Multi-index online detection system for edible fungi based on multi-modal image recognition

Abstract

The application discloses a multi-index online detection system for edible mushrooms based on multi-modal image recognition and mainly relates to the field of edible mushroom index detection. The application is aimed at the existing edible mushroom detection method, which only uses static detection without setting a dynamic detection process, and cannot identify problem bacteria bodies that have no obvious defects on the surface but have internal mold risk. The application precisely matches the conveying rhythm under different speed gradients of the production line through an image recognition rhythm matching unit, avoids image motion artifacts, missed shooting of bacteria bodies, detection omission problems under high-speed working conditions, and redundancy problems of massive invalid image data under low-speed working conditions, and fuses RGB texture images and super-high-precision micro-depth images through a boundary segmentation unit to precisely segment the adhered edible mushroom individuals, avoid morphological misjudgment of the bacteria bodies and detection omission problems. In addition, the application integrates multi-modal time sequence image data through a hidden disease risk identification unit to realize dynamic time sequence identification of internal hidden mold and early diseases of the edible mushrooms.

Description

Technical Field

[0001] This invention relates to the field of edible fungi index detection technology, specifically an online detection system for multiple indicators of edible fungi based on multimodal image recognition. Background Technology

[0002] With the rapid popularization of industrialized and large-scale production models for edible fungi, automated sorting and quality testing have become core links in improving the quality and efficiency of the edible fungi industry. Traditional edible fungi testing methods have long relied on manual visual identification and sampling inspection, which have many defects such as low testing efficiency, large subjective errors, inconsistent sorting standards, and high rates of missed and false detections, and cannot adapt to the production pace of modern high-speed production lines.

[0003] In other words, during the multi-index testing of edible fungi, it is difficult to match the rhythm of multimodal image acquisition equipment with the production line and to perform boundary segmentation, resulting in multimodal images not being fully and effectively identified. When using multimodal image recognition, it is only through static detection without setting up a dynamic detection process, which fails to identify problematic fungi that have no obvious defects on the surface but have a risk of mold growth inside.

[0004] To address the aforementioned technical shortcomings, we propose an online detection system for multiple indicators of edible fungi based on multimodal image recognition. This system aims to accurately detect multiple explicit indicators of edible fungi, such as morphology, size, whiteness, surface defects, and maturity. Simultaneously, it can identify latent mold risks in advance and complete batch traceability, achieving proactive identification and control of detection interference risks throughout the entire process. Summary of the Invention

[0005] The purpose of this invention is to solve the problems mentioned above by proposing an online detection system for multiple indicators of edible fungi based on multimodal image recognition.

[0006] The objective of this invention can be achieved through the following technical solution: an online detection system for multiple indicators of edible fungi based on multimodal image recognition, the system comprising: The image recognition beat matching unit is used to match the image acquisition beat with the production line beat during the detection of indicators in the edible fungi production line, so as to adapt to the production line; to conduct idling experiment simulation, to perform beat adaptation analysis in the speed change stage and the speed continuous stage, and to adjust and set the matching beat point based on the analysis results; The boundary segmentation unit is used to distinguish the boundaries of edible fungi during multimodal image recognition. Specifically, it uses the morphology of the current variety of edible fungi as the recognition logic to perform image recognition, collect multimodal image data, analyze the data to obtain suspected boundary positions and non-boundary positions, and screen the boundary positions based on the feature comparison of suspected boundary positions to determine the boundary positions and construct the boundary. At the same time, it identifies the positions of impurities and uploads them to the system for multi-index collection and detection. The latent disease risk identification unit is used to conduct phased and continuous detection of edible fungi. It sets dynamic detection periods, collects multimodal time-series image datasets, and the multimodal time-series image datasets include RGB texture images, hyperspectral images, and infrared temperature images of edible fungi during the dynamic detection period. It also performs cooperative analysis and detection based on the multimodal time-series image datasets to determine moldy areas and uploads them to the system.

[0007] Furthermore, the idling experiment simulation content in the image recognition beat matching unit is as follows: Record the conveying speed of the edible fungus production line, determine the range of conveying speed based on the current production line output target, determine the image acquisition points, set up standard reference objects on the production line, and conduct simulated idling experiments.

[0008] Furthermore, the specific process of beat adaptation analysis is as follows: Static data was collected from a standard reference object during the simulated idling of the production line, and the texture of the standard reference object and the surface of the production line was recorded. The production line operates within the conveying speed range set for each stage, and images are acquired through image acquisition points to construct production line image time sequence data. If the actual number of times the standard reference object passes through the production line image time-series data is inconsistent with the number of times the image is recognized, it indicates that the image recognition cycle is not suitable for the current conveyor speed adjustment stage. The cycle point of the current image acquisition equipment is adjusted, the time nodes of production line speed change are recorded, the speed deviation span of each node is identified, speed change redundancy is set according to the current speed deviation span, a speed fluctuation span standard is formed, the node of reference object in the production line passing through the image acquisition node is re-determined according to the speed fluctuation span standard, and the cycle point of the image acquisition equipment is re-planned accordingly.

[0009] Furthermore, when the conveying speed is not suitable, the identification set is extracted and divided into a front set and a back set with images containing standard reference objects. If the clarity of the surface texture of the pipeline in the collected images in the front set deviates from that in the static collection state, and the clarity deviation continues to worsen as time increases, a hysteresis compensation signal is generated and uploaded to the system. If artifacts gradually appear between the images acquired in the subsequent set and the standard reference image, a continuous adaptation signal is generated and uploaded to the system. After receiving the signal, the system increases the shooting frequency of the current image acquisition device and increases the number of beat points while keeping the current acquisition cycle length unchanged, so as to ensure that the beat points of the image acquisition device are adapted to the reference objects in the production line through the beat.

[0010] Furthermore, the multimodal image data in the boundary segmentation unit consists of: the RGB texture image and the ultra-high precision micro-depth image of edible fungi within the boundary segmentation image.

[0011] Furthermore, the methods for obtaining suspected boundary locations and non-boundary locations are as follows: Based on the acquisition of multimodal image data, the system collects minute depth differences in the adhesion gaps within the boundary segmentation image and identifies texture gradient features within the boundary segmentation image. If the minute depth difference in the adhesion gaps within the boundary segmentation image exceeds the depth difference threshold, the corresponding location is extracted and marked as a suspected boundary location. If the minute depth difference in the adhesion gaps within the boundary segmentation image does not exceed the depth difference threshold, the corresponding location is extracted and marked as a non-boundary location.

[0012] Furthermore, the specific process of comparing suspected boundary location features is as follows: Based on the extracted texture gradient features, feature comparison is performed on the suspected boundary positions of edible fungi. If there is a texture misalignment deviation, the corresponding suspected boundary position is marked as the confirmed boundary position, and the corresponding confirmed boundary positions are connected to construct the boundary. If there is no texture misalignment deviation, the corresponding suspected boundary position is excluded and uniformly marked as a non-boundary position.

[0013] Furthermore, the process for identifying the location of the impurity is as follows: Non-boundary locations are summarized, and the spatial morphology of edible fungi corresponding to all non-boundary locations is analyzed. If the morphology is consistent with that of the current variety of edible fungi, it is distinguished by the boundary and multiple indicators are collected and detected. If the morphology is inconsistent with that of the current variety of edible fungi, the morphological difference location is identified, marked as an impurity attachment location, uploaded to the system, and no indicators are collected at the impurity attachment location during the multi-indicator collection and detection.

[0014] Furthermore, the coordination analysis and detection process in the latent disease risk identification unit is as follows: Based on the temporal data of RGB texture images, the temporal changes of surface texture features of edible fungi are extracted. If the surface roughness of edible fungi continues to increase, or the corresponding surface brightness decreases synchronously, the corresponding edible fungi are marked as objects at risk of mold growth. Then, the parameter absorption spectrum of each position of the objects at risk of mold growth is collected through the temporal data of hyperspectral images, and the spectral reflectance of the parameter absorption spectrum is collected.

[0015] Furthermore, if the spectral reflectance of the absorption spectrum at the current location continues to decrease, the dimensional distribution of the mold-risk object in the same time series within the infrared temperature image is extracted simultaneously. If the temperature at the location where the spectral reflectance continues to decrease is higher than that of the surrounding locations and continues to increase, the corresponding location is marked as a moldy area. Simultaneously, the multimodal time series image dataset of the mold-risk object is uploaded to the system. After receiving the dataset, the system sorts and traces the current edible fungi to determine whether mold has occurred in the same batch of edible fungi.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. Accurately matches the conveyor cycle time under different speed gradients on the production line, avoiding image motion artifacts, missed bacterial images, and detection omissions under high-speed conditions, as well as the problem of massive invalid image data redundancy under low-speed conditions. This significantly reduces image recognition computing costs and system energy consumption, improves detection real-time performance, and accurately identifies two types of cycle time mismatches: lagging and leading at production line speed switching nodes. By expanding the shooting cycle and increasing the number of cycle time points through differentiated adjustment methods, the cycle time error can be accurately corrected.

[0017] 2. It accurately identifies the boundary position of tiny adhesion gaps, distinguishes between single fungal cells and adherent fungal cells, and between normal and abnormal textures, greatly improving the accuracy of individual morphology judgment, size measurement, and defect location of edible fungi. At the same time, it effectively identifies the location of impurities attached to the fungal cells, automatically avoids interference areas when detecting multiple indicators, and collects indicators only for the effective fungal cell area, eliminating the interference of impurities and foreign objects on the test results.

[0018] 3. It enables precise location of moldy areas, marking of risky fungi, and traceability of batch diseases, allowing for rapid tracing of the overall quality status of edible fungi in the same batch. This allows for quality control from the source. By using a multimodal time-series data fusion detection mode, it accurately identifies problematic fungi that have no obvious defects on the surface but have internal mold risks, achieving comprehensive detection of both obvious defects and latent diseases, and constructing a full-dimensional edible fungi quality detection system. Attached Figure Description

[0019] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0020] Figure 1 This is a flowchart illustrating the principle of the method of the present invention; Figure 2 This is a system architecture block diagram of the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0023] Example 1: Please refer to Figures 1-2 As shown, the online detection system for multiple indicators of edible fungi based on multimodal image recognition uses multimodal image recognition technology to perform online detection of multiple indicators of edible fungi. During the detection process, the system identifies and controls the interference risks that may exist in the detection, so as to improve the accuracy of online detection of multiple indicators and avoid detection errors that may lead to incorrect decisions in the sorting of edible fungi production line. The system is equipped with an image recognition beat matching unit, which is used to match the image acquisition beat with the production line beat when performing index detection on the edible fungus production line, so as to adapt to the movement of the production line, avoid image motion artifacts caused by beat mismatch, and also avoid missed beats, resulting in missed detection of edible fungi. Record the conveying speed of the edible fungus production line, determine the conveying speed range based on the current production line output task, that is, make speed predictions based on the expected output of each stage, determine the image acquisition points, set up standard reference objects on the production line, and conduct simulated idling experiments. During the simulated idle run of the production line, static acquisition of standard reference objects is performed using the equipment set at the image acquisition points. The texture of the standard reference objects and the surface of the production line is recorded. The production line is operated within the conveyor speed range set at each stage, and images are acquired through the image acquisition points to construct the production line image time sequence data. The actual number of times the standard reference object passes through the production line image time sequence data is matched with the number of times the image is recognized. If they match, it indicates that the image recognition cycle time is suitable for the current conveyor speed adjustment stage. If they do not match, it indicates that the image recognition cycle time is not suitable for the current conveyor speed adjustment stage. The cycle points of the current image acquisition equipment are adjusted, the time nodes of production line speed changes are recorded, the speed deviation span of each node is identified, speed change redundancy is set according to the current speed deviation span to form a speed fluctuation span standard, and the nodes where reference objects in the production line pass through the image acquisition nodes are obtained again according to the speed fluctuation span standard. The cycle points of the image acquisition equipment are then replanned based on this. If there are images without reference objects after replanning, the cycle points are fine-tuned to ensure that image acquisition is compatible with the production line. At the same time, it avoids the large amount of image data caused by continuous acquisition, which would lead to unnecessary image recognition costs when identifying edible fungi and matching cycle points. Images containing standard reference objects are extracted from the time-series data of production line images, and adjacent left and right images are collected to construct a recognition set. The number of adjacent left and right images extracted is set manually by recognizing the production line speed. If the images containing standard reference objects have artifacts, it indicates that the current image recognition cycle is not suitable for the current conveyor speed. If the images containing standard reference objects do not have artifacts, it indicates that the current image recognition cycle is suitable for the current conveyor speed. When the conveying speed is not compatible, the recognition set is extracted and divided into a front set and a back set based on the images containing standard reference objects. If the clarity of the surface texture of the production line in the images collected in the front set deviates from that in the static acquisition state, and the clarity deviation continues to worsen as the time sequence increases, it indicates that the current image acquisition device's beat point is later than the production line's passing node. That is, the image recognition beat is lagging behind the production line's passing beat. A lag compensation signal is generated and uploaded to the system. After receiving the signal, the system expands the shooting cycle of the current image acquisition device, keeping the number of beat points in the entire acquisition cycle unchanged and lengthening the interval beat cycle length. If artifacts gradually appear between the acquired images and the standard reference images in the later stages of the sequence, it indicates that the current image acquisition device's cycle point intervals are not compatible with the current production line speed. A continuous adaptation signal is generated and uploaded to the system. Upon receiving this signal, the system increases the shooting frequency of the current image acquisition device and, while maintaining the current acquisition cycle length, increases the number of cycle points to ensure that the image acquisition device's cycle points are compatible with the passing cycle of the reference objects in the production line. It should be noted that a cycle point represents the start and end of image acquisition. During slow passage, cycle points are executed gradually; if the passage speed is high and the number of cycle points is large, the execution of cycle points is rapid. After completing the image recognition beat matching, the system has a built-in boundary segmentation unit, which is used to distinguish the boundaries of edible fungi during the multimodal image recognition process. This avoids the edible fungi edges from sticking together in the production line, which would cause the image recognition to be unable to determine the morphology of the edible fungi, resulting in detection omissions or failure to identify edible fungi. During the operation of the production line, image recognition is performed on the edible fungi transported by the production line. The recognition logic is based on the morphology of the current variety of edible fungi. When the edible fungi have the same morphology and the spatial proportion increases, the image is extracted for boundary segmentation. The RGB texture image and ultra-high precision micro-depth image of edible fungi in the boundary segmentation image are obtained and uniformly labeled as multimodal image data. Based on the multimodal image data, the small depth difference of the adhesion gap in the boundary segmentation image is collected, and the texture gradient features in the boundary segmentation image are identified. If the small depth difference of the adhesion gaps in the boundary segmentation image exceeds the depth difference threshold, it indicates that there may be a non-single edible fungus at the current location. The corresponding location is extracted and marked as a suspected boundary location. If the small depth difference of the adhesion gaps in the boundary segmentation image does not exceed the depth difference threshold, it indicates that the current location is a single edible fungus. The corresponding location is extracted and marked as a non-boundary location. Based on the extracted texture gradient features, feature comparison is performed on the suspected boundary positions of edible fungi. If there is a texture misalignment deviation, the corresponding suspected boundary position is marked as the confirmed boundary position, and the corresponding confirmed boundary positions are connected to construct the boundary. If there is no texture misalignment deviation, the corresponding suspected boundary position is excluded and uniformly marked as a non-boundary position. The non-boundary locations are summarized, and the spatial morphology of the edible fungi corresponding to all non-boundary locations is analyzed. If the morphology is consistent with that of the current variety of edible fungi, it is distinguished by the boundary and multiple indicators are collected and detected. If the morphology is inconsistent with that of the current variety of edible fungi, the morphological difference location is identified, marked as the location of the impurity, uploaded to the system, and no indicators are collected at the location of the impurity during the collection and detection of multiple indicators. After the system completes multimodal image recognition and control, it extracts and detects multiple indicators of edible fungi based on the collected image data. The specific indicators collected through image recognition are: morphology and size: cap diameter, stipe length, roundness, thickness, deformity; color / whiteness, color uniformity, browning degree; surface defects: damage, cracks, insect holes, mold spots, impurities, foreign objects; maturity grading, variety appearance differentiation, mold identification; quantity statistics, sorting grade (first-grade / second-grade / ungraded mushrooms).

[0024] Example 2: Based on the previous example, sampling bias is avoided when collecting and testing multiple indicators of edible fungi. However, in the multi-indicator sampling and testing stage, only static sampling of the surface is performed, which cannot perform dynamic detection of internal indicators of edible fungi. For example, the surface may not show mold at the time, but there is already a risk of mold inside. The system is equipped with a hidden disease risk identification unit, which is used to conduct continuous phased detection of edible fungi to avoid the situation where mold is not detected during production line testing but mold appears during storage, thus reducing the accuracy of multi-indicator detection of edible fungi on the production line. Dynamic monitoring of edible fungi in the production line is carried out. A dynamic monitoring period is set, and RGB texture images, hyperspectral images and infrared temperature images of edible fungi are acquired during the dynamic monitoring period to form a multimodal time-series image dataset. The multimodal time-series image dataset is then analyzed and detected in conjunction with each other. Based on the temporal data of RGB texture images, the temporal changes of surface texture features of edible fungi are extracted. If the surface texture roughness of edible fungi continues to increase, or the corresponding surface brightness decreases synchronously, the corresponding edible fungi are marked as objects at risk of mold growth. Conversely, if the surface texture roughness of edible fungi does not continue to increase, and the corresponding surface brightness does not change, the edible fungi are continuously monitored until the end of the dynamic detection period. The texture roughness is represented by the continuous increase in the number of texture branches on the surface of edible fungi, which are disordered and irregular. The parametric absorption spectra of objects at risk of mold growth are collected using time-series hyperspectral image data. The parameters are represented by polysaccharides, proteins, and moisture. The fluctuation characteristics of the parametric absorption spectra are identified, namely spectral reflectance. If the spectral reflectance of the absorption spectrum at the current location continues to decrease, the dimensional distribution of the mold-prone objects in the same time series within the infrared temperature image is extracted simultaneously. If the temperature at the location where the spectral reflectance continues to decrease is higher than that of the surrounding locations and continues to increase, the corresponding location is marked as a moldy area. Simultaneously, the multimodal time series image dataset of the mold-prone objects is uploaded to the system. After receiving the dataset, the system sorts and traces the current edible fungi to determine whether mold has occurred in the same batch of edible fungi.

[0025] In summary, by employing standardized reference static calibration, simulated idling time-series data construction, and cycle time adaptation verification, the technical shortcomings of traditional fixed acquisition cycles in adapting to dynamic production line conditions are addressed. Multimodal detection data is constructed by fusing RGB texture images and ultra-high precision micro-depth images. Through initial depth difference judgment and fine texture gradient judgment, accurate segmentation of the boundaries of densely adhered edible fungi individuals is achieved, resolving morphological misjudgments and detection omissions caused by fungal adhesion. Furthermore, through the fusion analysis of RGB, hyperspectral, and infrared temperature multimodal time-series image data, dynamic and continuous detection of latent mold growth and early diseases within edible fungi is achieved.

[0026] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. An online detection system for multiple indicators of edible fungi based on multimodal image recognition, characterized in that, The system includes: The image recognition beat matching unit is used to match the image acquisition beat with the production line beat during the detection of indicators in the edible fungi production line, so as to adapt to the production line; to conduct idling experiment simulation, to perform beat adaptation analysis in the speed change stage and the speed continuous stage, and to adjust and set the matching beat point based on the analysis results; The boundary segmentation unit is used to distinguish the boundaries of edible fungi during multimodal image recognition. Specifically, it performs image recognition based on the morphology of the current variety of edible fungi, collects multimodal image data, analyzes the data to obtain suspected boundary locations and non-boundary locations, and filters and determines the boundary location based on the feature comparison of suspected boundary locations. At the same time, it identifies the location of impurities and uploads them to the system for multi-index collection and detection. The latent disease risk identification unit is used to conduct phased and continuous detection of edible fungi. It sets dynamic detection periods, collects multimodal time-series image datasets, and the multimodal time-series image datasets include RGB texture images, hyperspectral images, and infrared temperature images of edible fungi during the dynamic detection period. It also performs cooperative analysis and detection based on the multimodal time-series image datasets to determine moldy areas and uploads them to the system.

2. The online detection system for multiple indicators of edible fungi based on multimodal image recognition according to claim 1, characterized in that, The idling experiment simulation content in the image recognition beat matching unit is as follows: Record the conveying speed of the edible fungus production line, determine the range of conveying speed based on the current production line output target, determine the image acquisition points, set up standard reference objects on the production line, and conduct simulated idling experiments.

3. The online detection system for multiple indicators of edible fungi based on multimodal image recognition according to claim 2, characterized in that, The specific process of beat adaptation analysis is as follows: Static data was collected from a standard reference object during the simulated idling of the production line, and the texture of the standard reference object and the surface of the production line was recorded. The production line operates within the conveying speed range set for each stage, and images are acquired through image acquisition points to construct production line image time-series data. If the actual number of times the standard reference object passes through the production line image time-series data is inconsistent with the number of times the image is recognized, it indicates that the image recognition cycle time is not suitable for the current conveyor speed adjustment stage. The cycle point of the current image acquisition equipment is adjusted, the time nodes of production line speed change are recorded, the speed deviation span of each node is identified, speed change redundancy is set according to the current speed deviation span to form a speed fluctuation span standard, the node of reference object in the production line passing through the image acquisition node is re-determined according to the speed fluctuation span standard, and the cycle point of the image acquisition equipment is re-planned accordingly.

4. The online detection system for multiple indicators of edible fungi based on multimodal image recognition according to claim 3, characterized in that, When the conveying speed is not suitable, the identification set is extracted and the images containing the standard reference are divided into the front set and the back set. If the clarity of the surface texture of the pipeline in the acquired images in the front set deviates from that of the surface texture of the pipeline in the static acquisition state, and the clarity deviation continues to increase with the time sequence, a hysteresis compensation signal is generated and uploaded to the system. If artifacts gradually appear between the images acquired in the subsequent set and the standard reference image, a continuous adaptation signal is generated and uploaded to the system. After receiving the signal, the system increases the shooting frequency of the current image acquisition device and increases the number of beat points while keeping the current acquisition cycle length unchanged, so as to ensure that the beat points of the image acquisition device are adapted to the reference objects in the production line through the beat.

5. The online detection system for multiple indicators of edible fungi based on multimodal image recognition according to claim 1, characterized in that, The multimodal image data in the boundary segmentation unit consists of: RGB texture images and ultra-high precision micro-depth images of edible fungi within the boundary segmentation image.

6. The online detection system for multiple indicators of edible fungi based on multimodal image recognition according to claim 5, characterized in that, The methods for obtaining suspected boundary locations and non-boundary locations are as follows: Based on multimodal image data acquisition, the small depth difference of the adhesion gaps in the boundary segmentation image is collected, and the texture gradient features in the boundary segmentation image are identified at the same time. If the slight depth difference of the adhesion gaps in the boundary segmentation image exceeds the depth difference threshold, the corresponding position is extracted and marked as a suspected boundary position. If the slight depth difference of the adhesion gaps in the boundary segmentation image does not exceed the depth difference threshold, the corresponding position is extracted and marked as a non-boundary position.

7. The online detection system for multiple indicators of edible fungi based on multimodal image recognition according to claim 6, characterized in that, The specific process for comparing suspected boundary location features is as follows: Based on the extracted texture gradient features, feature comparison is performed on the suspected boundary positions of edible fungi. If there is a texture misalignment deviation, the corresponding suspected boundary position is marked as the confirmed boundary position, and the corresponding confirmed boundary positions are connected to construct the boundary. If there is no texture misalignment deviation, the corresponding suspected boundary position is excluded and uniformly marked as a non-boundary position.

8. The online detection system for multiple indicators of edible fungi based on multimodal image recognition according to claim 7, characterized in that, The process for identifying the location of impurities is as follows: Non-boundary locations are summarized, and the spatial morphology of edible fungi corresponding to all non-boundary locations is analyzed. If the morphology is consistent with that of the current variety of edible fungi, it is distinguished by the boundary and multiple indicators are collected and detected. If the morphology is inconsistent with that of the current variety of edible fungi, the morphological difference location is identified, marked as an impurity attachment location, uploaded to the system, and no indicators are collected at the impurity attachment location during the multi-indicator collection and detection.

9. The online detection system for multiple indicators of edible fungi based on multimodal image recognition according to claim 1, characterized in that, The detection process for the coordination analysis in the latent disease risk identification unit is as follows: Based on the temporal data of RGB texture images, the temporal changes of surface texture features of edible fungi are extracted. If the surface roughness of edible fungi continues to increase, or the corresponding surface brightness decreases synchronously, the corresponding edible fungi are marked as objects at risk of mold growth. Then, the parameter absorption spectrum of each position of the objects at risk of mold growth is collected through the temporal data of hyperspectral images, and the spectral reflectance of the parameter absorption spectrum is collected.

10. The online detection system for multiple indicators of edible fungi based on multimodal image recognition according to claim 9, characterized in that, If the spectral reflectance of the absorption spectrum at the current location continues to decrease, the dimensional distribution of the mold-prone objects in the same time series within the infrared temperature image is extracted simultaneously. If the temperature at the location where the spectral reflectance continues to decrease is higher than that of the surrounding locations and continues to increase, the corresponding location is marked as a moldy area. Simultaneously, the multimodal time series image dataset of the mold-prone objects is uploaded to the system. After receiving the dataset, the system sorts and traces the current edible fungi to determine whether mold has occurred in the same batch of edible fungi.

Images