lemma tip morphology. This method is not only inefficient but also susceptible to the influence of operator subjectivity, experience level, and fatigue, resulting in significant classification errors. It is particularly difficult to distinguish closely related rice varieties with high genetic similarity, failing to meet the needs of large-scale, high-precision rice variety classification.

[0003] With the development of machine vision technology, some image recognition-based rice classification schemes have emerged, but existing schemes still have significant limitations. Most existing schemes use single-path image acquisition, which can only obtain a single visual image of the rice seed and cannot capture the microscopic distinguishing features of the rice seed, resulting in incomplete feature extraction and difficulty in improving classification accuracy. At the same time, the feature extraction and classification inference architecture of existing schemes are poorly designed, with weak ability to distinguish highly similar rice seeds, and poor model adaptability. When adding new rice seed categories, the model needs to be fully retrained, resulting in high usage and maintenance costs and difficulty in adapting to multi-scenario detection needs.

[0004] Therefore, developing an image recognition-based intelligent classification system and method for rice varieties that can solve the above-mentioned pain points and achieve rapid, accurate, and non-destructive intelligent classification of rice varieties has become an urgent need in the field of agricultural seed testing. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the defects of the above-mentioned technologies and provide an intelligent classification system and method for rice varieties based on image recognition.

[0006] This invention discloses an intelligent rice variety classification system and method based on image recognition. The intelligent rice variety classification system based on image recognition includes a dual-optical-path controllable image acquisition module, a rice variety-specific preprocessing module, a dual-branch feature decoupling and fusion extraction module, a few-sample classification reasoning module, and a result output and verification module. The above modules are connected in sequence. The dual-optical-path controllable image acquisition module simultaneously acquires two target images of the rice variety to be detected. Each module sequentially completes the preprocessing, feature extraction, rice variety identification, result output, and verification operations of the corresponding image.

[0007] Preferably, the dual-optical-path controllable image acquisition module includes a coaxial white light source unit, a side-positioned linearly polarized light source unit, an image acquisition unit, and an automatic alignment stage unit; the coaxial white light source unit and the side-positioned linearly polarized light source unit provide corresponding illumination for the acquisition of two target images, and the automatic alignment stage unit is used to carry the rice seeds to be tested and complete the posture normalization placement of the rice seeds to be tested.

[0008] Preferably, the rice seed-specific preprocessing module is used to sequentially perform adaptive segmentation of the target and background, noise removal, rice seed principal axis pose normalization, texture detail enhancement, and feature dimension alignment processing on the received two target images, and output standardized dual-input images.

[0009] Preferably, the dual-branch feature decoupling and fusion extraction module includes an appearance morphology feature extraction branch, a microtexture feature extraction branch, and a rice seed key feature forced activation unit; the appearance morphology feature extraction branch and the microtexture feature extraction branch respectively extract rice seed features of corresponding dimensions, and the rice seed key feature forced activation unit is used to screen and strengthen feature channels that are strongly correlated with rice varieties.

[0010] Preferably, the few-shot classification reasoning module adopts an architecture that combines metric learning and incremental learning to complete the category identification and confidence calculation of the rice variety to be detected, while realizing the rapid adaptation of new rice variety categories without the need to retrain the model.

[0011] On the other hand, an intelligent classification method for rice varieties based on image recognition is provided. The method sequentially performs the entire process of rice seed image acquisition, image preprocessing, feature extraction, variety identification, and result output verification. It synchronously acquires two target images of the rice seed to be detected through a dual-optical-path acquisition architecture, and completes the decoupling extraction and fusion of rice seed features through a dual-branch architecture, finally outputting the variety identification result of the rice seed to be detected.

[0012] Preferably, in the step of rice seed image acquisition, illumination is provided by a coaxial white light source and a side-positioned linearly polarized light source, and the apparent morphology white light image and the micro-texture polarized image of the hull of the rice seed to be tested are acquired simultaneously, while the posture normalization of the rice seed to be tested is completed.

[0013] Preferably, in the image preprocessing step, the two acquired target images are sequentially processed with adaptive segmentation of the target and background, noise removal, rice seed principal axis pose normalization, texture detail enhancement, and feature dimension alignment to output a standardized dual-input image.

[0014] Preferably, in the feature extraction step, the macroscopic morphological features and microscopic identification features of the rice variety to be detected are extracted by a dual-branch architecture, while feature channels that are strongly correlated with rice variety are screened and strengthened, thereby completing the decoupled extraction and adaptive fusion of rice variety features.

[0015] Preferably, in the steps of category identification and result output verification, the category identification and confidence calculation of the rice variety to be detected are completed, the rice variety identification result is output, and a manual verification process is triggered for identification results that are lower than the preset confidence threshold.

[0016] The advantages of this invention compared to the prior art are: First, the system deeply integrates modules such as dual-optical-path image acquisition, dedicated preprocessing, decoupled feature fusion extraction, and few-shot classification inference, achieving fully automated intelligent classification. This fundamentally overcomes reliance on human experience and significantly improves classification efficiency, accuracy, and consistency. Second, it innovatively adopts a dual-optical-path acquisition mechanism to simultaneously acquire information on the apparent morphology of rice seeds and the microscopic texture of the husks. Through dual-path feature fusion, it greatly enriches the dimensions of distinguishing features and significantly improves the accuracy of distinguishing highly similar closely related rice species. Third, through decoupled feature extraction and few-shot classification inference design, the system can quickly adapt to new rice varieties while ensuring the accuracy of feature extraction, without requiring full retraining of the model. This reduces maintenance costs and enhances the system's generalization ability. Furthermore, the entire process uses non-contact image recognition, achieving non-destructive classification and protecting the viability of rice seeds. The system also supports portable terminal integration and distributed deployment, flexibly adapting to various practical application scenarios. Finally, relying on standardized preprocessing procedures and classification result verification mechanisms, the standardization, reliability, and full-process traceability of the output results are further guaranteed, comprehensively improving the system's practicality and reliability. Attached Figure Description

[0017] Figure 1 An overall architecture diagram of an image recognition-based intelligent classification system for rice varieties provided in an embodiment of the present invention; Figure 2 A schematic diagram of the internal structure of a dual-optical-path controllable image acquisition module of an image recognition-based intelligent rice variety classification system provided in an embodiment of the present invention; Figure 3 A schematic diagram of the internal structure of a dual-branch feature decoupling fusion extraction module in an image recognition-based intelligent rice variety classification system provided in an embodiment of the present invention; Figure 4 This is an overall flowchart of an intelligent rice variety classification method based on image recognition. Figure 5 A flowchart of image preprocessing steps for an intelligent rice variety classification method based on image recognition, provided in an embodiment of the present invention; Figure 6 The flowchart illustrates the feature extraction and classification reasoning process of an intelligent rice variety classification method based on image recognition, as provided in this embodiment of the invention. Detailed Implementation

[0018] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0019] The present invention discloses an intelligent rice seed classification system and method based on image recognition, which is applied to the fields of agricultural seed testing and machine vision technology. The core is to realize non-destructive, rapid and high-precision intelligent classification of rice seed varieties, and solve the technical pain points of existing technologies that rely on human experience for rice seed classification, resulting in low efficiency and large subjective errors, as well as the insufficient accuracy of existing image recognition schemes in distinguishing closely related rice varieties with high genetic similarity and weak model generalization ability.

[0020] The image recognition-based intelligent rice variety classification system comprises, in sequence, a dual-optical-path controllable image acquisition module, a rice-specific preprocessing module, a dual-branch feature decoupling and fusion extraction module, a few-sample classification inference module, and a result output and verification module. In practical implementation, these five modules can be integrated into a single portable testing terminal, enabling on-site testing in various scenarios such as fields, seed warehouses, and testing laboratories. Alternatively, a distributed deployment approach can be adopted, with the dual-optical-path controllable image acquisition module deployed at the testing site and the remaining modules deployed on a cloud server, connected via wireless communication to meet the centralized testing needs of large batches of rice samples.

[0021] The dual-optical-path controllable image acquisition module is used to simultaneously acquire two target images of the rice seed to be detected, and at the same time complete the posture normalization of the rice seed to be detected, eliminating detection errors caused by shooting angle and stacking obstruction. The dual-optical-path controllable image acquisition module includes a coaxial white light source unit, a side-positioned linearly polarized light source unit, an image acquisition unit, and an automatic alignment stage unit. In practical implementation, the coaxial white light source unit uses a ring-shaped arrangement of white light-emitting diodes, fixedly installed directly below the lens of the image acquisition unit. The light emission direction of the coaxial white light source unit is completely coaxial with the optical axis of the lens of the image acquisition unit, providing a uniform and shadow-free illumination environment for the acquisition of white light images of the apparent morphology, clearly presenting the overall outline, grain shape, and other macroscopic appearance information of the rice seed to be detected. The side-positioned linearly polarized light source unit is fixedly installed to the side of the automatic alignment stage unit. The light emission direction of the side-positioned linearly polarized light source unit forms a preset fixed angle with the bearing plane of the automatic alignment stage unit. The light emission end of the side-positioned linearly polarized light source unit has a built-in linear polarizer, which can output linearly polarized light to provide directional illumination for the acquisition of polarized images of the glumes' micro-textures, highlighting microscopic identification information such as glumes' surface texture and endosperm crystal distribution that cannot be clearly captured by conventional white light, thus solving the defect of existing single-light-path acquisition methods that cannot obtain the core microscopic identification features of rice seeds from the source of acquisition. The image acquisition unit employs an industrial-grade area scan camera, which can synchronously acquire images under two different lighting conditions via a trigger signal. This ensures that the acquisition angle, acquisition range, and acquisition time of the two target images are completely consistent, avoiding image misalignment and feature mismatch caused by different acquisition times. The automatic alignment stage unit uses high-transmittance, non-slip tempered glass as the supporting surface to support single or multiple orderly arranged rice seeds to be tested. The automatic alignment stage unit has a built-in electric alignment mechanism that can adjust the main axis of the rice seeds to be tested to a preset uniform direction through mechanical limiting, completing the posture normalization of the rice seeds to be tested. This completely eliminates detection errors caused by differences in shooting angles and rice seed stacking obstruction, ensuring the consistency of input data in the subsequent feature extraction process.

[0022] The rice seed-specific preprocessing module is signal-connected to the dual-optical-path controllable image acquisition module. This module preprocesses the two target images output by the dual-optical-path controllable image acquisition module, outputting standardized dual-input images. The preprocessing operations sequentially include adaptive segmentation of the target and background, noise removal, rice seed principal axis pose normalization, texture detail enhancement, and feature dimension alignment. In actual implementation, the adaptive segmentation of the target and background uses an adaptive threshold segmentation algorithm. Based on the pixel value distribution of local image regions, it automatically calculates the segmentation threshold, accurately separating the target area of ​​the rice seed to be detected from the background area of ​​the image. Invalid pixels in the background area are removed, leaving only the pixels in the target area where the rice seed is located, reducing interference from invalid data in subsequent processing. The noise removal operation uses a median filtering algorithm to smooth the segmented image, removing salt-and-pepper noise and random Gaussian noise generated during image acquisition, preventing noise signals from obscuring the subtle identification features of the rice seed. The rice seed principal axis pose normalization operation, based on the target... The target region contour of the rice seed is calculated, and the principal axis direction of the grains of the rice seed to be detected is calculated. The image is rotated so that the principal axis of the grains is completely consistent with the preset horizontal direction, further enhancing the pose consistency of all rice seed images to be detected. The texture detail enhancement operation adopts a contrast-limited adaptive histogram equalization algorithm to directionally enhance the contrast of the texture region of the rice seed in the image, highlighting the subtle texture features of the rice seed and improving the accuracy of subsequent feature extraction. The feature dimension alignment processing adjusts the size, resolution, and pixel value distribution range of the two target images to be completely consistent, ensuring that the feature dimensions of the two images are completely matched, providing a unified and standardized input basis for subsequent dual-branch feature extraction.

[0023] After noise removal, in order to unify the pixel value distribution range of images acquired under different lighting conditions and improve the stability and convergence speed of subsequent model processing, this implementation method performs pixel value normalization on the images. The normalization calculation formula used is as follows: in, These are the original pixel values ​​of the image after noise removal. It is the minimum value among all pixel values ​​of the currently processed image. It is the maximum value among all pixel values ​​in the currently processed image. These are the normalized pixel values ​​after normalization. The purpose of this formula is to uniformly map the pixel values ​​of different images to the range of 0 to 1, completely eliminating the differences in pixel value distribution caused by different acquisition batches and different lighting intensities, ensuring the consistency of data distribution of all input images, and improving the stability and reliability of the subsequent feature extraction process.

[0024] The dual-branch feature decoupling and fusion extraction module is signal-connected to the rice seed-specific preprocessing module. This module includes an appearance morphology feature extraction branch, a micro-texture feature extraction branch, and a rice seed key feature forced activation unit. The appearance morphology feature extraction branch extracts the macroscopic morphological features of the rice seed to be detected, the micro-texture feature extraction branch extracts the microscopic identification features of the rice seed, and the rice seed key feature forced activation unit filters feature channels strongly correlated with rice variety, applies adaptive weight enhancement to these channels, and suppresses irrelevant noise channels. This process ultimately completes the decoupling extraction and adaptive fusion of the dual features, outputting a rice variety identification fusion feature. In practical implementation, the apparent morphology feature extraction branch receives the apparent morphology white light image from the standardized dual-channel input image output by the rice seed-specific preprocessing module. It extracts the macroscopic morphological features of the rice seed under test through multi-layer convolution operations. These macroscopic morphological features include the grain length and width, lemma tip morphology, and grain contour curvature. The microscopic texture feature extraction branch receives the glume microscopic texture polarization image from the standardized dual-channel input image output by the rice seed-specific preprocessing module. It extracts the microscopic identification features of the rice seed under test through multi-layer convolution operations. These microscopic identification features include the glume texture density and endosperm crystal form distribution. Through the decoupled design of the two branches, the macroscopic morphological features and microscopic identification features of the rice seed can be extracted independently and specifically, avoiding mutual interference between features of different dimensions and significantly improving the accuracy and specificity of feature extraction.

[0025] The key feature forced activation unit for rice varieties is connected to the signals of the apparent morphology feature extraction branch and the microtexture feature extraction branch, respectively. In actual implementation, the key feature forced activation unit for rice varieties calculates the correlation value between each feature channel and the rice variety classification task using a pre-trained gradient-weighted class activation mapping algorithm. Based on the correlation value, feature channels strongly correlated with rice varieties are selected, and then adaptive weight enhancement is applied to the strongly correlated feature channels, while weight suppression is applied to irrelevant noise channels. To achieve adaptive weight adjustment of feature channels, this implementation adopts the following weight calculation and enhancement formula: in, This represents the weight enhancement coefficient corresponding to the nth feature channel. The preset weight adjustment coefficient is used to control the overall magnitude of weight enhancement. The correlation value between the nth feature channel and the rice variety classification task represents the degree of correlation. A higher correlation value indicates a greater contribution of that feature channel to the rice variety classification. This formula assigns higher weights to feature channels strongly correlated with rice varieties, specifically strengthening the expression intensity of core distinguishing features, while reducing the weights of noise channels irrelevant to the classification task and suppressing interference from ineffective features. This addresses the core pain point of existing general models' insufficient distinguishability of highly similar closely related rice varieties from the feature extraction level. After completing the feature channel weight enhancement processing, the dual-branch feature decoupling and fusion extraction module adaptively splices and fuses the features output by the appearance morphology feature extraction branch and the microtexture feature extraction branch, outputting the final rice variety identification fusion feature.

[0026] The few-shot classification inference module is signal-connected to the dual-branch feature decoupling fusion extraction module. The few-shot classification inference module receives the rice variety identification fusion features output by the dual-branch feature decoupling fusion extraction module, completes the variety identification and confidence calculation of the rice variety to be detected, and outputs the rice variety identification result. The few-shot classification inference module adopts an architecture combining metric learning and incremental learning. In actual implementation, the metric learning architecture learns the feature distance measurement method between different rice varieties through pre-training, which can accurately distinguish highly similar rice varieties with minimal feature differences, significantly improving classification accuracy. The incremental learning architecture enables rapid adaptation to new rice varieties. When it is necessary to add the classification ability of a certain rice variety to the model, only a small number of sample images of the new rice variety are needed to complete the model adaptation update, without requiring full retraining of the model. This significantly reduces the adaptation cost of new rice varieties and effectively avoids the model forgetting problem caused by full retraining, significantly improving the model's generalization ability and applicability.

[0027] The result output and verification module is signal-connected to the few-shot classification inference module. This module receives the rice variety identification results output by the few-shot classification inference module, stores and displays the identification results, and triggers a manual verification process for identification results below a preset confidence threshold. In actual implementation, the result output and verification module can display, in real time, complete information such as the collected image of the rice variety to be detected, the extracted core identification features, the identified rice variety, and the corresponding classification confidence level through a supporting display device. It can also store all detection results in a local database, supporting subsequent historical data queries, statistical analysis, and report export operations. In this embodiment, the result output and verification module uses the following judgment logic to trigger the manual verification process, and the corresponding judgment formula is: in, The confidence score value corresponding to the classification result output by the few-shot classification inference module. A confidence threshold is preset for operators. This formula sets a reliability verification threshold for the classification results. When the formula is true, meaning the confidence value of the classification result is lower than the preset confidence threshold, it indicates that the reliability of the current classification result is insufficient. The result output and verification module automatically triggers the manual verification process, reminding operators to manually verify the rice seed to be tested through a pop-up prompt, thus completely avoiding misclassification problems and ensuring the accuracy and reliability of the final classification results.

[0028] This invention also discloses an intelligent rice variety classification method based on image recognition. This method is applied to the aforementioned intelligent rice variety classification system based on image recognition. The method sequentially performs the entire process of rice seed image acquisition, image preprocessing, feature extraction, variety identification, and result output verification. It simultaneously acquires two target images of the rice variety to be detected through a dual-optical-path acquisition architecture, and completes the decoupled extraction and fusion of rice seed features through a dual-branch architecture, finally outputting the variety identification result of the rice variety to be detected. The specific implementation steps of the intelligent rice variety classification method based on image recognition are as follows: Step one involves constructing a standard rice seed image dataset covering multiple rice varieties, and completing image annotation and augmentation operations within the dataset. In practice, this standard rice seed image dataset contains images of various rice varieties, covering samples from different origins, maturity levels, and storage durations. It also includes images of multiple genetically similar closely related rice varieties, ensuring the dataset's comprehensiveness and representativeness, providing a reliable data foundation for model training. The annotation operation assigns the corresponding true label for each rice variety to each image in the dataset, providing accurate supervisory signals for supervised training of the model. The augmentation operation uses conventional image augmentation methods such as image rotation, horizontal flipping, vertical flipping, brightness adjustment, and contrast adjustment to expand the sample size of the dataset, avoiding overfitting during model training and further improving the model's generalization ability.

[0029] Step two involves simultaneously acquiring a white light image of the apparent morphology of the rice seed to be tested and a polarized image of the micro-texture of the hull through a dual-path controllable image acquisition module. In practice, the operator places individual rice seeds onto the bearing plane of the automatically aligned stage unit of the dual-path controllable image acquisition module. After starting the device, the automatically aligned stage unit uses a built-in electric alignment mechanism to normalize the orientation of the rice seeds, adjusting the main axis of all the seeds to a preset uniform direction. Subsequently, corresponding illumination is provided by the coaxial white light source unit and the side-mounted linearly polarized light source unit, respectively. The acquisition of both images is completed through the synchronous trigger signal of the image acquisition unit, obtaining a white light image of the apparent morphology of the rice seed and a polarized image of the micro-texture of the hull, ensuring that the acquisition angle and range of the two images are completely consistent.

[0030] Step three involves preprocessing the acquired white light image of the apparent morphology and the polarization image of the glumes' micro-texture using a rice-specific preprocessing module, outputting standardized dual-path input images. In practice, the preprocessing operations sequentially include adaptive segmentation of the target and background, noise removal, rice seed principal axis pose normalization, texture detail enhancement, and feature dimension alignment. These preprocessing operations remove invalid background information and acquisition noise from the images, unify the size, resolution, and pixel value distribution range of all images, and output standardized dual-path input images, providing high-quality, highly consistent input data for subsequent feature extraction.

[0031] Step four involves inputting the standardized dual-path input image into the dual-branch feature decoupling and fusion extraction module. The core features for rice variety identification are enhanced through the rice seed key feature forced activation unit, completing the decoupling extraction and adaptive fusion of the dual-path features, and outputting the rice variety identification fusion feature. In actual implementation, the macroscopic morphological features of the rice seed to be detected are extracted through the apparent morphology feature extraction branch, and the microscopic identification features of the rice seed to be detected are extracted through the microscopic texture feature extraction branch. The rice seed key feature forced activation unit filters feature channels strongly correlated with the rice variety, applies adaptive weight enhancement to the strongly correlated feature channels, and suppresses irrelevant noise channels. After feature enhancement processing, the features extracted from the dual branches are adaptively spliced ​​and fused to output the final rice variety identification fusion feature.

[0032] Step 5: The few-shot classification inference module receives the rice variety identification fusion features, completes the variety identification and confidence calculation of the rice variety to be detected, and outputs the rice variety identification result. In actual implementation, the few-shot classification inference module adopts an architecture combining metric learning and incremental learning. Based on the input rice variety identification fusion features, it calculates the feature distance between the rice variety to be detected and the standard rice variety, matches the rice variety corresponding to the rice variety to be detected, and calculates the confidence value corresponding to the classification result, outputting the final rice variety identification result. When it is necessary to add the classification ability of rice varieties to the model, only a small number of sample images of the new rice varieties need to be input to complete the model adaptation update, without the need to retrain the model completely.

[0033] Step six involves receiving the rice variety identification results through the result output and verification module, storing and displaying the identification results, and triggering a manual verification process for identification results below a preset confidence threshold. In actual implementation, the result output and verification module displays the collected image of the rice variety to be detected, the identified rice variety, and the corresponding classification confidence level in real time through a display device. Simultaneously, it stores all detection results in a local database for easy subsequent querying and statistics. When the confidence level of the classification result falls below the preset confidence threshold, the result output and verification module automatically triggers a manual verification process, reminding the operator to perform manual verification to ensure the accuracy of the final classification result.

[0034] This invention, through the design of the aforementioned system and method, completely overcomes the technical limitations of existing single image recognition schemes. It employs a dual-optical-path acquisition architecture to simultaneously acquire the macroscopic appearance and microscopic texture features of rice varieties, enriching the identification feature dimensions of rice varieties from the data source. Through a dual-branch decoupled and fused feature extraction architecture, combined with a forced activation unit for key rice variety features, it directionally strengthens core identification features strongly correlated with rice variety types, significantly improving the accuracy of distinguishing highly similar closely related rice varieties. Simultaneously, the use of a few-sample incremental learning architecture greatly reduces the adaptation cost for new rice variety types, enhances the model's generalization ability and applicability, and enables non-destructive, rapid, and high-precision intelligent classification of rice varieties, possessing strong industrial application value and promotional significance.

[0035] The advantages of this invention compared to existing technologies are as follows: The intelligent rice seed classification system and method based on image recognition disclosed in this invention have significant beneficial effects compared to existing technologies. Firstly, it effectively solves the pain point of existing rice seed classification relying on human experience. Through the collaborative work of a dual-light-path controllable image acquisition module, a rice seed-specific preprocessing module, a dual-branch feature decoupling and fusion extraction module, a few-sample classification reasoning module, and a result output and verification module, it achieves automated intelligent rice seed classification, eliminating the need for highly experienced professionals, significantly improving classification efficiency, avoiding subjective human errors, and ensuring classification accuracy and consistency. Secondly, it overcomes the limitations of single-light-path acquisition. The dual-light-path controllable image acquisition module simultaneously acquires the white light image of the apparent morphology of the rice seed to be detected and the polarized image of the micro-texture of the hull. Combined with the dual-branch feature decoupling and fusion extraction module, it enriches the dimensions of rice seed identification features, effectively improving the differentiation accuracy of highly similar closely related rice varieties. Third, the feature extraction and classification architecture is optimized. The dual-branch decoupled design improves the accuracy of feature extraction, and the few-sample classification inference module can quickly adapt to new rice varieties without requiring full model retraining, reducing usage and maintenance costs and improving the system's generalization ability. Fourth, non-contact image recognition is adopted to achieve non-destructive classification of rice varieties, ensuring the subsequent use value of rice varieties. The system can be integrated into portable terminals or deployed in a distributed manner, adapting to multi-scenario detection needs and demonstrating strong practicality. Fifth, through the standardized operation and result output of the rice variety-specific preprocessing module and the verification mechanism of the verification module, the reliability and standardization of classification results are further ensured, and the traceability of the classification process is improved.

[0036] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A smart rice variety classification system based on image recognition, characterized in that: The image recognition-based intelligent rice variety classification system includes a dual-optical-path controllable image acquisition module, a rice variety-specific preprocessing module, a dual-branch feature decoupling and fusion extraction module, a few-sample classification reasoning module, and a result output and verification module. The above modules are connected in sequence. The dual-optical-path controllable image acquisition module simultaneously acquires two target images of the rice variety to be detected. Each module sequentially completes the preprocessing, feature extraction, rice variety identification, result output, and verification operations of the corresponding image.

2. The intelligent rice variety classification system based on image recognition according to claim 1, characterized in that: The dual-optical-path controllable image acquisition module includes a coaxial white light source unit, a side-positioned linearly polarized light source unit, an image acquisition unit, and an automatic alignment stage unit. The coaxial white light source unit and the side-positioned linearly polarized light source unit provide corresponding illumination for the acquisition of two target images, and the automatic alignment stage unit is used to carry the rice seeds to be tested and complete the posture normalization of the rice seeds to be tested.

3. The intelligent rice variety classification system based on image recognition according to claim 1, characterized in that: The rice seed-specific preprocessing module is used to sequentially perform adaptive segmentation of the target and background, noise removal, rice seed principal axis pose normalization, texture detail enhancement, and feature dimension alignment on the two received target images, and output standardized dual-input images.

4. The intelligent rice variety classification system based on image recognition according to claim 1, characterized in that: The dual-branch feature decoupling and fusion extraction module includes an appearance morphology feature extraction branch, a microtexture feature extraction branch, and a rice seed key feature forced activation unit. The appearance morphology feature extraction branch and the microtexture feature extraction branch respectively extract rice seed features of corresponding dimensions, and the rice seed key feature forced activation unit is used to screen and strengthen feature channels that are strongly correlated with rice varieties.

5. The intelligent rice variety classification system based on image recognition according to claim 1, characterized in that: The few-shot classification reasoning module adopts an architecture that combines metric learning and incremental learning to complete the category identification and confidence calculation of the rice varieties to be detected, while realizing the rapid adaptation of new rice varieties without the need to retrain the model.

6. A method for intelligent classification of rice varieties based on image recognition, characterized in that: The image recognition-based intelligent classification method for rice varieties is applied to the image recognition-based intelligent classification system for rice varieties as described in any one of claims 1 to 5. The method sequentially performs the entire process of rice seed image acquisition, image preprocessing, feature extraction, variety identification, and result output verification. It simultaneously acquires two target images of the rice variety to be detected through a dual-optical-path acquisition architecture, and completes the decoupling extraction and fusion of rice seed features through a dual-branch architecture, finally outputting the variety identification result of the rice variety to be detected.

7. The intelligent rice variety classification method based on image recognition according to claim 6, characterized in that: In the rice seed image acquisition step, illumination is provided by a coaxial white light source and a side-positioned linearly polarized light source, and the apparent morphology white light image and the micro-texture polarized image of the hull of the rice seed to be detected are acquired simultaneously, while the posture normalization of the rice seed to be detected is completed.

8. The intelligent rice variety classification method based on image recognition according to claim 6, characterized in that: In the image preprocessing step, the two acquired target images are sequentially processed for adaptive segmentation of the target and background, noise removal, rice seed principal axis pose normalization, texture detail enhancement, and feature dimension alignment, and a standardized dual-input image is output.

9. The intelligent rice variety classification method based on image recognition according to claim 6, characterized in that: In the feature extraction step, the macroscopic morphological features and microscopic identification features of the rice variety to be detected are extracted through a dual-branch architecture. At the same time, feature channels that are strongly correlated with rice varieties are screened and strengthened to complete the decoupled extraction and adaptive fusion of rice variety features.

10. The intelligent rice variety classification method based on image recognition according to claim 6, characterized in that: In the steps of category identification and result output verification, the category identification and confidence calculation of the rice variety to be detected are completed, the rice variety identification result is output, and the manual verification process is triggered for the identification result that is lower than the preset confidence threshold.

Images