Pear variety identifying and sorting device and method based on deep transfer learning algorithm

Abstract

The invention discloses a variety pear identifying and sorting device and method based on a deep transfer learning algorithm. The device comprises a conveying and rotating mechanism, a fusion sensing unit, a processing and control unit and a sorting executing mechanism. According to the method, a dynamic anti-interference optical coherence tomography sensor is used for collecting an original one-dimensional interference spectrum signal sequence in the pear movement process, and the signal originally fuses the surface appearance and the shallow internal characteristics; then, directly inputting the processed signal into a variety identification neural network model which is pre-trained and finely adjusted by adopting a deep transfer learning algorithm for end-to-end analysis, and outputting a variety identification result; and finally, the sorting action is controlled according to an identification result. According to the method, the multi-source data synchronization problem is radically solved through a single-sensor scheme, the small sample training bottleneck is overcome by utilizing transfer learning, and high-precision, high-robustness and low-cost automatic identification and sorting of pear varieties in a real production line high-speed dynamic environment are realized.

Description

Technical Field

[0001] This invention belongs to the field of automated post-harvest processing technology for agricultural products, specifically relating to an online identification and sorting device and method for fruits such as pears. Background Technology

[0002] In the commercial processing stage of the pear industry, achieving automated and high-precision sorting based on variety is of great significance for enhancing product added value, ensuring brand purity, and meeting the refined demands of the market. To achieve this goal, existing technologies generally employ two or more of the following approaches: First, machine vision-based methods primarily utilize high-resolution industrial cameras to capture visible features of pears, such as surface color, shape, and texture, and then classify them using image processing algorithms. Second, multi-sensor information fusion methods, which introduce near-infrared spectroscopy, hyperspectral imaging, or laser sensors into the vision system to attempt to acquire internal quality information (such as sugar content and acidity) or deeper physical information about the pears, aiming to improve variety differentiation by fusing surface and internal features.

[0003] However, these existing technologies have revealed significant shortcomings and bottlenecks in practical industrial applications. First, for pear varieties with highly similar appearances (such as fragrant pears from different origins), relying solely on surface visual features is insufficient for reliable differentiation, resulting in limited accuracy. Second, and more critically, while multi-sensor solutions (such as cameras and spectrometers) theoretically provide richer information, they face severe challenges in the spatiotemporal synchronization and registration of multi-source heterogeneous data under high-speed, continuous, and dynamic production line conditions (where pears simultaneously undergo transport revolution and rotation). Minor trigger delays between sensors, mechanical vibrations, and instantaneous changes in fruit posture can easily lead to inaccurate temporal and spatial correspondences between different feature data from the same pear, resulting in a weak foundation for subsequent data fusion algorithms and a significant decrease in recognition stability and reliability from the laboratory to the production line. Furthermore, the aforementioned deep learning-based methods typically rely on large-scale, high-quality brand-labeled sample libraries for model training, but acquiring such data in agricultural settings is costly and time-consuming, posing another major obstacle to technology implementation. Simultaneously, complex multi-sensor systems also bring problems such as high hardware costs and cumbersome system calibration and maintenance.

[0004] Therefore, existing technologies struggle to achieve high-precision and robust online identification of pear varieties at a reasonable cost while meeting the pace of industrial production. Exploring a novel technological approach that fundamentally circumvents the multi-sensor synchronization problem and enables efficient learning using limited labeled data has become a critical technological bottleneck that urgently needs to be overcome, making improvement imperative. Summary of the Invention

[0005] To address the shortcomings of existing pear variety sorting technologies, such as difficulties in asynchronous fusion of multi-sensor information and reliance on large amounts of labeled data, this invention provides a device and method for identifying and sorting pear varieties based on a deep transfer learning algorithm. This addresses the technical challenge of achieving high-precision, high-robustness, and low-cost automatic variety identification in a high-speed, dynamic industrial environment.

[0006] The solution to the technical problem of this invention is as follows: a method for identifying and sorting pear varieties based on a deep transfer learning algorithm, comprising the following steps: S1. When an individual pear moves and rotates through the detection area, its surface is scanned using an optical coherence tomography (OCT) sensor to collect the original one-dimensional interference spectral signal sequence; S2. The processed original one-dimensional interference spectral signal sequence is input into a pre-trained variety identification neural network model; wherein, the variety identification neural network model is trained using a deep transfer learning algorithm, and the training includes: pre-training the basic neural network on a general dataset to obtain general feature extraction capabilities, and then fine-tuning the pre-trained model on a pear variety labeled spectral dataset; S3. The variety identification neural network model processes the input signal sequence and outputs the pear variety identification result; S4. Based on the variety identification result, the sorting execution mechanism is controlled to sort the pears.

[0007] Preferably, in step S1, while acquiring the original one-dimensional interference spectrum signal sequence, a reference calibration signal generated by a fixed reference surface inside the sensor is also acquired simultaneously, and the acquired signal is corrected for motion artifacts in real time based on the reference calibration signal.

[0008] Preferably, in step S2, the variety identification neural network model is a one-dimensional convolutional neural network; the one-dimensional convolutional neural network integrates an attention mechanism module, which is used to adaptively weight the spectral dimension or feature channel dimension of the signal sequence.

[0009] Preferably, in the fine-tuning step of the deep transfer learning algorithm, the loss function used includes a triplet loss function, which is used to optimize the feature space so that the feature distance of samples of the same variety decreases and the feature distance of samples of different varieties increases.

[0010] Another pear variety identification and sorting device based on a deep transfer learning algorithm is adopted, comprising: a conveying and rotating mechanism for moving and rotating individual pears through a detection station; an optical coherence tomography (OCT) sensor, set at the detection station, for acquiring the original one-dimensional interference spectrum signal sequence of the pears; a processing and control unit, communicatively connected to the sensor, which internally deploys a variety identification neural network model trained using a deep transfer learning algorithm; the processing and control unit is configured to: process the signal sequence, use the model to perform variety identification, and generate sorting instructions; and a sorting execution mechanism, set at the sorting station and communicatively connected to the processing and control unit, for sorting the pears according to the sorting instructions.

[0011] Preferably, the optical coherence tomography sensor is a swept-frequency optical coherence tomography module, and its reference arm integrates a fixed reference mirror to provide a reference signal for real-time motion artifact correction.

[0012] Preferably, the processing and control unit is further configured with a model incremental learning module, which is used to perform incremental fine-tuning based on the existing model using a small amount of sample data of newly added varieties, so as to expand the range of identifiable varieties.

[0013] Preferably, the training process of the variety identification neural network model includes: firstly, performing self-supervised or supervised pre-training on a dataset containing multiple types of optical or spectral signals; then, fine-tuning the model using pear variety-labeled spectral data for domain adaptation; and during the fine-tuning stage, employing a hard sample mining strategy to select training samples.

[0014] The beneficial effects of this invention are as follows: 1. By employing native fusion sensing of a single sensor (OCT) and direct processing of the original signal using an end-to-end deep neural network, significant optimizations in recognition accuracy, processing efficiency, and signal stability are directly achieved. In a high-speed dynamic production line environment, the recognition accuracy of pear varieties is increased to over 98%; by avoiding the complex calculations of traditional image reconstruction and multi-source fusion, the single recognition inference time is shortened to less than 15 milliseconds; and by utilizing a fixed reference mirror for real-time correction, the signal stability (signal-to-noise ratio) under vibration conditions is improved by more than 20 dB.

[0015] 2. In terms of practical application, the system hardware configuration has been simplified, replacing the traditional multi-sensor arrays such as cameras and spectrometers with a single high-performance sensor, significantly reducing the implementation threshold and the complexity of long-term maintenance. Simultaneously, it has promoted the optimization of overall system cost and energy consumption, with hardware cost estimated to be reduced by approximately 40%; due to the simplified computation process, overall system power consumption has been reduced by over 60%, making it possible to deploy high-precision models on low-cost edge computing devices.

[0016] 3. The system maintains stability under broader and more demanding conditions. For example, thanks to multi-scale feature fusion and metric learning strategies, it maintains a high recognition rate even when there is slight contamination on the pear surface or when the features of different varieties are extremely similar, demonstrating its strong robustness and practicality. It provides a reliable path for subsequent technology iterations and application expansion. The incremental learning module enables the system to quickly adapt to new varieties at low cost, showcasing good scalability. The complete perception-processing integration provides a proven underlying implementation scheme for transferring this paradigm to the detection of other agricultural products such as apples and citrus. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall structure of the pear variety identification and sorting device provided in the embodiment of the present invention.

[0018] Figure 2 yes Figure 1 Enlarged structural diagram of the middle-row fruit roller.

[0019] Figure 3 This is a schematic diagram of the structure of the dynamic anti-interference optical coherence tomography sensor provided in an embodiment of the present invention.

[0020] Figure 4 This is a block diagram illustrating the principle of real-time motion artifact correction in software provided in this embodiment of the invention.

[0021] Figure 5 This is a schematic diagram of the end-to-end variety identification neural network model architecture provided in an embodiment of the present invention.

[0022] Figure 6 This is a flowchart of the pear variety identification and sorting process.

[0023] The diagram is labeled as follows: 10-Conveying and rotating mechanism; 20-Fusion sensing unit; 30-Processing and control unit; 40-Sorting execution mechanism; 101-Discharging roller; 102-Drive motor; 103-Upstream conveyor belt; 104-Downstream conveyor belt; 110-Discharging tray; 111-Toothed head; 113-Side frame; 112-Guide body; 114-Follower pulley; 115-Follower roller; 116-Fixed toothed plate; 21-Dynamic anti-interference optical coherence tomography sensor; 211-Sweep laser source; 212-Fiber optic coupler; 213-Sample arm; 214-Reference arm; 2131-Collimator; 2132-Two-dimensional MEMS scanning galvanometer; 2141-Fixed reference mirror; 215-Balanced photodetector; 41-Cylinder; 42-Baffle; 43-Guide seat. Detailed Implementation

[0024] The technical solution of the present invention will now be described in detail and clearly with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0025] In the post-harvest commercial processing of pears, achieving automated, high-precision online identification and sorting based on pear varieties is crucial for enhancing product added value and brand management. Existing technologies lack an effective solution for accurately identifying different pear varieties on a high-speed industrial sorting line in a low-cost and highly reliable manner. This invention aims to solve the technical challenge of achieving high-precision and high-efficiency online identification of pear varieties under dynamic operating conditions using limited labeled samples. The following, in conjunction with the accompanying drawings, details the specific implementation of the technical solution of this invention using a preferred embodiment. This embodiment discloses a device and method for identifying and sorting pear varieties based on a deep transfer learning algorithm.

[0026] Figure 1 This is a schematic diagram illustrating the overall structure and workflow of an embodiment of the pear variety identification and sorting device described in this invention. As shown, the device mainly includes: a frame (unlabeled), a conveying and rotating mechanism 10, a fusion sensing unit 20 located at the detection station, a processing and control unit 30, and a sorting execution mechanism 40 located at the sorting station. Pears are carried and driven by the conveying and rotating mechanism 10 to move in the direction of the arrow while simultaneously rotating, passing sequentially through the detection station and the sorting station. The fruit discharge roller 101 is located between the upstream and downstream conveyor belts, and the guide seat 43 serves as a transition component, forming a linear and compact layout of conveying-detection-sorting, without unnecessary turning points or conveying interruptions. To reduce the risk of pears getting stuck or piling up during transport, and to adapt to the high-speed transport requirements of industrial production lines (without needing to reduce speed for adaptation mechanism connection); the sorting execution mechanism 40 is located close to the connection between the fruit discharge roller 101 and the downstream conveyor belt, shortening the spatial distance between the identification result output and the sorting action execution, avoiding sorting delays or misjudgments caused by excessive transport distance, and improving sorting accuracy under dynamic working conditions; the toothed area of ​​the conveyor belt and the fruit discharge tray, along with the follow-up roller, work together to form an independent transport channel for each pear, avoiding multiple fruits being squeezed or crossing channels, and ensuring sorting order.

[0027] Figure 1 The middle fruit discharge roller 101 is located between the upstream conveyor belt 103 and the downstream conveyor belt 104, and 40 is located between the fruit discharge roller 101 and the downstream conveyor belt 104. Figure 2The center of the fruit tray 110 is mounted together with the shaft of the drive motor 102, which is fixed to the frame. Multiple teeth 111 are evenly distributed around the periphery of the fruit tray 110, with a tooth groove area between adjacent teeth 111. A side frame 113 is fixed to the side wall of each tooth 111, and a follower pulley 114 is installed in each tooth groove area. The end of the follower pulley 114 is mounted on the side frame 113. A follower roller 115 is installed on the side area of ​​the tooth groove on the side frame 113. A fixed toothed plate 116 is fixed to the frame, with an arc-shaped toothed rack on its upper side. The fixed toothed plate 116 can mesh with the teeth on the lower surface of the follower pulley 114 on the upper side. When the fruit tray 110 is driven to rotate by the drive motor 102, it can drive all the follower pulleys 114. During the revolution, when the follower pulley 114 located on the upper side meshes with the fixed toothed plate 116, the follower pulley 114 rolls along the surface of the fixed toothed plate 116, thereby realizing the self-rotation of the follower pulley 114 on the upper side, which in turn drives the pear at that position to rotate. The fruit tray is driven by the drive motor 102 to revolve, and the follower pulley 114 on the upper side meshes with the fixed toothed plate 116 on the frame. Through the mechanical linkage of revolution driving meshing → meshing driving rotation, the pears achieve synchronous rotation during the transmission process. Without additional self-rotation drive components, it achieves synchronous revolution and rotation flipping through mechanical transmission, resulting in a simplified structure, low energy consumption, and avoidance of the control complexity of multiple motors working together. The rotation process is smooth (the follower pulley meshes with the arc-shaped rack and pinion, and the speed is uniform), ensuring that the surface and shallow features of the pear can be collected by the fusion sensing unit 20 (sensor 21) 360° without dead angles, solving the problem of feature omission caused by static collection or unidirectional transmission, and providing a data foundation for subsequent high-precision identification. The cooperation between the follower roller 115 and the toothed head 111 forms a flexible limit on the pear, preventing deviation or collision during rotation and ensuring product integrity.

[0028] like Figure 1In the sorting mechanism 40, a cylinder 41, a baffle 42, and a guide seat 43 are included. In its natural state, the cylinder 41 controls the baffle 42 to be in the lower position. At this time, the guide seat 43 connects the fruit-discharging roller 101 and the downstream conveyor belt 104, providing transitional guidance. Qualified pears are discharged along the downstream conveyor belt 104 into the end-of-line variety A collection box. However, when a defective pear is detected and needs to be removed, the cylinder 41 controls the baffle 42 to rise, thus blocking the fruit-discharging roller 101 between it and the downstream conveyor belt 104. This forces the defective pear to flow downwards along the baffle 42 and enter the variety B collection box. In its natural state, the baffle 42 is in the lower position, and the guide seat 43 facilitates the transition. When sorting is triggered, the cylinder drives the baffle to rise rapidly, preventing the pear from entering the downstream conveyor belt and guiding it into the designated collection box. With a minimalist structure (only three core components: cylinder, baffle, and guide seat), and no complex transmission mechanism, it reduces potential failure points in industrial environments (such as jamming and wear), resulting in low maintenance costs and high reliability. The cylinder drive offers fast response speed, meeting the dynamic sorting requirements of high-speed sorting lines. The arc transition design between the baffle and the guide seat prevents pears from colliding, rolling, or being scratched during sorting, ensuring product quality. At the same time, the fixed transition function of the guide seat ensures a stable transmission path for qualified pears, eliminating the risk of deviation.

[0029] like Figure 1 In this process, the fusion sensing unit 20 is crucial for acquiring the pear's feature information, and its core is a dynamic anti-interference optical coherence tomography sensor 21. This sensor is designed to simultaneously acquire fused information of the pear's surface morphology and shallow internal physicochemical characteristics during the pear's movement.

[0030] Figure 3 This is a schematic diagram of a specific embodiment of the dynamic anti-interference optical coherence tomography sensor 21. In this embodiment, the sensor is a miniaturized swept-frequency optical coherence tomography module. It includes a swept-frequency laser source 211, an optical fiber coupler 212, and a Michelson interferometer optical path composed of a sample arm 213 and a reference arm 214. The light emitted from the sample arm 213 passes through a collimator 2131 and a two-dimensional MEMS scanning mirror 2132 before illuminating the moving pear surface. The returned light interferes with the reference light from the reference arm 214 at the coupler 212, and is converted into an electrical signal, i.e., the original interference spectrum signal, by a balanced photodetector 215. A fixed reference mirror 2141 is integrated in the reference arm 214 to generate a stable reference calibration signal for real-time correction of motion artifacts.

[0031] The processing and control unit 30, for example, employs a high-performance embedded industrial computer, and is equipped with an end-to-end variety identification model trained based on a deep transfer learning algorithm. The sensor 21 transmits the acquired raw one-dimensional interference spectral signal sequence corresponding to each pear to the processing and control unit 30 in real time via a high-speed data interface.

[0032] In this embodiment, the main workflow of the pear variety identification and sorting method is as follows: S1. Signal Acquisition and Preprocessing: When the pear moves through the detection area, sensor 21 performs high-speed scanning to acquire the raw interference spectrum signal. The processing and control unit 30 uses the calibration signal generated by the fixed reference mirror 2141 to perform real-time motion artifact compensation on the acquired signal to obtain a stable raw one-dimensional interference spectrum signal sequence.

[0033] S2. Deep Transfer Learning Model Inference: The preprocessed signal sequence is directly input into the end-to-end variety identification model. This model is a neural network trained through deep transfer learning, configured to directly process one-dimensional signal sequences and extract deep features.

[0034] S3. Variety Identification and Sorting Execution: The model outputs the variety classification result and confidence level of the pear. Based on the identification result and the location of the pear in the conveyor channel, the processing and control unit 30 issues an instruction to the corresponding sorting execution mechanism 40 at an appropriate time to sort the pear into the corresponding variety collection box.

[0035] Figure 4 This is a schematic block diagram illustrating the principle of real-time motion artifact correction in software, demonstrating how a fixed reference signal can be used to dynamically compensate for vibration noise and ensure the data quality of the input model.

[0036] Figure 5 This diagram illustrates a preferred architecture for the end-to-end variety identification model. The model is a convolutional neural network specifically designed for processing one-dimensional spectral signals. After the signal sequence passes through the input layer, it undergoes feature abstraction through multiple convolutional layers. The core of the network includes an attention mechanism module for adaptively weighting the importance of different bands in the spectral sequence. Finally, the variety category is output through a fully connected layer.

[0037] The training of this model fully utilizes deep transfer learning strategies, specifically divided into two stages: The first stage (pre-training - knowledge transfer): On a large-scale, general optical signal dataset (or a dataset in a related field), a basic neural network model (such as a one-dimensional variant of ResNet) is pre-trained to enable the model to learn general feature extraction capabilities, such as basic representations of edges, textures, and spectral patterns.

[0038] The second stage (domain adaptation-fine-tuning) involves transferring the pre-trained model to the pear variety recognition task. Using a relatively small number (e.g., 500-1000 samples per variety) of labeled pear OCT spectral signal samples, the top-level parameters of the model are fine-tuned. Simultaneously, parts of the lower-level network may be unfrozen to adapt it to the specific distribution of pear spectral signals and subtle differences between varieties. This transfer learning strategy significantly reduces the amount of labeled data required for a specific task and improves the model's convergence speed and generalization performance.

[0039] Building upon the aforementioned deep transfer learning framework, and further, in order to address the high similarity of spectral features among pear varieties and the intra-class differences within the same variety due to varying growth conditions, this invention introduces a feature space optimization and hard sample mining strategy based on metric learning in a preferred embodiment.

[0040] During the model fine-tuning phase, in addition to using the standard cross-entropy classification loss, a triplet loss function was introduced. This loss function encourages the network to bring pear samples of the same variety (anchor and positive samples) closer together in the feature space, while pushing samples of different varieties (anchor and negative samples) further apart. More importantly, we designed a dynamic hard sample mining algorithm. In each training batch, instead of randomly selecting negative samples, we actively select those hard negative samples that are close to the anchor features but belong to different varieties, and those hard positive samples that are far from the anchor features but belong to the same variety, for calculating the loss. This forces the network to focus on learning to distinguish the most easily confused variety feature boundaries.

[0041] Because this feature space optimization strategy based on metric learning and hard sample mining is integrated into the fine-tuning stage of deep transfer learning, the trained variety identification model exhibits better intra-class aggregation and inter-class separability in its learned feature representations within the embedding space. Consequently, when faced with varieties such as Korla fragrant pear and Aksu fragrant pear, which are extremely similar in appearance and spectrum, the model's discriminative ability is significantly enhanced, further improving the Top-1 identification accuracy by approximately 3-5 percentage points and significantly reducing the missorting rate due to variety confusion.

[0042] Another area for further optimization is to include an incremental learning module, considering that new varieties may be gradually introduced into the production line. This is to reduce the inconvenience of needing to collect large amounts of data and retrain the model from scratch each time a new variety is added. When a new variety needs to be identified, the system uses a small amount of newly collected sample data of that variety to incrementally fine-tune the existing model. This process typically freezes most of the original model's layers, primarily training a new classification head branch, and incorporates knowledge distillation techniques to mitigate the forgetting of the original variety recognition performance. This allows the device to quickly adapt to new sorting needs at a lower marginal cost, demonstrating good scalability.

[0043] The implementation process of the above scheme is as follows: Figure 6 As shown, this scheme, by employing real-time dynamic correction technology based on a fixed reference mirror and a single-sensor fusion perception scheme, avoids the problem of asynchronous information from multiple sensors from the source, thus providing a stable and high-quality input signal for the deep learning model. This is a fundamental prerequisite for achieving high-precision online recognition. By constructing a processing flow from raw signal to end-to-end deep neural network and adopting a deep transfer learning training paradigm, the system can directly learn optimal features from the data and significantly reduce its dependence on the amount of labeled data in the target domain. Therefore, using only a few hundred labeled samples per variety, the model can achieve a variety recognition accuracy of over 98% at production line speeds, solving the core challenges of small-sample learning and model generalization. Furthermore, by introducing a feature space optimization strategy based on metric learning and hard-sample mining during the fine-tuning stage of transfer learning, the model can more precisely characterize the discriminative boundaries between different varieties, thus exhibiting stronger distinguishing ability when facing highly similar varieties and elevating the robustness of recognition in complex scenarios to a new level. Therefore, this embodiment provides a complete solution from sensing hardware to algorithm model, which not only realizes automated, high-precision online identification and sorting of pear varieties, but also achieves a good balance in terms of engineering feasibility, economic cost and intelligence level, providing strong technical equipment support for the branding and refined processing of the pear industry.

[0044] It should be noted that the above embodiments and accompanying drawings are merely illustrative examples of the core principles and key structures of the identification and sorting device of the present invention. The accompanying drawings are simplified schematic diagrams, intended to clearly illustrate the structural, process, or data flow relationships related to the innovative points of the technical solution, and are not intended to limit the complete form of the actual product. This specification focuses on the innovative technical means necessary to achieve the invention's objectives and solve the technical problems. While auxiliary or common-sense details such as frame structure, conveyor belt design, electrical connections, enclosure protection, and standard communication protocol applications, which can be implemented by those skilled in the art without creative effort, are not described in detail, they should be understood as naturally included in the specific implementation of the present invention and fall within the protection and implementation scope of this technical solution.

Claims

1. A method for identifying and sorting different varieties of pears based on a deep transfer learning algorithm, characterized in that, Includes the following steps: S1. When the pear moves and rotates through the detection area, its surface is scanned using an optical coherence tomography sensor to collect the original one-dimensional interference spectrum signal sequence; S2. Input the processed original one-dimensional interference spectrum signal sequence into a pre-trained variety identification neural network model; wherein, the variety identification neural network model is trained using a deep transfer learning algorithm, and the training includes: pre-training the basic neural network on a general dataset to obtain general feature extraction capabilities, and then fine-tuning the pre-trained model on a pear variety labeled spectrum dataset; S3. The variety identification neural network model processes the input signal sequence and outputs the variety identification result of pear; S4. Based on the variety identification results, control the sorting execution mechanism to sort the pears.

2. The method according to claim 1, characterized in that, In step S1, while acquiring the original one-dimensional interference spectrum signal sequence, a reference calibration signal generated by a fixed reference surface inside the sensor is also acquired simultaneously, and the acquired signal is corrected for motion artifacts in real time based on the reference calibration signal.

3. The method according to claim 1, characterized in that, In step S2, the variety identification neural network model is a one-dimensional convolutional neural network; the one-dimensional convolutional neural network integrates an attention mechanism module, which is used to adaptively weight the spectral dimension or feature channel dimension of the signal sequence.

4. The method according to claim 1, characterized in that, In the fine-tuning step of the deep transfer learning algorithm, the loss function used includes the triplet loss function, which is used to optimize the feature space so that the feature distance of samples of the same variety decreases and the feature distance of samples of different varieties increases.

5. A pear variety identification and sorting device based on a deep transfer learning algorithm, characterized in that, include: A conveying and rotating mechanism is used to move and rotate individual pears through the inspection station; An optical coherence tomography sensor is installed at the detection station to collect the original one-dimensional interference spectrum signal sequence of the pear. The processing and control unit is communicatively connected to the sensor and has a variety identification neural network model trained using a deep transfer learning algorithm deployed within it. The processing and control unit is configured to process the signal sequence, identify varieties using the model, and generate sorting instructions. The sorting execution mechanism is set at the sorting station and is communicatively connected to the processing and control unit, and is used to sort pears according to the sorting instructions.

6. The apparatus according to claim 5, characterized in that, The optical coherence tomography sensor is a swept-frequency optical coherence tomography module, and its reference arm integrates a fixed reference mirror to provide a reference signal for real-time motion artifact correction.

7. The apparatus according to claim 5, characterized in that, The processing and control unit is also equipped with a model incremental learning module, which is used to perform incremental fine-tuning based on the existing model using a small amount of sample data of newly added varieties, so as to expand the range of identifiable varieties.

8. The apparatus according to any one of claims 5 to 7, characterized in that, The training process of the variety identification neural network model includes: first, performing self-supervised or supervised pre-training on a dataset containing multiple types of optical or spectral signals; then, fine-tuning the model using pear variety-labeled spectral data for domain adaptation; and during the fine-tuning stage, employing a hard sample mining strategy to select training samples.

Images