A small sample target detection device and method for time-series species evolution

Abstract

The present invention discloses a small sample target detection device and method for time-series species evolution. The device comprises: a mutation module, which is used to generate basic candidate regions and new candidate regions for basic samples and new samples in a small sample, respectively, and then merge them to obtain merged candidate regions; a detector, which is used to identify the basic samples and new samples in the small sample according to the merged candidate regions. The present invention solves the problems of poor inheritance of information and knowledge of basic categories in the detection samples and lack of non-specific changes to new categories, and generates sufficient high-quality positive sample candidate regions, effectively optimizing the detection performance of the small sample target detection device.

Description

Technical Field

[0001] The invention relates to a small sample target detection device for time-series species evolution and also to a corresponding small sample target detection method, belonging to the technical field of computer vision. Background Art

[0002] In the past few decades, the success of deep learning has promoted breakthroughs in traditional object detection problems. However, most deep learning methods rely on a large amount of labeled data. When labeled data is scarce, these models may suffer from serious overfitting problems and show poor generalization ability.

[0003] In real-world scenarios, labeled data is very expensive, and samples of some categories are even difficult to collect. To address this problem, few-shot object detection has become a frontier research issue. In the few-shot learning problem, the model is trained based on the knowledge of the underlying categories with rich instances, using a very small number of new category images as support to solve new tasks. However, research progress in few-shot learning has mainly focused on image classification and recognition tasks, while the performance of the more challenging few-shot object detection task is far from satisfactory.

[0004] In the existing technology, the main problem faced by small-sample target detection tasks is that overfitting is prone to occur due to the small number of new category training samples. One of the important reasons is that in the existing small-sample target detection technology, the new category training samples have poor inheritance of information and knowledge of the basic category, and lack specific changes to the new category, which leads to the inability of small-sample target detection technology to generate enough high-quality positive sample candidate regions. Summary of the invention

[0005] The primary technical problem to be solved by the present invention is to provide a small sample target detection device for time-series species evolution.

[0006] Another technical problem to be solved by the present invention is to provide a small sample target detection method for time-series species evolution.

[0007] In order to achieve the above-mentioned object of the invention, the present invention adopts the following technical solutions:

[0008] According to a first aspect of an embodiment of the present invention, a small sample target detection device for time-series species evolution is provided, comprising:

[0009] The mutation module is used to generate basic candidate regions and new candidate regions for the basic samples and new samples in the small samples respectively, and then merge them to obtain merged candidate regions;

[0010] The detector is used to identify basic samples and new samples in small samples based on the merged candidate regions.

[0011] Preferably, the small sample target detection device further comprises:

[0012] The recombination module is used to recombine the distribution of basic samples and new samples to generate instantiated individuals;

[0013] The instantiated individuals are input as newly added new samples into the detector for training.

[0014] Preferably, the mutation module comprises:

[0015] The traditional region generation network is used to output a set of candidate boxes to form a basic candidate region;

[0016] The mutation region generation network is used to output a set of candidate boxes to form a new candidate region;

[0017] The merging submodule is used to select candidate boxes belonging to the basic sample or the new sample whose confidence level is higher than a preset threshold from the basic candidate area and the new candidate area to form a merged candidate area.

[0018] Preferably, the traditional region generation network and the mutation region generation network are set in parallel;

[0019] The initial parameters of the mutation region generation network are obtained by training the parameters of the traditional region generation network.

[0020] Preferably, the recombination module comprises:

[0021] The selection submodule is used to select the best similar basic sample distribution of each new sample distribution through the similarity between the new sample and the basic sample distribution to form a similar set;

[0022] An updating submodule is used to update the distribution feature information of the corresponding new sample according to the distribution feature information of the basic sample in the similar set, so as to obtain the abstract category information of the corresponding new sample;

[0023] The generation submodule is used to instantiate different individuals according to the abstract category information of each new sample.

[0024] Preferably, when the small sample target detection device is trained, the traditional region generation network is fully trained with basic samples to obtain fixed network parameters of the traditional region generation network;

[0025] When the mutation region generation network is trained by new samples, the initialization network parameters of the mutation region generation network are equal to the fixed network parameters of the traditional region generation network.

[0026] Preferably, the selection submodule is used to measure the similarity between the new sample distribution and the basic sample distribution by using the Wasserstein distance; and select the optimal similar basic sample distribution of each new sample distribution according to the greedy algorithm and the similarity between the new sample distribution and the basic sample distribution to form a similarity set.

[0027] Preferably, the distribution characteristic information of the basic samples includes the mean and covariance of the distribution characteristics of the basic samples;

[0028] The distribution characteristic information of the new sample includes the mean and covariance of the distribution characteristic of the new sample;

[0029] The abstract category information of the new sample includes the mean and covariance of the distribution characteristics of the abstract category.

[0030] Preferably, the generating submodule is specifically used to obtain a Gaussian distribution of the abstract category according to the abstract category information of the new sample;

[0031] According to the Gaussian distribution, sampling generates different individuals of instantiation of the abstract category.

[0032] According to a second aspect of an embodiment of the present invention, a small sample target detection method for time-series species evolution is provided, comprising the following steps:

[0033] The mutation module generates basic candidate regions and new candidate regions for the basic samples and new samples in the small samples respectively, and then merges them to obtain merged candidate regions;

[0034] The detector identifies basic samples and new samples in small samples based on the merged candidate regions.

[0035] In the present invention, after the mutation module fuses the basic sample and the new sample to obtain the merged candidate region, the recombination module fuses the information of the basic sample into the new sample, solving the problems of poor inheritance of information and knowledge of the basic category in the detection sample and lack of non-specific changes to the new category, and generating sufficient high-quality positive sample candidate regions, effectively optimizing the detection performance of the small sample target detection device. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a schematic diagram of the structure of a small sample target detection device in an embodiment of the present invention;

[0037] Figure 2 A schematic diagram showing the comparison of training and test samples of a new category in an embodiment of the present invention;

[0038] Figure 3 This is a flow chart of a small sample target detection method in an embodiment of the present invention;

[0039] Figure 4 Schematic diagram of the structure of a small sample target detection system in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The technical content of the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] The main task of target detection is to classify and locate targets in images. Existing small-sample target detection technology is trained based on a large number of annotated images, which limits its application and promotion in certain scenarios. For example, defect samples in industrial production are generally difficult to collect, and it is difficult to build a large-scale data set; in autonomous driving applications, there are only a few car samples in extremely bad weather conditions such as fog and rain, which makes it difficult to achieve car target detection in ordinary scenarios; in most medical applications, it is quite difficult to obtain high-quality medical data and high-quality annotations. In these scenarios where there are only a small number of samples as training sets for new categories, in order to achieve a certain detection accuracy for both the basic categories with a large number of annotations and the new categories with a small number of annotations, it is necessary to use improved small-sample target detection technology for target detection.

[0042] like Figure 1 As shown, the small sample target detection device provided by the embodiment of the present invention at least includes:

[0043] The mutation module S2 is used to generate basic candidate regions and new candidate regions for the basic samples and new samples in the small samples respectively, and then merge them to obtain merged candidate regions;

[0044] The detector O is used to identify the basic samples and new samples in the small samples according to the merged candidate regions.

[0045] In the embodiment of the present invention, the mutation module S2 fuses the basic sample and the new sample to generate a merged candidate region. The detector O can detect the basic category and the new category. The small sample target detection device provided by the present invention solves the problem of poor inheritance of information and knowledge of the basic category in the detection sample and lack of non-specific changes to the new category, thereby generating sufficient high-quality positive sample candidate regions and optimizing the detection performance of the small sample target detection device.

[0046] like Figure 1 and 2As shown, in an embodiment of the present invention, the basic samples include: images of dogs and images of cars, and the corresponding basic categories are: dogs and cars; the new samples are images of horses, and the corresponding new categories are: horses. In the above practical problems, such as autonomous driving, the basic categories are: common traffic lights at intersections, and the new categories are: some dangerous signs on mountain roads. It should be noted that the present invention takes dogs, cars and horses as examples and only makes adaptive explanations. All technical solutions described in the present invention that are used to solve various practical problems are within the protection scope of the present invention.

[0047] The following will be Figure 1 The small sample shown is used as a training sample to specifically explain the training process of the small sample target detection device provided by the present invention, thereby illustrating the structural setting of the small sample target detection device.

[0048] The small sample set C is divided into the basic category set C containing a large number of samples (b) and a small set of new categories C (n) ,in, Correspondingly, the training data set D can also be divided into D (b) ={x (b) ,y (b)} and D (n) ={x (n) ,y (n)}Two groups, among which D (n) The number is much smaller than D (b) Specifically, x represents an image, y = {(cls, box)}, cls represents the category label cls∈C of the object in image x, and box represents the bounding box of image x. The purpose of the small sample detection problem is to (n) Test image The detection model can predict the corresponding y (n) .

[0049] For example, in Figure 2 In the training samples shown, the basic categories are: dog, car and cat; the new category is horse, where the color of horse is mostly white. The small sample object detection device provided by the embodiment of the present invention is trained by the above training samples, and then the small sample object detection device is tested using a black horse image as a test sample, and the test output is "horse".

[0050] like Figure 1 As shown, the small sample target detection device provided by the embodiment of the present invention also includes: a backbone network E, which is used to extract feature representations of small samples as input of the mutation module S2.

[0051] In the embodiment of the present invention, the backbone network E extracts features from the basic samples and new samples to obtain corresponding feature representations. (b), extract its feature representation f through the backbone network E (b) . Given a new sample x (n) , extract its feature representation f through the backbone network E (n) , then f (b) and f (n) Input to mutation module S2.

[0052] like Figure 1 As shown, the mutation module S2 includes:

[0053] The traditional region generation network S21 is used to output a set of candidate boxes to form a basic candidate region;

[0054] The mutation region generation network S22 is used to output a set of candidate boxes to form a new candidate region;

[0055] The merging submodule S23 is used to select candidate boxes belonging to the basic sample or the new sample whose confidence level is higher than a preset threshold from the basic candidate regions and the new candidate regions to form a merged candidate region.

[0056] In the embodiment of the present invention, the traditional region generation network S21 may be a region generation network RPN.

[0057] After receiving the features of the basic samples and the new samples, the traditional region generation network S21 and the mutation region generation network S22 output the corresponding candidate box sets respectively. In the basic candidate region output by the traditional region generation network S21, the candidate boxes containing the basic categories are sorted according to their confidence; and the candidate boxes of the first set value are selected from high to low. In the new candidate region output by the mutation region generation network S22, the candidate boxes containing the new categories are sorted according to their confidence; and the candidate boxes of the second set value are selected from high to low. In the basic candidate region output by the traditional region generation network S21, the confidence of the candidate boxes containing the basic categories will be higher. In the new candidate region output by the mutation region generation network S22, the confidence of the candidate boxes containing the new categories will be higher.

[0058] There are several boxes on an image, each box is a part of the image, and the boxes output by the traditional region generation network S21 or the mutation region generation network S22 are candidate boxes. Each candidate box corresponds to a confidence level. The confidence level is the confidence level of whether the corresponding candidate box contains a basic category or a new category.

[0059] The basic candidate area includes a basic sample candidate frame set and a new sample candidate frame set, but tends to give a higher confidence to the candidate frame containing the basic sample.

[0060] The new candidate region also includes a basic sample candidate frame set and a new sample candidate frame set, but tends to give a higher confidence to the candidate frame containing the new sample.

[0061] like Figure 1 As shown, the traditional region generation network S21 and the mutation region generation network S22 are set in parallel;

[0062] The mutation region generation network S22 is obtained by training the parameters of the traditional region generation network S21.

[0063] When the small sample target detection device is trained:

[0064] After the traditional region generation network S21 is fully trained with basic samples, fixed network parameters of the traditional region generation network S21 are obtained;

[0065] When the mutation region generation network S22 is trained by new samples, the initialization network parameters of the mutation region generation network S22 are equal to the fixed network parameters of the traditional region generation network S21.

[0066] In the embodiment of the present invention, basic training is first performed on the traditional region generation network S21:

[0067] The traditional region generation network S21 is trained on the basic samples to generate basic candidate regions, where the candidate boxes corresponding to the basic samples are represented as P (b) =R (b) (f (b) θ (b) ), where each candidate box is represented by the position offset of the associated predefined anchor point, R (b) Represents the traditional region generation network S21. The training set for initially training the traditional region generation network S21 only contains basic samples.

[0068] When training the traditional region generation network S21, the anchor point is determined to be a positive sample or a negative sample based on the intersection-and-union ratio of the anchor point and the real basic sample annotation box. The intersection-and-union ratio of the positive sample anchor point and the annotation box is large and exceeds a certain threshold. Calculate the loss of positive and negative samples and the position offset loss of the anchor point to get L rpn .

[0069] like Figure 1As shown, the region of interest module ROI is set behind the mutation module S2. The ROI is the specific first few layers of the detector O, which is used to process the features of the candidate box area into the same scale. After the basic candidate region obtained by S21 is input into the detector O, the confidence score of each candidate box belonging to each category will be output. When training the detector O, the specific judgment category of the candidate box is obtained according to the intersection and union ratio of the candidate box and the annotation box, and positive samples and negative samples are obtained. The positive sample is a candidate box belonging to the current judgment category, and the negative sample is a candidate box that does not belong to the current judgment category. At this time, the positive sample and negative sample obtained by the basic sample training are respectively represented as P (b)+ and P (b)- The above positive and negative samples are input into the detector O to train the classification loss L of the detector O. cls and the regression loss L reg , the corresponding objective function is:

[0070]

[0071] In formula (1), O (b) Represents the base class detector, L (b) is the overall training loss of the base category, L rpn is the loss of the region generation network RPN, L cls is the classification loss of detector O, L reg is the regression loss of detector O.

[0072] L rpn Including the loss of regional generation network RPN binary classification and regression, L cls The specific form is multi-classification cross entropy loss, L reg The specific form is the regression box smooth L1 loss.

[0073] When training the network, it is necessary to identify both positive samples that belong to the category and negative samples that do not belong to the category. In this way, the network can learn the classification boundaries of different categories, among which negative samples play the role of regular constraints.

[0074] Then, the mutation phase training is performed for the traditional region generation network S21:

[0075] After the basic training, the network parameters θ of the traditional region generation network S21 are obtained (b) , and the network parameters θ (b) Fixed, does not participate in subsequent iterative training, that is, a fixed network parameter θ is obtained (b) . Fix the network parameters θ (b) When the training set contains new samples, the traditional region generation network S21 outputs the basic candidate region, where the candidate box corresponding to the new sample is represented by P(bn) =R (b) (f (n) θ (b) ).

[0076] Then, set the initialization network parameter θ0 of the mutation region generation network S22 (n) is θ (b) . In the network parameter θ0 (n) Based on the new sample x (n) and the corresponding feature representation f (n) Add the training set and input it into the mutation region generation network S22. After t iterations, the network parameter θ is obtained. t (n) , thus, after the training, the network parameters θ of the mutation region generation network S22 are obtained (n) At this time, the mutation region generation network S22 outputs a new candidate region, where the candidate box corresponding to the new sample is represented by P (n) =R (n) (f (n) θ (n) ), the candidate box corresponding to the basic sample is represented as P (nb) =R (n) (f (b) θ (n) ).

[0077] Finally, the merging submodule S23 sorts the candidate boxes in the basic candidate area and the new candidate area according to the confidence of the included category. The confidence of each candidate box in the included category only represents whether the corresponding candidate box contains the category, and does not determine the type of the category. The merging submodule S23 selects the candidate boxes whose confidence of the included category is higher than the preset threshold from the candidate boxes to form the merged candidate area. Will output the corresponding positive sample and negative samples

[0078] The mutation module S2 is a sequentially updated mutation region generation network. Unlike the recombination module where the new category is completely inherited from the base category with a certain randomness, the mutation also takes into account the new features of the new category.

[0079] like Figure 1 As shown, the small sample target detection device further includes:

[0080] The recombination module S1 is used to recombine the distribution of the basic sample and the new sample to generate instantiated individuals;

[0081] The instantiated individual is input as a newly added new sample into the detector O of the small sample target detection device for training.

[0082] In the embodiment of the present invention, the recombination module S1 is only used in the training process of the small sample target detection device. After the training is completed, the recombination module S1 does not work in the actual detection.

[0083] The instantiated individuals and the merged candidate region samples selected in S23 are used as training samples and input into the detector O of the small sample target detection device. Thus, the diversity of training is increased.

[0084] Finally, reorganization training is performed for the reorganization module S1:

[0085] The recombinant module S1 comprises:

[0086] The selection submodule S11 is used to select the best similar basic sample distribution of each new sample distribution according to the similarity between the new sample and the basic sample distribution to form a similar set;

[0087] An updating submodule S12 is used to update the distribution feature information of the corresponding new sample according to the distribution feature information of the basic sample in the similar set, so as to obtain the abstract category information of the corresponding new sample;

[0088] The generation submodule S13 is used to instantiate different individuals according to the abstract category information of each new sample.

[0089] The selection submodule S11 is specifically used to measure the similarity between the new sample distribution and the basic sample distribution through the Wasserstein distance; and select the optimal similar basic sample distribution of each new sample distribution according to the greedy algorithm and the similarity between the new sample distribution and the basic sample distribution to form a similar set.

[0090] The distribution feature information of the basic sample includes: the mean and covariance of the distribution feature of the basic sample; the distribution feature information of the new sample includes: the mean and covariance of the distribution feature of the new sample; the abstract category information includes: the mean and covariance of the abstract category distribution.

[0091] The similarity is calculated by the following formula:

[0092]

[0093] In formula (2), i and j represent the two category indices whose similarity is calculated, μ i and μ j represents the mean of the corresponding category, △ i and j represents the covariance of the corresponding category, and φ represents the Wasserstein distance for calculating the similarity of the distribution of two categories.

[0094] The updating submodule S12 updates the distribution in the following manner:

[0095]

[0096]

[0097] In formulas (3) to (4), Represents the mean of the abstract category distribution corresponding to the new sample after update, Represents the covariance of the abstract category distribution corresponding to the new sample after update, and Represent the mean and covariance of the new sample feature distribution, respectively. and Distribution represents the mean and covariance of the distribution of any underlying sample feature in the similarity set.

[0098] The generating submodule 13 is specifically used to obtain the Gaussian distribution of the abstract category according to the mean and covariance of the abstract category distribution;

[0099] According to the Gaussian distribution, sampling generates different individuals of instantiation of the abstract category.

[0100] The instantiation method is:

[0101]

[0102] In formula (5), N(·) represents Gaussian distribution and P represents the probability of the abstract category.

[0103] The generation submodule 13 is based on the mean The variance is The abstract category of Gaussian distribution is generated according to probability P, and the feature sampling result of the abstract category is generated, that is, different individuals of instantiation are obtained.

[0104] In the tth iteration of the recombination training, the recombination module S1 outputs the synthetic data set and uses its sampled individuals to form the positive sample matrix

[0105] Reorganize the candidate regions output by module S1: positive sample matrix And the candidate regions output by mutation module S2: Positive samples and negative samples Input the detector O and perform parameter training. The corresponding objective function is:

[0106]

[0107] In formula (6), O (n) Represents the detector of the new category, L t (n) is the loss computed by detector O on new categories.

[0108] In the embodiment of the present invention, the instantiated individuals generated at the end of the training process are input as samples of the newly added new category into the detector O of the small sample target detection device for training.

[0109] The recombination module S1 is a non-parametric selective gene recombination module, which allows a new category to accelerate learning by recombining and reusing the visual information inherited from the base categories. In addition, the gene recombination operation introduces randomness when the new combination genes are instantiated, which greatly promotes the diversity of new categories.

[0110] In the basic training phase, the embodiment of the present invention trains the corresponding network with abundant basic samples, which can easily locate suitable candidate regions and determine their categories. In the mutation training phase and the recombination training phase, the network uses sparse new samples as input and attempts to transfer knowledge of basic categories through mutation and recombination.

[0111] In the mutation training phase and the recombination training phase, the recombination module S1 and the mutation module S2 are used to generate a variety of reasonable candidate regions to promote the full training of the detector O. Since all new categories are independent of each other and the basic categories, the recognition process of the detector O can be regarded as isolation guided by the actual annotation information of the sample, which is regarded as the result of natural selection.

[0112] The embodiment of the present invention uses two candidate region generation networks to generate merged candidate regions, and generates positive sample candidate regions based on the recombination and instantiation of basic samples and new samples. Therefore, the temporal species evolution network can simultaneously consider past category experience and the specific situation of the current category.

[0113] The embodiment of the present invention provides a simple and effective small sample target detection framework with evolutionary training. Starting from the perspective of inheritance and mutation in natural evolution, in order to solve the problem that most small sample target detection methods cannot generate enough high-quality positive sample candidate regions due to poor inheritance of basic category knowledge and lack of non-specific modifications of new categories, resulting in unsatisfactory detection performance, a recombination module and a mutation module are proposed to increase the diversity and rationality of positive sample candidate regions respectively. These operations are simple and effective, can be applied to any two-stage detection framework, do not interfere with the original modules, and have excellent performance.

[0114] The following takes image recognition in autonomous driving as an example to illustrate the application of the small sample target detection device provided by the present invention in actual detection:

[0115] The small sample target detection device of the present invention is provided in the autonomous driving vehicle, and the traffic lights and traffic signs of special sections encountered during driving are obtained during driving. When driving on urban roads: the traffic light image obtained by the autonomous driving vehicle is used as the basic sample and input into the small sample target detection device. First, the backbone network E extracts the features of the basic sample, and the obtained basic sample feature information is then simultaneously passed through the traditional region generation network S21 and the mutation region generation network S22. The traditional region generation network S21 and the mutation region generation network S22 output the corresponding basic candidate region and the new candidate region for the feature information of the traffic light image of the ordinary road section. At this time, the confidence that each candidate frame in the basic candidate region contains "traffic light" is high. The merging submodule S23 selects a certain proportion of candidate frames with high confidence, sends them to the region of interest module ROI, and outputs the corresponding positive samples and negative samples. Finally, the detector O determines that the input image is an ordinary traffic light image through classification and regression.

[0116] When the situation is on a special road section such as a mountain road: the image of the special traffic sign obtained by the autonomous driving vehicle is a new sample and is input into the small sample target detection device. First, the backbone network E extracts the features of the new sample to obtain the feature information of the new sample, and then passes through the traditional region generation network S21 and the mutation region generation network S22 at the same time. The traditional region generation network S21 and the mutation region generation network S22 output the corresponding basic candidate region and the new candidate region for the feature information of the traffic sign image of the special road section. At this time, the confidence that each candidate frame in the new candidate area contains the "special road section sign" is high. The merging submodule S23 selects a certain proportion of candidate frames with high confidence, and after being sent to the region of interest module ROI, the positive and negative samples corresponding to each category are output after ROI processing. Finally, the detector O determines that the input image is a special road section traffic sign image after classification and regression.

[0117] Next, the small sample target detection device provided by the embodiment of the present invention is tested on the PASCAL VOC dataset and the COCO dataset respectively, and compared with the existing detection model.

[0118] Among them, the existing detection models include but are not limited to:

[0119] 1. Meta-learning based models, such as FSRW, Meta-RCNN, and MetaDet;

[0120] 2. Models based on fine-tuning, such as FRCN+joint, FRCN+ft, FRCN+ft-full, and TFA.

[0121] Table 1 shows the results of side evaluation of the small sample target detection device provided by the embodiment of the present invention and the existing detection model on the PASCAL VOC dataset, where shot is the number of samples and bold indicates the best performance.

[0122] It can be seen from Table 1 that the small sample target detection device provided by the embodiment of the present invention has detection performance that is much better than the existing detection model when the number of samples is small.

[0123] Table 2 shows the results of side evaluation of the small sample target detection device provided by an embodiment of the present invention and the existing detection model on the COCO dataset.

[0124] It can be seen from Table 2 that, similarly, the small sample target detection device provided by the embodiment of the present invention has detection performance that is much better than the existing detection model when the number of samples is small.

[0125] Table 1

[0126]

[0127] Table 2

[0128]

[0129] like Figure 3 As shown, the embodiment of the present invention also provides a small sample target detection method for time-series species evolution, which at least includes the following steps:

[0130] 101. The mutation module generates a basic candidate region and a new candidate region for the basic sample and the new sample in the small sample, respectively, and then merges them to obtain a merged candidate region;

[0131] 102. The detector identifies basic samples and new samples in small samples based on the merged candidate regions.

[0132] The small sample target detection method provided in the above embodiment is implemented based on the embodiment of the small sample target detection device provided above. For specific implementation steps, please refer to the description in the device embodiment, which will not be repeated here.

[0133] In order to implement the small sample target detection method provided by the present invention, the present invention also provides a small sample target detection system. Figure 4As shown, the small sample target detection system includes a memory 41 and a processor 42, and may further include a communication component, a sensor component, a power component, a multimedia component and an input / output interface according to actual needs. Among them, the memory, communication component, sensor component, power component, multimedia component and input / output interface are all connected to the processor 42. In an embodiment of the present invention, the memory 41 may be a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, etc., and the processor may be a central processing unit (CPU), a graphics processing unit (GPU), a field programmable gate array (FPGA), an application-specific integrated circuit (ASIC), a digital signal processing (DSP) chip, etc. Other communication components, sensor components, power components, multimedia components, etc. can all be implemented using common components in existing smart phones, which will not be described in detail here.

[0134] On the other hand, in the above-mentioned small sample target detection system, the processor 42 reads the computer program in the memory 41 to perform the following operations:

[0135] The mutation module generates basic candidate regions and new candidate regions for the basic samples and new samples in the small samples respectively, and then merges them to obtain merged candidate regions;

[0136] The detector identifies basic samples and new samples in small samples based on the merged candidate regions.

[0137] The above is a detailed description of the small sample target detection method and device for time-series species evolution provided by the present invention. For those skilled in the art, any obvious changes made to it without departing from the essence of the present invention will constitute an infringement of the patent right of the present invention and will bear the corresponding legal liability.

Claims

1. A small sample target detection device for time-series species evolution, characterized by include: A mutation module is used to generate basic candidate regions and new candidate regions for the basic samples and new samples in the small sample, respectively, and then merge them to obtain merged candidate regions, and generate positive sample candidate regions according to the recombination and instantiation of the basic samples and the new samples; wherein the mutation module includes a traditional region generation network for outputting a set of candidate boxes to form a basic candidate region and a mutation region generation network for outputting a set of candidate boxes to form a new candidate region; the parameters of the mutation region generation network are obtained by training the parameters of the traditional region generation network; The detector is used to identify the basic samples and new samples in the small samples based on the merged candidate regions; A recombination module is used to recombine the distribution of basic samples and new samples to generate instantiated individuals; the instantiated individuals are input into the detector as newly added new samples for training; wherein the recombination module includes: a selection submodule, used to select the optimal similar basic sample distribution of each new sample distribution through the similarity between the new sample and the basic sample distribution to form a similar set; an update submodule, used to update the distribution feature information of the corresponding new sample according to the distribution feature information of the basic sample in the similar set, and obtain the abstract category information of the corresponding new sample; a generation submodule, used to instantiate different individuals according to the abstract category information of each new sample.

2. The small sample target detection device as claimed in claim 1, characterized in that The mutation module also includes a merging submodule, which is used to select candidate boxes belonging to the basic sample or the new sample whose confidence is higher than a preset threshold from the basic candidate area and the new candidate area to form a merged candidate area.

3. The small sample target detection device according to claim 2, characterized in that: The traditional region generation network and the mutation region generation network are set in parallel.

4. The small sample target detection device as claimed in claim 3, characterized in that When the small sample target detection device is trained: After the traditional region generation network is fully trained with basic samples, fixed network parameters of the traditional region generation network are obtained; When the mutation region generation network is trained by new samples, the initialization network parameters of the mutation region generation network are equal to the fixed network parameters of the traditional region generation network.

5. The small sample target detection device according to claim 1, characterized in that: The selection submodule measures the similarity between the new sample distribution and the basic sample distribution by the Wasserstein distance, and selects the best similar basic sample distribution of each new sample distribution according to the greedy algorithm and the similarity between the new sample distribution and the basic sample distribution to form a similar set.

6. The small sample target detection device according to claim 1, characterized in that: The distribution characteristic information of the basic samples includes: the mean and covariance of the distribution characteristics of the basic samples; The distribution feature information of the new sample includes: the mean and covariance of the distribution feature of the new sample; The abstract category information of the new sample includes: the mean and covariance of the distribution characteristics of the abstract category.

7. The small sample target detection device according to claim 6, characterized in that: The generation submodule obtains the Gaussian distribution of the abstract category according to the abstract category information of the new sample; and samples and generates different instantiated individuals of the abstract category according to the Gaussian distribution.

8. A small sample target detection method for time-series species evolution, implemented based on the small sample target detection device according to any one of claims 1 to 7, characterized in that The steps include: The mutation module generates basic candidate regions and new candidate regions for the basic samples and new samples in the small samples respectively, and then merges them to obtain merged candidate regions; The detector identifies basic samples and new samples in small samples based on the merged candidate regions.

Images