Real-time detection system for fruit maturity

Abstract

The invention relates to the technical field of detection, and provides a fruit maturity real-time detection system which comprises a visual detection module, a touch detection module, a data fusion and decision module and a control and execution module. The visual detection module obtains a target fruit image and completes maturity grade identification and confidence output. The tactile detection module carries out non-destructive contact when the confidence coefficient is insufficient, and two elastic parameters reflecting peel and pulp characteristics are extracted. And the data fusion and decision module performs maturity joint judgment according to the visual result and the tactile parameter, and adjusts the judgment parameter through a preset rule under a conflict condition. And the control and execution module receives the final judgment result and completes response operation.

Description

Technical Field

[0001] The present invention relates to the field of detection technology, and more particularly to a real-time detection system for fruit maturity. Background Art

[0002] With the continuous development of smart agricultural technologies, rapid recognition and classification of peach ripeness, a fruit with highly variable appearance and texture, is crucial for automated harvesting and grading in orchards. Currently, commonly used detection methods are primarily based on computer vision technology, such as in Reference 1 (Wang Y., Jin X., Zheng J., Zhang X., Wang X., He X., Polovka M. (2023). An energy-efficient classification system for peach ripeness using YOLOv4 and flexible piezoelectric sensor, Computers and Electronics in Agriculture, Vol. 207, 107888.). This method extracts features such as fruit surface color and shape and uses deep learning models (such as YOLOv4 and YOLOv5) to classify and identify ripeness levels. This method offers the advantages of non-contact, high speed, and ease of deployment, and demonstrates good recognition performance under conditions of uniform lighting and adequate fruit exposure.

[0003] However, peaches, a fruit with slowly changing skin color, noticeable surface fuzz, and uneven ripening, are often significantly affected by occlusion, lighting angle, and appearance heterogeneity. This can lead to low confidence levels in some images. Against this backdrop, researchers have explored new solutions from the perspective of tactile perception. In Reference 2 (Nnodim CT, et al., Design and Simulation of a Tactile Sensor for Fruit Ripeness Detection, WCECS 2019), Nnodim et al. proposed a mechanical model based on a dual-spring structure. By simplifying mangoes of varying ripeness into targets with variable elastic moduli, and combining simulations with validation, they established a quantitative relationship between contact force ratio and fruit hardness, demonstrating the potential of tactile sensors for fruit ripeness estimation. Their research framework demonstrates the importance of the mechanical response of fruit in ripeness identification.

[0004] Although the combination of vision and touch has attracted attention, the existing technology is still at the primary stage of using overall stiffness as a tactile indicator. It has failed to establish a hierarchical mechanical modeling structure, making it difficult to finely distinguish the maturity of the peel and flesh. It also lacks a fusion and adjustment mechanism for multimodal detection conflicts, making it difficult to meet the actual usage needs in complex orchard scenarios. Summary of the Invention

[0005] To overcome the aforementioned shortcomings of existing technologies, the present invention provides a real-time fruit ripeness detection system. This system utilizes the YOLOv4 model for preliminary image recognition. When recognition confidence is low, it uses a flexible capacitive sensor to acquire contact force-displacement responses. This system constructs a dual-layer elastic model of the peel and flesh, extracting tactile feature parameters for ripeness determination. The system incorporates a confidence-driven multimodal fusion mechanism and a sliding window dynamic correction module, effectively addressing the inaccurate judgments of existing methods due to fuzzy image features, single-metric tactile perception, and a lack of adaptive strategies.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A real-time fruit maturity detection system is characterized by comprising a visual detection module, a tactile detection module, a data fusion and decision module, and a control and execution module, wherein the visual detection module and the tactile detection module are both connected to the data fusion and decision module, which is then connected to the control and execution module; the visual detection module is used to acquire and analyze fruit images and output a preliminary maturity grade and a confidence value; the tactile detection module is used to perform non-destructive contact detection on peaches when the confidence value is lower than a preset threshold, record the force-displacement response signal during the contact process, and extract two elastic parameters reflecting the structural characteristics of the peel and flesh respectively through a segment fitting method; the data fusion and decision module is used to jointly determine the fruit maturity grade based on the elastic parameters and visual judgment results, and when the difference between the visual judgment grade and the tactile judgment grade exceeds a judgment conflict threshold, dynamically adjust the classification parameters according to preset fusion rules and an error statistical model; and the control and execution module is used to control a mechanical execution device to complete a picking operation according to the judgment result.

[0008] As a further solution of the present invention, the tactile detection module includes the following implementation steps:

[0009] Step A1, using a flexible tactile sensor mounted on the end of the robotic arm to apply non-destructive contact to the surface of the peach;

[0010] Step A2: During the contact process, the relationship between the electrical signal of the pressure sensor and the pressing displacement is recorded in real time to form a force-displacement curve;

[0011] Step A3, performing double-layer tactile modeling on the force-displacement curve, dividing the curve into a first pressure interval reflecting the surface characteristics of the peel and a second pressure interval reflecting the overall structure of the flesh;

[0012] Step A4, performing linear fitting on the first pressing interval and the second pressing interval respectively to obtain a first elastic coefficient and a second elastic coefficient;

[0013] Step A5: searching and matching the first elastic coefficient and the second elastic coefficient in a pre-calibrated reference model, and outputting a result of determining the maturity level of tactile perception.

[0014] As a further solution of the present invention, the step of constructing the reference model includes:

[0015] First, a certain number of fresh peach samples were selected, covering the full maturity range from unripe to fully ripe. The samples were required to be undamaged, of moderate size, and from a uniform source. For each sample, three or more fruit tree experts conducted a comprehensive visual and tactile assessment of its maturity according to a standardized evaluation standard. Each sample was then assigned a uniform grade: Grade I, Grade II, Grade III, Grade IV, and Grade V. All samples were labeled, forming a fully annotated dataset.

[0016] After expert calibration is complete, tactile testing is performed on all samples sequentially. Using the flexible capacitive tactile sensor of the present invention, a robotic arm controls the clamping end to perform non-destructive pressure on the fruit surface under set contact speed and maximum penetration depth conditions. The system records the pressure sensing signal and corresponding displacement data during the pressing process, constructing a complete force-displacement response curve. To reduce the impact of individual fluctuations, each sample is sampled three times, and the average value is taken as the final valid curve. This process ensures the acquisition of comparable standard tactile feature data.

[0017] Based on the acquired force-displacement curve, the system strategically divides the force into two penetration zones, corresponding to the peel and flesh, respectively. A linear fit is performed within each zone, yielding two fitting slopes, defined as the first and second elastic coefficients. A set of three-dimensional data is generated for all samples: {first elastic coefficient, second elastic coefficient, maturity level}. During data processing, abnormal signals (such as samples with excessively large fitting residuals or test failures) are eliminated to ensure the representativeness and stability of the underlying data for modeling.

[0018] All cleaned sample data is input into the modeling module to construct a mapping relationship between the feature space of the first and second elasticity coefficients and the maturity level. A variety of modeling strategies can be used. If a rule-matching strategy is used, a statistical analysis of the first and second elasticity coefficients of all samples is performed by level, and the mean interval of each level is extracted as the boundary to generate a table-based classification standard. If a discriminant model strategy is used, training is performed based on a support vector machine (SVM), K-nearest neighbor (KNN), or decision tree model to output a discriminant boundary function.

[0019] After modeling is complete, the resulting rule classification table or trained model file is embedded in the present invention's tactile discrimination module, serving as a pre-calibrated reference model for use during system runtime. During real-time testing, when new first and second elastic coefficient values are acquired, the system uses the model to find the corresponding grade interval or perform a discrimination calculation, outputting the maturity level of tactile perception.

[0020] As a further solution of the present invention, the flexible tactile sensor in the tactile detection module adopts a capacitive structure, including:

[0021] The flexible base layer is made of polydimethylsiloxane material and has regularly arranged micro-nano structures on its surface;

[0022] The conductive electrode layer comprises two layers of flexible electrodes, one on top and one on the bottom, located on both sides of the flexible base layer;

[0023] The protective covering layer is made of food-grade silicone material and covers the outer surface of the sensor to protect the fruit from damage;

[0024] The signal conditioning circuit is connected to the conductive electrode layer and is used for converting the capacitance change into a voltage signal output.

[0025] As a further solution of the present invention, the data fusion and decision module includes the following implementation steps:

[0026] Step B1, receiving the preliminary maturity level and confidence value output by the visual detection module, and the tactile characteristic parameters output by the tactile detection module, wherein the tactile characteristic parameters include a first elastic coefficient and a second elastic coefficient;

[0027] Step B2, setting a visual judgment confidence threshold and a tactile judgment validity threshold;

[0028] Step B3: When the visual judgment confidence value is higher than the threshold, the visual inspection result is directly used as the final judgment;

[0029] Step B4: when the visual judgment confidence value is lower than the threshold and the tactile feature parameter meets the validity condition, the tactile detection result is used as the final judgment;

[0030] Step B5: When the difference between the visual judgment level and the tactile judgment level exceeds the judgment conflict threshold set by the system, the weighted fusion algorithm is triggered and the final maturity level is calculated using the weighted fusion algorithm. The mathematical expression of the weighted fusion algorithm is:

[0031] G = α·Gv+(1-α)·Gt;

[0032] Where G is the final maturity level, Gv is the visual judgment level, Gt is the tactile judgment level, and α is the dynamic weight coefficient, which is related to the confidence of visual judgment and the stability of tactile feature parameters.

[0033] The dynamic weight coefficient is calculated according to the following formula:

[0034]

[0035] Wherein, σ is the standard deviation of the first elastic coefficient and the second elastic coefficient within the sliding time window.

[0036] As a further aspect of the present invention, the validity condition is that the first and second elastic coefficients acquired by the sensor have fitting residuals below a preset threshold, parameter values are within the valid range of the pre-calibrated model, and the results of repeated sampling are highly consistent. Specifically, this validity condition is obtained by the system through annotated sample statistics during the initial training phase, and the stability and matching degree of each test result are evaluated in real time during runtime, ensuring that the tactile judgment logic only independently outputs the maturity level when the data is reliable, thereby ensuring the accuracy and robustness of the fusion decision.

[0037] As a further aspect of the present invention, the conflict threshold is determined by collecting the visual and tactile ratings for each sample, along with the actual maturity level manually annotated by experts, and calculating the difference between the visual and tactile results. By statistically analyzing the distribution of these differences across all samples and combining the system's false positive rate with decision stability, an acceptable upper limit for discrepancy is identified to determine whether there is a conflict between the two detection methods. When the difference exceeds this statistical threshold and the system's false positive rate increases significantly, a non-negligible discrepancy between the visual and tactile judgments is considered.

[0038] As a further solution of the present invention, the data fusion and decision module also includes a dynamic error self-correction mechanism, which is implemented by the following steps:

[0039] Step C1, establish a sliding time window and record the error between the historical detection results and the actual calibration results;

[0040] Step C2, calculating the statistical characteristics of the error within the sliding window, including the mean and standard deviation;

[0041] Step C3, setting a correction trigger threshold, when the absolute value of the mean is greater than the correction trigger threshold or the standard deviation is greater than the correction trigger threshold, starting the self-correction program;

[0042] Step C4 adjusts the visual inspection module's classification threshold and the mapping relationship between the first and second elastic coefficients based on the error distribution characteristics. The classification threshold is adjusted as follows: when the mean error value is consistently high, the current level boundary is appropriately lowered to increase sensitivity; when the mean error value is low or misjudgments frequently occur at the boundary between adjacent levels, the boundary value is appropriately raised to enhance judgment robustness. To adjust the elastic coefficients, the system refits the feature ranges corresponding to different maturity levels within the statistical window and, accordingly, updates the interval boundaries of the first and second elastic coefficients in the mapping model, achieving dynamic optimization of tactile judgment accuracy.

[0043] Step C5: After the adjustment is completed, the new parameters are applied to the subsequent detection process and the error statistics are updated.

[0044] As a further solution of the present invention, the classification threshold is a confidence threshold or a dividing value used in the visual recognition module to map the image recognition results into five maturity levels.

[0045] As a further solution of the present invention, the visual detection module includes:

[0046] Image acquisition unit, which acquires fruit images through an installed high-resolution industrial camera;

[0047] An image preprocessing unit performs normalization, color correction, and region cropping on the acquired image;

[0048] Feature extraction and recognition unit, which uses a deep learning-based object detection algorithm to identify the location of fruits and extract features;

[0049] The maturity preliminary classification unit maps the recognition results to predefined maturity levels and calculates the corresponding confidence values.

[0050] As a further solution of the present invention, the feature extraction and recognition unit adopts the YOLOv4 deep learning model, which is trained with labeled peach maturity samples and uses transfer learning technology to improve the robustness of the model under different lighting conditions and occlusion conditions, and divides the maturity into five levels: Grade I maturity, Grade II maturity, Grade III maturity, Grade IV maturity and Grade V maturity.

[0051] As a further solution of the present invention, the control and execution module includes a picking working mode, which is used in orchard scenes. It determines whether to perform picking operations by judging the maturity of the fruit, and controls the robotic arm and end effector to complete the picking action.

[0052] As a further solution of the present invention, the end effector in the picking working mode is a robotic claw with an integrated tactile sensor, which can complete tactile detection while clamping the peaches. The tactile sensor is installed on the inside of the claw, and the contact surface is covered with non-slip flexible material. The closing force of the claw is dynamically adjusted by the controller according to real-time tactile signal feedback to avoid squeezing damage to the fruit.

[0053] As a further solution of the present invention, the system's sequential workflow is as follows: first, the visual module quickly identifies the peaches in the target area and makes a preliminary maturity judgment, and outputs a confidence value; if the confidence value is higher than a preset threshold, the judgment result is directly output; if the confidence value is lower than the preset threshold, the tactile detection process is triggered; after the tactile detection is completed, the data fusion and decision module integrates the two types of perception information to make a final judgment; finally, the control and execution module controls the mechanical device to perform corresponding operations based on the judgment result.

[0054] Compared with the prior art, the beneficial effects of the fruit maturity real-time detection system of the present invention are:

[0055] This invention introduces a multi-level modeling mechanism based on force-displacement response curves into the tactile detection module, dividing the contact process into two stages: the epidermis and the flesh. Representative elastic coefficients are extracted for each stage, effectively reflecting the mechanical response characteristics of the fruit at different structural levels. Compared to traditional methods based on single-point indentation values or overall stiffness estimation, the dual-parameter modeling method of the present invention can accurately distinguish between inconsistencies between external and internal ripening. It is particularly suitable for fruit varieties where the peel and flesh mature at different rates, significantly improving the meticulousness of tactile detection and the accuracy of maturity determination.

[0056] The present invention triggers the fusion process based on the degree of difference between visual and tactile judgments, and uses a sliding window method to statistically analyze historical error trends, adaptively adjusting the mapping relationship between the classification boundaries of the visual recognition model and tactile parameters. This mechanism establishes a closed-loop optimization path, enabling the system to offset accuracy drift caused by factors such as environmental changes and sample offsets during long-term operation. Compared with existing fusion strategies that mostly use static weighting or fixed rules, the present invention possesses self-learning capabilities and robustness, and the overall maturity determination of the system is more accurate and more stable. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 Schematic diagram of the experimental device structure in Reference 2.

[0058] Figure 2 Schematic diagram of the sample after compression experiment in Reference 2.

[0059] Figure 3This is the Yolo4 identification map in Reference 1.

[0060] Figure 4 This is a system framework diagram of a real-time fruit maturity detection system of the present invention.

[0061] Figure 5 This is a process decision diagram of a real-time fruit maturity detection system of the present invention.

[0062] Figure 6 This is a force-displacement curve diagram of a real-time fruit maturity detection system of the present invention.

[0063] Figure 7 This is a real-life picture of peaches used in the real-time fruit maturity detection system of the present invention. DETAILED DESCRIPTION

[0064] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0065] In step 1, when a peach enters the system's detection range, the visual inspection module initiates its operation. A high-resolution industrial camera mounted at the end of the robotic arm or above the sorting line captures image data of the peach. This image is then transmitted to the image preprocessing unit for normalization, color correction, and region cropping to improve image quality and facilitate subsequent analysis.

[0066] In step two, the preprocessed image is fed into the Feature Extraction and Recognition Unit. This unit uses the YOLOv4 deep learning algorithm to identify the peaches' locations in the image and extract their appearance features, such as color and texture, that correlate with maturity. The Preliminary Maturity Classification Unit then maps these features to predefined maturity levels (I to V) and calculates the corresponding confidence level, representing the visual system's confidence in the resulting judgment.

[0067] In step three, the visual inspection results are transmitted to the confidence assessment unit of the data fusion and decision-making module. The system then checks whether the confidence level of the visual judgment exceeds a preset threshold. If the confidence level is above the threshold, the system deems the visual judgment reliable enough and directly uses it as the final maturity determination, skipping the tactile inspection phase and proceeding to the execution phase. If the confidence level is below the threshold, the tactile inspection process is triggered.

[0068] In step 4, when tactile detection is triggered, the control and execution module instructs the robotic arm to move, bringing the flexible tactile sensor on the end effector into contact with the peach's surface. During contact, the force control unit ensures that appropriate pressure is applied without damaging the peach. Simultaneously, the tactile detection module begins collecting force-displacement curve data. The sensor presses into the peach's surface at a constant speed to a predetermined depth, recording changes in the pressure signal throughout the process.

[0069] In step five, the force-displacement curve data is input into a dual-layer tactile modeling and analysis unit, which automatically divides the curve into a first pressure range representing the peel's characteristics and a second pressure range representing the flesh's characteristics. The system performs a linear fit on these two ranges to calculate the first and second elastic coefficients. These two parameters effectively reflect the peach's firmness and internal structure, thereby indicating its ripeness.

[0070] The following is a Python code example that uses two-layer tactile modeling to collect, process, and determine the ripeness of a peach's force-displacement response signal. Note that this example is only a starting point and may need to be adjusted in practice based on actual conditions and device interfaces.

[0071]

[0072]

[0073]

[0074]

[0075]

[0076]

[0077]

[0078]

[0079]

[0080] This code is only an example and needs to be modified and adjusted appropriately according to specific circumstances in actual applications.

[0081] In step six, the calculated elastic coefficients are transmitted to the multimodal data fusion unit in the data fusion and decision module. This unit first verifies whether the tactile parameters meet validity criteria, such as whether the sum of squared residuals is less than a set standard. The system then compares the visual and tactile judgment results to determine whether the difference between the two exceeds a conflict threshold.

[0082] Step seven: When the difference between the visual and tactile judgment results is not much, the system will give priority to the result with higher confidence; when the difference is significant, the system will start the weighted fusion algorithm, dynamically assign weights according to the reliability of the two perception methods, and calculate the final maturity level judgment result.

[0083] In step eight, the system's dynamic error self-correction mechanism continuously records the error between historical detection results and actual calibration results through a sliding time window. When the error statistics exceed the correction trigger threshold, a parameter adjustment program automatically initiates, fine-tuning the vision module's classification threshold and the mapping relationship between tactile parameters, enabling the system to adapt to environmental changes and device drift.

[0084] Step nine: The final maturity determination result is transmitted to the control and execution module and converted into specific execution instructions according to the current working mode.

[0085] Example 1

[0086] In this example, Zhongyou No. 5 peaches were selected as the target fruit. A comprehensive experimental verification of the present invention's structure, functionality, and decision logic was conducted by building a maturity detection platform integrating visual and tactile perception. The platform configuration includes a visual recognition model based on YOLOv4, a flexible capacitive tactile sensor based on a nanostructured PDMS film, a collaborative robotic arm with integrated sensors, a Jetson Xavier NX edge computing platform, and a dedicated data fusion and control execution module.

[0087] During the sample preparation phase, 150 fresh peaches of varying maturity were collected from the orchard. Three or more experts independently scored them, categorizing them into five maturity levels based on visual, tactile, and color criteria: Grade I (unripe), Grade II (slightly ripe), Grade III (medium ripe), Grade IV (ripe), and Grade V (overripe). All samples were labeled with numbers and basic information such as the plot and harvest time was recorded to ensure complete data traceability.

[0088] The experimental system performs inspections according to the following process: First, a high-resolution industrial camera captures a single sample image. After normalization, illumination enhancement, and region cropping, the image is fed into the YOLOv4 network for detection and recognition. The system outputs the peach's target bounding box location, preliminary ripeness level, and classification confidence. If the confidence value is above 0.85, it is directly used as the final output; if it is below this threshold, the system automatically enters the tactile inspection process.

[0089] At this point, the robotic arm drives the end effector, which houses the embedded tactile sensor, to press vertically downward toward the sample surface at a speed of 0.5 mm / s, with a maximum penetration depth of no more than 3 mm and a contact force limited to 0.2–1.2 N. The sensor collects the capacitance signal in real time and converts it into a mechanical response curve. The force-displacement curve is automatically divided into two intervals: 0.5–1.5 mm for the peel and 1.5–2.8 mm for the flesh. Two linear slopes are fitted for each interval, defined as the first and second elastic coefficients.

[0090] The dual parameters are compared with the maturity mapping model established in the system. The model is constructed based on expert calibration samples and an error band is added on the mapping boundary to improve robustness. If the tactile parameters are within the preset valid range and the error is less than the set standard (the sum of squared residuals is less than 0.03N 2 ), it is judged as a valid input and can participate in the fusion judgment.

[0091] The fusion judgment rules are as follows: If the visual and tactile judgment levels differ by no more than one level, the system prioritizes the one with higher confidence. If the difference is significant (a level difference greater than two), the fusion algorithm is triggered. The fusion result is calculated as a weighted average, where the dynamic weight coefficient is determined by the stability of the tactile data, and the system automatically calculates the standard deviation.

[0092] At the same time, the system uses a sliding window mechanism (window length 30 test samples) to record the difference between the predicted level and the actual labeled level in each round. If the average error value exceeds 0.8 or the standard deviation exceeds 0.5 levels, the parameter callback mechanism is automatically triggered to adjust the maturity level probability classification threshold output by the YOLOv4 network or adjust the level division range of the tactile parameters to eliminate drift bias. The recognition results are shown in Table 1:

[0093] Table 1

[0094]

[0095] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0096] Finally: 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, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A real-time detection system for fruit maturity, characterized in that: The system comprises a visual detection module, a tactile detection module, a data fusion and decision module, and a control and execution module, wherein the visual detection module and the tactile detection module are both connected to the data fusion and decision module, and the data fusion and decision module is connected to the control and execution module; the visual detection module is used to acquire and analyze fruit images and output a preliminary maturity grade and a confidence value; the tactile detection module is used to perform non-destructive contact detection on the peaches when the confidence value is lower than a preset threshold, record the force-displacement response signal during the contact process, and respectively extract two elastic parameters reflecting the structural characteristics of the peel and the flesh through a segment fitting method; the data fusion and decision module is used to jointly determine the maturity grade of the fruit based on the elastic parameters and the visual judgment result, and when the difference between the visual judgment grade and the tactile judgment grade exceeds the judgment conflict threshold, dynamically adjust the classification parameters according to the preset fusion rules and the error statistical model; The control and execution module is used to control the mechanical execution device to complete the picking operation according to the judgment result.

2. A fruit maturity real-time detection system according to claim 1, characterized in that: The tactile detection module includes the following implementation steps: Step A1, using a flexible tactile sensor mounted on the end of the robotic arm to apply non-destructive contact to the surface of the peach; Step A2: During the contact process, the relationship between the electrical signal of the pressure sensor and the pressing displacement is recorded in real time to form a force-displacement curve; Step A3, performing double-layer tactile modeling on the force-displacement curve, dividing the curve into a first pressure interval reflecting the surface characteristics of the peach peel and a second pressure interval reflecting the overall structure of the peach flesh; Step A4, performing linear fitting on the first pressing interval and the second pressing interval respectively to obtain a first elastic coefficient and a second elastic coefficient; Step A5: searching and matching the first elastic coefficient and the second elastic coefficient in a pre-calibrated reference model, and outputting a result of determining the maturity level of tactile perception.

3. A fruit maturity real-time detection system according to claim 1, characterized in that: The data fusion and decision-making module includes the following implementation steps: Step B1, receiving the preliminary maturity level and confidence value output by the visual detection module, and the tactile characteristic parameters output by the tactile detection module, wherein the tactile characteristic parameters include a first elastic coefficient and a second elastic coefficient; Step B2, setting a visual judgment confidence threshold and a tactile judgment validity threshold; Step B3: When the visual judgment confidence value is higher than the threshold, the visual inspection result is directly used as the final judgment; Step B4: when the visual judgment confidence value is lower than the threshold and the tactile feature parameter meets the validity condition, the tactile detection result is used as the final judgment; Step B5: When the difference between the visual judgment level and the tactile judgment level exceeds the judgment conflict threshold set by the system, the weighted fusion algorithm is triggered and the final maturity level is calculated using the weighted fusion algorithm. The mathematical expression of the weighted fusion algorithm is: G = α·Gv+(1-α)·Gt; Where G is the final maturity level, Gv is the visual judgment level, Gt is the tactile judgment level, and α is the dynamic weight coefficient.

4. A fruit maturity real-time detection system according to claim 3, characterized in that: The data fusion and decision module also includes a dynamic error self-correction mechanism, which is implemented by the following steps: Step C1, establish a sliding time window and record the error between the historical detection results and the actual calibration results; Step C2, calculating the statistical characteristics of the error within the sliding window, including the mean and standard deviation; Step C3, setting a correction trigger threshold, when the absolute value of the mean is greater than the correction trigger threshold or the standard deviation is greater than the correction trigger threshold, starting the self-correction program; Step C4, adjusting the classification threshold of the visual inspection module and the mapping relationship between the first elastic coefficient and the second elastic coefficient according to the error distribution characteristics; Step C5: After the adjustment is completed, the new parameters are applied to the subsequent detection process and the error statistics are updated.

5. A fruit maturity real-time detection system according to claim 1, characterized in that: The visual inspection module includes: Image acquisition unit, which acquires fruit images through an installed high-resolution industrial camera; An image preprocessing unit performs normalization, color correction, and region cropping on the acquired image; Feature extraction and recognition unit, which uses a deep learning-based object detection algorithm to identify the location of fruits and extract features; The maturity preliminary classification unit maps the recognition results to predefined maturity levels and calculates the corresponding confidence values.

6. A fruit maturity real-time detection system according to claim 5, characterized in that: The feature extraction and recognition unit adopts the YOLOv4 deep learning model, which is trained with labeled peach maturity samples. The transfer learning technology is used to improve the robustness of the model under different lighting conditions and occlusion conditions, and the maturity is divided into five levels: level I maturity, level II maturity, level III maturity, level IV maturity and level V maturity.

7. A fruit maturity real-time detection system according to claim 1, characterized in that: The flexible tactile sensor in the tactile detection module adopts a capacitive structure and includes: The flexible base layer is made of polydimethylsiloxane material and has regularly arranged micro-nano structures on its surface; The conductive electrode layer comprises two layers of flexible electrodes, one on top and one on the bottom, located on both sides of the flexible base layer; The protective covering layer is made of food-grade silicone material and covers the outer surface of the sensor to protect the fruit from damage; The signal conditioning circuit is connected to the conductive electrode layer and is used for converting the capacitance change into a voltage signal output.

8. A fruit maturity real-time detection system according to claim 1, characterized in that: The control and execution module includes a picking working mode, which is used in orchard scenarios. It determines whether to perform picking operations by judging the maturity of peaches, and controls the robotic arm and end effector to complete the picking action.

Images