Multi-object recognition and sorting system and method

Abstract

The present invention relates to a multi-object recognition and sorting system and method. This invention utilizes a small sample recognition algorithm as a parameter-free model to identify unknown objects. This not only improves the accuracy of object feature extraction through feature fusion, but also eliminates the need to retrain the entire network for the introduction of new objects, requiring only feature fusion. This system therefore exhibits improved generalization capabilities for identifying new objects, improving overall system efficiency and addressing issues such as inaccurate object recognition, heavy manual sorting workload, and high error rates.

Description

Technical Field

[0001] The present invention relates to the technical field of object recognition and sorting, and in particular to a multi-object recognition and sorting system and method. Background Art

[0002] Industrial intelligence plays a vital role in improving the efficiency of industrial production lines and assembly lines. With the rapid development of e-commerce and the increase in labor costs, the use of industrial intelligent equipment and robots is becoming more and more widespread.

[0003] Currently, most robot recognition systems operate through a teach pendant or offline programming. Changes in the type or location of the target object require re-teaching or reprogramming, resulting in a lack of environmental awareness and adaptability. Compared to traditional recognition systems, convolutional neural networks offer greater robustness, improving both intelligent environmental perception and the accuracy of multi-object recognition.

[0004] However, traditional convolutional neural networks need to retrain the network for newly added objects, which reduces the efficiency of the system. Summary of the Invention

[0005] The purpose of the present invention is to provide a multi-object recognition and sorting system and method, which can improve the accuracy of extracting object features. Moreover, when introducing new objects, there is no need to retrain the entire network, only feature fusion is required. Therefore, it has better generalization ability for the recognition of new objects and can improve the recognition and sorting efficiency of multiple objects.

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

[0007] A multi-object recognition and sorting system, comprising:

[0008] A data set automatic collection module is used to collect an object image set and pre-process the object image set; the object image set contains a variety of objects to be identified;

[0009] an image processing-recognition module, connected to the automatic data set collection module, for extracting feature information of each object in the preprocessed object image set using a convolutional neural network model based on a small sample recognition algorithm, for determining fusion features based on the feature information, and for determining classification information of each object in the preprocessed object image set based on the fusion features;

[0010] The robotic arm sorting module is connected to the image processing-recognition module and is used to complete the sorting work of objects according to the classification information.

[0011] Preferably, the data set automatic collection module includes:

[0012] A camera, configured to capture an image set of the object;

[0013] An image processing unit is connected to the camera and the image processing-recognition module respectively, and is used to pre-process the object image set according to the object mask image to obtain an image containing only the object.

[0014] Preferably, the image processing-recognition module includes:

[0015] a feature extraction unit connected to the image processing unit, configured to extract feature information of each object in the preprocessed object image set using a convolutional neural network model based on a small sample recognition algorithm;

[0016] a mapping weight determination unit, connected to the feature extraction unit, for determining a mapping weight for feature information of each object based on an image mask corresponding to the extracted feature information of each object;

[0017] a mapping feature determination unit connected to the mapping weight determination unit, configured to determine a mapping feature based on the mapping weight of each object feature information, an image mask corresponding to the extracted feature information of each object, and the feature information of each object;

[0018] a projection weight determination unit, connected to the feature extraction unit, for determining the projection weight of each dimension in the feature information of each object;

[0019] a fusion feature unit, connected to the mapping feature determination unit and the projection weight determination unit, respectively, and configured to determine a fusion feature according to the mapping feature and the projection weight;

[0020] An object category template storage unit, configured to store a preset object category template; the object category template includes an object category and an object template fusion feature;

[0021] a similarity value determining unit, connected to the object category template storage unit and the fusion feature unit, and configured to determine a similarity value between the fusion feature and the object template fusion feature;

[0022] The object recognition unit is connected to the object category template storage unit and the similarity value determination unit, respectively, and is used to determine the category of each object in the object image set according to the similarity value; wherein, when the similarity value is greater than a preset value, the object category in the object category template corresponding to the maximum similarity value is determined to be the category of the object to be recognized; when the similarity value is less than the preset value, the category of the object to be recognized is determined to be a new object category, and the new object category is stored in the object category template storage unit as a new object category template.

[0023] Preferably, the robotic arm sorting module includes:

[0024] A sorting processing unit connected to the image processing-recognition module is used to generate sorting instructions based on the classification information of each object; the sorting instructions include: object picking instructions and object placement instructions;

[0025] An execution unit is connected to the sorting processing unit and is used to complete object sorting according to the sorting instruction.

[0026] Preferably, the execution unit includes:

[0027] a robotic arm connected to the sorting processing unit and configured to perform movement actions according to the sorting instructions;

[0028] A mechanical claw is fixedly arranged on the mechanical arm and connected to the sorting processing unit, and is used for performing a claw release action according to the sorting instruction.

[0029] In addition, the present invention also provides a multi-object recognition and sorting method for use in the multi-object recognition and sorting system provided above, the method comprising:

[0030] Collecting an object image set and preprocessing the object image set; the object image set contains a variety of objects to be identified;

[0031] Based on a small sample recognition algorithm, a convolutional neural network model is used to extract feature information of each object in the preprocessed object image set;

[0032] determining a fusion feature based on the feature information;

[0033] determining classification information of each object in the preprocessed object image set according to the fusion features;

[0034] The objects are sorted according to the classification information.

[0035] Preferably, determining the fusion feature based on the feature information specifically includes:

[0036] determining a mapping weight for the feature information of each object based on an image mask corresponding to the extracted feature information of each object;

[0037] Determining mapping features based on mapping weights of feature information of each object, an image mask corresponding to the extracted feature information of each object, and the feature information of each object;

[0038] Determine the projection weight of each dimension in the feature information of each object;

[0039] A fusion feature is determined according to the mapping feature and the projection weight.

[0040] Preferably, determining the classification information of each object in the pre-processed object image set according to the fusion feature specifically includes:

[0041] Obtaining a preset object category template; the object category template includes object category and object template fusion features;

[0042] Determining a similarity value between the fused feature and the object template fused feature;

[0043] The category of each object in the object image set is determined according to the similarity value.

[0044] Preferably, determining the category of each object in the object image set according to the similarity value specifically includes:

[0045] Get the preset value of similarity;

[0046] When the similarity value is greater than the preset value, determining that the object category in the object category template corresponding to the maximum similarity value is the category of the object to be identified;

[0047] When the similarity value is less than the preset value, the category of the object to be identified is determined to be a new object category, and the new object category is stored as a new object category template.

[0048] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0049] The multi-object recognition and sorting system and method provided by the present invention use the small sample recognition algorithm as a parameter-free model to realize the recognition of unknown objects. Not only does it improve the accuracy of extracting object features through feature fusion, but the introduction of new objects does not require retraining the entire network, only feature fusion is required. Therefore, it has better generalization ability for the recognition of new objects, improves the efficiency of the entire system, and thus solves the problems of inaccurate object recognition, large workload of manual sorting, and high error rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0051] Figure 1 This is a diagram showing the implementation framework of the multi-object recognition and sorting system provided by the present invention;

[0052] Figure 2A diagram illustrating the implementation of a multi-object recognition and sorting system according to an embodiment of the present invention;

[0053] Figure 3 A schematic diagram of the structure of a neural network model provided by an embodiment of the present invention;

[0054] Figure 4 This is a flow chart of the multi-object recognition and sorting method provided by the present invention. DETAILED DESCRIPTION

[0055] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments 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.

[0056] The purpose of the present invention is to provide a multi-object recognition and sorting system and method, which can improve the accuracy of extracting object features. Moreover, when introducing new objects, there is no need to retrain the entire network, only feature fusion is required. Therefore, it has better generalization ability for the recognition of new objects and can improve the recognition and sorting efficiency of multiple objects.

[0057] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0058] like Figure 1 As shown, the multi-object recognition and sorting system provided by the present invention includes: an automatic data set collection module 100, an image processing-recognition module 101 and a robotic arm sorting module 102.

[0059] The automatic data set collection module 100 is used to collect and pre-process an object image set containing a variety of objects to be identified.

[0060] The image processing-recognition module 101 is connected to the data set automatic collection module 100 and is used to extract the feature information of each object in the pre-processed object image set using a convolutional neural network model based on a small sample recognition algorithm, and is used to determine the fusion feature based on the feature information, and is also used to determine the classification information of each object in the pre-processed object image set based on the fusion feature. The specific structure of the neural network model used in this embodiment is as follows: Figure 2 shown.

[0061] The robotic arm sorting module 102 is connected to the image processing-recognition module 101 and is used to complete the sorting of objects according to the classification information.

[0062] In order to improve the accuracy of object recognition, the automatic data set collection module 100 used in the present invention can be configured to include: a camera and an image processing unit. In this embodiment, the camera can be a Kinect camera.

[0063] A camera is used to collect an image set of an object.

[0064] The image processing unit is connected to the camera and the image processing-recognition module 101 respectively, and is used to pre-process the object image set according to the object mask image to obtain an image containing only the object.

[0065] Furthermore, in order to improve the accuracy of object recognition while improving the efficiency of object recognition, the image processing-recognition module 101 adopted in the present invention includes: a feature extraction unit, a mapping weight determination unit, a mapping feature determination unit, a projection weight determination unit, a fusion feature unit, an object category template storage unit, a similarity value determination unit and an object recognition unit.

[0066] The feature extraction unit is connected to the image processing unit and is used to extract feature information of each object in the preprocessed object image set based on a small sample recognition algorithm and a convolutional neural network model.

[0067] The mapping weight determination unit is connected to the feature extraction unit and is used to determine the mapping weight of the feature information of each object based on the image mask corresponding to the extracted feature information of each object.

[0068] The mapping feature determination unit is connected to the mapping weight determination unit and is used to determine the mapping feature according to the mapping weight of each object feature information, the image mask corresponding to the extracted feature information of each object, and the feature information of each object.

[0069] The projection weight determination unit is connected to the feature extraction unit and is used to determine the projection weight of each dimension in the feature information of each object.

[0070] The fusion feature unit is connected to the mapping feature determination unit and the projection weight determination unit respectively, and is used to determine the fusion feature according to the mapping feature and the projection weight.

[0071] The object category template storage unit is used to store a preset object category template. The object category template includes an object category and an object template fusion feature.

[0072] The similarity value determination unit is connected to the object category template storage unit and the fusion feature unit, and is used to determine the similarity value between the fusion feature and the object template fusion feature.

[0073] The object recognition unit is connected to the object category template storage unit and the similarity value determination unit, and is configured to determine the category of each object in the object image set based on the similarity value. When the similarity value is greater than a preset value, the object category in the object category template corresponding to the maximum similarity value is determined as the category of the object to be recognized. When the similarity value is less than the preset value, the category of the object to be recognized is determined to be a new object category, and the new object category is stored in the object category template storage unit as a new object category template.

[0074] In order to improve the accuracy of object sorting, the robotic arm sorting module 102 used in the present invention may include: a sorting processing unit and an execution unit.

[0075] The sorting processing unit is connected to the image processing and recognition module 101 and is used to generate sorting instructions based on the classification information of each object. The sorting instructions include: object picking instructions and object placement instructions.

[0076] The execution unit is connected to the sorting processing unit and is used to complete the object sorting according to the sorting instruction.

[0077] Among them, the execution unit includes: a robotic arm and a robotic claw.

[0078] The robotic arm is connected to the sorting processing unit and is used to perform movement according to the sorting instructions.

[0079] The mechanical claw is fixed on the robot arm and connected to the sorting processing unit to perform the claw release action according to the sorting instruction. The installation position of the robot arm, mechanical claw and camera is as follows: Figure 3 As shown, Figure 3 In the figure, reference numeral 1 is the robotic arm, which can be extended and rotated. Reference numeral 2 is the working area, where objects to be identified and sorted are placed. Reference numeral 3 is the robotic gripper. In this embodiment, the robotic gripper is a two-fingered hand located at the end of robotic arm 1. Reference numeral 4 is the Kinect camera, which is fixed at a height of 1000 mm from the working area.

[0080] To further improve the integration of the entire system, the functions of the above-mentioned image processing-recognition module 101, sorting processing unit and execution unit can be implemented by a host computer or intelligent controller (such as a main control computer).

[0081] The following is based on Figure 3 Taking the system structure shown in FIG. 1 as an example, the specific implementation process of the multi-object recognition and sorting system provided above is described.

[0082] The automatic dataset collection module 100 uses an image acquisition device to automatically acquire images of target objects and transmits the acquired image information to the image processing-recognition module 101, thereby achieving the purpose of quickly acquiring training data and solving the problem of difficult dataset production. In addition, during the process of identifying and sorting objects to be identified, the automatic dataset collection module 100 uses the object images in the working area to form an object image set.

[0083] The specific working process of the data set automatic collection module 100 for collecting training samples is as follows:

[0084] An RGB-D camera is installed above the work platform to collect image data of multiple target objects (i.e., object samples used in training) at different positions, postures, and stacking conditions.

[0085] The robot uses a gripper to fix the object at the end of the robotic arm, and drives the robotic arm to change the object's posture, simulating the possible postures of the object in the real scene. To ensure that the weights trained by the convolutional neural network have good generalization, the object's posture and stacking conditions are collected as much as possible;

[0086] When performing object recognition, the test image only contains the object area. Therefore, an image containing only the object is created based on the object mask image to reduce the appearance of background noise, and images of the same category are placed in the same folder.

[0087] Image processing-recognition module 101 is based on a small sample recognition algorithm, using an improved convolutional neural network to calculate the similarities and differences between samples, and by training the weights corresponding to the features of each sample. Based on the obtained weights, a preliminary fusion is performed, and mapping is performed according to the features of different categories to generate fused features with greater differences. Finally, after training the fused features, the similarity value determination unit is used to calculate the category with the highest similarity among various objects to complete the recognition of the target object. For newly added objects, only the corresponding images need to be collected and added to the data set, without the need to retrain the entire network, which has strong adaptability to new objects.

[0088] Among them, such as Figure 2 As shown in the figure, the components of the convolutional neural network are as follows:

[0089] Input layer: input image matrix;

[0090] Convolutional layer: consists of five stacked convolution blocks, each of which contains 64 convolution kernels of size 3*3 for feature extraction;

[0091] Pooling layer: The size is 2*2, which is used to reduce the amount of calculation in the previous step and enhance the recognition and classification effect;

[0092] Activation layer: The activation function performs nonlinear operations to improve data processing efficiency and avoid the problem of gradient disappearance;

[0093] Output layer: Outputs the target recognition and classification results, which include the feature information of the recognition samples.

[0094] Compared with other neural networks, the above convolutional neural network can continuously integrate the relationship between basic features to obtain features that can represent the essence of the input image.

[0095] After obtaining feature information using the convolutional neural network model, it is necessary to determine the fusion features based on the feature information. The specific process of determining the fusion features is as follows:

[0096] The weight w corresponding to the features of each sample obtained by calculating the output layer of the convolutional neural network in , perform weighted fusion of different sample features. Then generate a set of weights w based on the relationship between the features of different types of samples ex , selecting features with high variance, ultimately obtaining features that best represent the differences between different object classes. For objects of the same class, the original features are selected using a linear transformation and the ReLU function. Features below the threshold are set to 0. The corresponding weight is calculated based on the number of samples that meet the feature criteria. A larger number indicates a more prominent feature and a larger corresponding weight.

[0097] The initial fusion features are obtained based on the importance within the feature class. The calculation method is as follows:

[0098]

[0099] Among them, N i represents the total number of samples of the i-th category, D represents the characteristic dimension of a single sample, Represents the mapping weight of the k-th dimension feature of the i-th class sample, v' i,j,k Represents the feature value of the kth dimension of the jth sample in the i-th category, a i,j,k is the corresponding mask. When the feature value is greater than 0, its value is 1, otherwise it is 0. The feature v of all dimensions after mapping i,k Together they form the C*D dimensional fusion feature F.

[0100] For the features of samples of different categories, the features with higher differences are selected by feature weights. Each category of the sample obtains a fused feature F, and then a set of projection weights w is calculated based on the differences between the features of samples of different categories. ex , projecting each category feature into the above convolutional neural network, the features that strongly distinguish samples are retained. The specific calculation process is as follows: first calculate the feature variance of each dimension, and then express different weights according to the difference of each feature:

[0101]

[0102] Among them, Var represents variance, V k Represents a column vector consisting of the k-th dimension features in a class of objects, represents the k-th dimension weight of the inter-class projection weight of the sample feature, and F' is the final fusion feature;

[0103] In the process of determining similarity (i.e., similarity value), this embodiment uses a similarity determination unit, which is used to calculate the similarity between the test sample image and various template images. The similarity determination unit includes a subtractor and two fully connected layers. Specifically, the subtractor is used to subtract the test sample feature vector from the feature vectors of each support sample. The difference between the feature vectors is then combined using two fully connected layers to obtain the similarity with each type of sample. The highest similarity category for each type of object is calculated to complete object recognition and classification.

[0104] For newly added objects, we only need to extract the features of the new objects and perform feature fusion based on the weights. The fused features are input into the similarity value determination unit to calculate the similarity with each object. There is no need to retrain the entire network, so it has better generalization ability for new objects.

[0105] The robot arm sorting module 102 is a module that transmits the identification and classification information obtained by the digital image processing-recognition module 101 to the robot arm in the form of instructions to complete the sorting of multiple objects. Specifically,

[0106] The main control computer generates the robot's "pick" and "release" trajectories through joint space trajectory planning based on the grasping posture of the target object and sends them to the robot side; the robot side records the grasping posture and completes the grasping operation.

[0107] In addition, the present invention also provides a multi-object recognition and sorting method, which can be applied to the multi-object recognition and sorting system provided above, such as Figure 4 As shown, the method includes:

[0108] Step 200: Collect an object image set and pre-process the object image set. The object image set contains multiple objects to be identified.

[0109] Step 201: Based on a small sample recognition algorithm, a convolutional neural network model is used to extract feature information of each object in the preprocessed object image set.

[0110] Step 202: Determine fusion features based on feature information.

[0111] Step 203: Determine the classification information of each object in the pre-processed object image set based on the fusion features.

[0112] Step 204: Complete the sorting of objects according to the classification information.

[0113] Wherein, step 202 specifically includes:

[0114] A mapping weight of the feature information of each object is determined based on an image mask corresponding to the extracted feature information of each object.

[0115] The mapping feature is determined according to the mapping weight of the feature information of each object, the image mask corresponding to the extracted feature information of each object, and the feature information of each object.

[0116] Determine the projection weight of each dimension in the feature information of each object.

[0117] The fusion features are determined based on the mapping features and projection weights.

[0118] Step 203 specifically includes:

[0119] Get the preset object category template. The object category template includes the object category and the object template fusion feature.

[0120] Determine the similarity between the fused features and the object template fused features.

[0121] Determine the category of each object in the object image set based on the similarity value, specifically including:

[0122] Gets the preset value for similarity.

[0123] When the similarity value is greater than a preset value, the object category in the object category template corresponding to the maximum similarity value is determined to be the category of the object to be identified.

[0124] When the similarity value is less than a preset value, the category of the object to be identified is determined to be a new object category, and the new object category is stored as a new object category template.

[0125] The specific calculation process of the multi-object recognition and sorting method provided by the present invention can be found in the calculation method disclosed in the above-mentioned multi-object recognition and sorting system.

[0126] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0127] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A multi-object recognition and sorting system, characterized in that: include: A data set automatic collection module is used to collect an object image set and pre-process the object image set; the object image set contains a variety of objects to be identified; an image processing-recognition module, connected to the automatic data set collection module, for extracting feature information of each object in the preprocessed object image set using a convolutional neural network model based on a small sample recognition algorithm, for determining fusion features based on the feature information, and for determining classification information of each object in the preprocessed object image set based on the fusion features; Among them, based on the small sample recognition algorithm, the convolutional neural network model is used to calculate the similarities and differences between samples, and the weights corresponding to the features of each sample are trained; based on the obtained weights, preliminary fusion is performed, and mapping is performed according to the features of different categories to generate fusion features with greater differences; finally, by training the fused features, the category with the highest similarity of each type of object is calculated to complete the recognition of the target object; for newly added objects, the features of the new objects are extracted, and the features are fused based on the weights, and the fused features are used to calculate the similarity with each object; among them, for the features of samples of different categories, features with high differences are selected by feature weights; each category of the sample obtains a fused feature, and then a set of projection weights are calculated based on the differences between the features of samples of different categories, and the features of each category are projected into the convolutional neural network model, and features with strong distinction between samples are retained; A robotic arm sorting module, connected to the image processing-recognition module, for sorting objects according to the classification information; The image processing-recognition module includes: a similarity value determination unit for determining a similarity value between the fusion feature and the object template fusion feature; the object template fusion feature belongs to a preset object category template; The similarity value determination unit includes a subtractor and two fully connected layers; the subtractor is used to subtract the feature vector of the test sample from the feature vectors of each support sample, and then the two fully connected layers are used to synthesize the differences in the feature vectors to obtain the similarity with each type of sample.

2. The multi-object recognition and sorting system according to claim 1, characterized in that: The data set automatic collection module includes: A camera, configured to capture an image set of the object; An image processing unit is connected to the camera and the image processing-recognition module respectively, and is used to pre-process the object image set according to the object mask image to obtain an image containing only the object.

3. The multi-object recognition and sorting system according to claim 2, characterized in that: The image processing-recognition module also includes: a feature extraction unit connected to the image processing unit, configured to extract feature information of each object in the preprocessed object image set using a convolutional neural network model based on a small sample recognition algorithm; a mapping weight determination unit, connected to the feature extraction unit, for determining a mapping weight for feature information of each object based on an image mask corresponding to the extracted feature information of each object; a mapping feature determination unit connected to the mapping weight determination unit, configured to determine a mapping feature based on the mapping weight of each object feature information, an image mask corresponding to the extracted feature information of each object, and the feature information of each object; a projection weight determination unit, connected to the feature extraction unit, for determining the projection weight of each dimension in the feature information of each object; a fusion feature unit, connected to the similarity value determination unit, the mapping feature determination unit, and the projection weight determination unit, respectively, and configured to determine a fusion feature according to the mapping feature and the projection weight; an object category template storage unit, connected to the similarity value determination unit, and configured to store the preset object category template; the object category template further comprising an object category; The object recognition unit is connected to the object category template storage unit and the similarity value determination unit, respectively, and is used to determine the category of each object in the object image set according to the similarity value; wherein, when the similarity value is greater than a preset value, the object category in the object category template corresponding to the maximum similarity value is determined to be the category of the object to be recognized; when the similarity value is less than the preset value, the category of the object to be recognized is determined to be a new object category, and the new object category is stored in the object category template storage unit as a new object category template.

4. The multi-object recognition and sorting system according to claim 1, characterized in that: The robotic arm sorting module includes: A sorting processing unit connected to the image processing-recognition module is used to generate sorting instructions based on the classification information of each object; the sorting instructions include: object picking instructions and object placement instructions; An execution unit is connected to the sorting processing unit and is used to complete object sorting according to the sorting instruction.

5. The multi-object recognition and sorting system according to claim 4, characterized in that: The execution unit includes: a robotic arm connected to the sorting processing unit and configured to perform movement actions according to the sorting instructions; A mechanical claw is fixedly arranged on the mechanical arm and connected to the sorting processing unit, and is used for performing a claw release action according to the sorting instruction.

6. A method for identifying and sorting multiple objects, characterized in that: The method is applied to the multi-object recognition and sorting system according to any one of claims 1 to 5; the method comprises: Collecting an object image set and preprocessing the object image set; the object image set contains a variety of objects to be identified; Based on a small sample recognition algorithm, a convolutional neural network model is used to extract feature information of each object in the preprocessed object image set; determining a fusion feature based on the feature information; determining classification information of each object in the preprocessed object image set according to the fusion features; completing the sorting of objects according to the classification information; Among them, based on the small sample recognition algorithm, the convolutional neural network model is used to calculate the similarities and differences between samples, and the weights corresponding to the features of each sample are trained; based on the obtained weights, preliminary fusion is performed, and mapping is performed according to different category features to generate fusion features with greater differences; finally, by training the fused features, the category with the highest similarity among various objects is calculated to complete the recognition of the target object; for newly added objects, the features of the new objects are extracted, and the features are fused based on the weights, and the fused features are used to calculate the similarity with each object.

7. The multi-object recognition and sorting method according to claim 6, characterized in that: The determining of the fusion feature based on the feature information specifically includes: determining a mapping weight for the feature information of each object based on an image mask corresponding to the extracted feature information of each object; Determining mapping features based on mapping weights of feature information of each object, an image mask corresponding to the extracted feature information of each object, and the feature information of each object; Determine the projection weight of each dimension in the feature information of each object; A fusion feature is determined according to the mapping feature and the projection weight.

8. The multi-object recognition and sorting method according to claim 6, characterized in that: Determining classification information of each object in the preprocessed object image set according to the fusion feature specifically includes: Obtaining a preset object category template; the object category template includes object category and object template fusion features; Determining a similarity value between the fused feature and the object template fused feature; The category of each object in the object image set is determined according to the similarity value.

9. The multi-object recognition and sorting method according to claim 8, characterized in that: Determining the category of each object in the object image set according to the similarity value specifically includes: Get the preset value of similarity; When the similarity value is greater than the preset value, determining that the object category in the object category template corresponding to the maximum similarity value is the category of the object to be identified; When the similarity value is less than the preset value, the category of the object to be identified is determined to be a new object category, and the new object category is stored as a new object category template.

Images