A compressed imaging recognition method and system for plankton monitoring

Abstract

The present invention discloses a compressed imaging recognition method for plankton monitoring, comprising the following steps: performing parameter optimization using a data set of in-situ plankton images; sequentially loading the optimized measurement mode combinations onto a digital micromirror in a hardware system, optically modulating the image of the current scene to obtain measurement results; and completing the recognition task using a classification network. The present invention can reduce redundant information in images, thereby reducing the amount of sampled data from plankton monitors and improving sampling efficiency. It utilizes image feature learning to quantify the importance of each feature, enabling the learning and selection of a corresponding number of key features to maximize inter-class differences in plankton images. During the dimensionality reduction process for the in-situ plankton images, the intra-class correlation and inter-class differences of the selected results are improved. Furthermore, composite high-level features are gradually extracted from low-level features to improve and ensure the accuracy of plankton detection.

Description

Technical Field

[0001] The present invention relates to the technical field of plankton monitoring, and in particular to a compression imaging recognition method and system for plankton monitoring. Background Art

[0002] With the development of smart ocean engineering, higher requirements are being placed on ocean information collection and transmission technologies. Smart ocean engineering, resulting from the deep integration of industrialization and informatization, is a systematic project for understanding and managing the ocean. Relying on basic marine information technology, it leverages next-generation information technology to deeply integrate the marine environment and human activities, forming a comprehensive ocean information system that enables accurate acquisition of ocean information, data sharing, and innovative applications. Research on ocean information collection technologies is becoming increasingly important, and real-time monitoring of plankton is a key research area. Real-time monitoring of plankton is crucial for building smart ocean engineering, achieving a thorough understanding of ocean information, assessing marine ecosystems, and preventing disasters.

[0003] The underwater optical in-situ plankton imager is mainly composed of an optical imaging lens and an image sensor. Some imagers are also equipped with a light source module to provide lighting for the dark underwater environment and improve the imaging effect. As a monitoring instrument, the imager can effectively image targets at the micrometer and millimeter scale. Even considering the absorption and attenuation of underwater light intensity, the optical imaging method can work normally. The image acquisition capability and data processing efficiency of the monitor directly affect its actual use effect. Among them, there are the following key problems: (1) Most plankton monitors need to work for a long time or even all day. When the image quality obtained at the front end is high, more data will be generated, which increases the transmission bandwidth demand and processing time. There is a problem that a large amount of data is difficult to store and transmit; when the image quality obtained at the front end is low, it is not conducive to the classification and recognition of the back end, which reduces the accuracy of monitoring and recognition. In addition to the problem of large data volume, the existing monitors also have the following problems: (2) During the sampling process, due to the effects of light absorption, scattering and turbulence, the sampled data contains a lot of noise interference, with little available information and low value density. (3) For highly complex and fragmented sampling results, there is a lack of reasonable data processing methods, making it impossible to achieve a high accuracy rate in plankton identification. Summary of the Invention

[0004] The purpose of the present invention is to solve the technical problem of low accuracy in plankton monitoring and recognition, and to propose a compressed imaging recognition method and system for plankton monitoring.

[0005] The technical problem of the present invention is solved by the following technical solutions:

[0006] A compressed imaging recognition method for plankton monitoring comprises the following steps:

[0007] S1, parameter optimization using a plankton in situ image dataset;

[0008] S2. Loading the optimized optimal measurement mode combination onto the digital micro-mirror of the hardware system in sequence, optically modulating the image of the current scene, and obtaining a measurement result;

[0009] S3, complete the recognition task through the classification network;

[0010] Wherein, step S1 includes:

[0011] S11. Using the discrete cosine transform basis, calculate and generate the orthogonal deterministic measurement matrix to improve the information entropy aggregation characteristics;

[0012] S12, through the adaptive feature selection method, the most critical compression features are sampled within the limited data to maximize the difference between plankton image classes;

[0013] In S12, in the adaptive feature selection method, image feature learning is used to set the cutoff parameters during statistical calculation, quantify the importance of each feature, learn and select a corresponding number of key features, and optimize to obtain the optimal measurement mode combination.

[0014] In some embodiments, the following technical features are also included:

[0015] Step S1 also includes: establishing a probability model and optimizing parameters in the probability model, and selecting a measurement matrix through probability distribution.

[0016] The probability model is a Bernoulli probability model. The method for establishing the Bernoulli probability model includes: associating features with importance scales by generating a feature sampling probability vector X.

[0017] Further,

[0018] The Bernoulli probability model establishment method further includes: assuming that the probability of the i-th frequency feature being selected is Pr[X i =1]=p i , according to p i Determine the random variable X i The value of X i There are only two states, selected and unselected, which are represented by {X i =1,X i =0}, and finally X is the sum of all random variables X i The set of probability p i The value of is determined by Gumbel-Softmax sampling; a vector [η i1 ,ηi2 ], by comparing η i1 and η i2 The value of X is determined i The value of is as follows:

[0019]

[0020] Then use the Softmax operation to adjust η i1 and η i2 Normalize it so that it can represent probability, and then use the Gumbel sampling method to make the sampling differentiable and closest to the true distribution. Finally, the complete Gumbel-Softmax process is shown in the following formula:

[0021] p i =softmax{[ln(η i1 )+G i ] / temperature}

[0022] where η i1 is a randomly given initial value, G i It is the sampling result of independent and identical distribution and obeys Gumbel distribution, and temperature is a hyperparameter greater than 0.

[0023] In some embodiments, the following technical features are also included:

[0024] The parameters that need to be optimized in the probability model include: the probability value p that characterizes a single compression feature i and the actual number of features selected l(x); both are iteratively optimized through continuous chain derivative backpropagation.

[0025] The optimized calculation formula in the probability model is as follows:

[0026]

[0027] in, and are the probability parameters p in the jth and j+1th rounds respectively. i The value of α is A positive learning rate, To solve the partial derivative operation, L cla is the loss function generated by classification error;

[0028]

[0029] Among them, num f is the number of frequency features that should be sampled at the current sampling rate, l(x) is the number of 1s in X, and is also the number of frequency features actually selected, lj (x) and l j+1 (x) are the values ​​of the number of features l(x) selected in the jth and j+1th rounds, respectively.

[0030] The classification network is a benchmark network model established based on Swin Transformer, that is, a recognition model; step S3 specifically includes: inputting the measurement results into the optimized recognition model to obtain the final classification result.

[0031] The optimization of the recognition model specifically includes:

[0032] The recognition model is optimized using a deep learning algorithm to minimize the error loss function generated by the classification model and the feature selection model. The loss function is shown in the following formula:

[0033]

[0034] Among them, M θ represents the recognition model, θ represents the model parameters, X represents the feature sampling probability vector, S represents the input after source decoding, and c = M θ (S), where c is the predicted category of the output, argmin represents the variable value that minimizes the objective function f(x), is the loss function generated by the classification model, is the cross entropy loss function used to calculate the classification loss; c is the predicted category, l is the true category label; L num is the error caused by selecting the number of features; λ is a hyperparameter used to balance classification error and feature selection error;

[0035]

[0036] Here, ln is the natural logarithm.

[0037] The present invention also adopts the following technical solutions:

[0038] A compressed imaging recognition system for plankton monitoring comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the aforementioned method is implemented.

[0039] A computer-readable storage medium stores a computer program, wherein the computer program implements the aforementioned method when executed by a processor.

[0040] The beneficial effects of the present invention compared with the prior art include:

[0041] The compressed imaging and recognition method for plankton monitoring proposed in the present invention uses a discrete cosine transform basis as a measurement mode to design a deterministic compressed imaging method, which can reduce redundant information in the image and obtain compressed data directly in the sampling stage, thereby reducing the amount of sampling data of the plankton monitor and improving sampling efficiency; based on the above-mentioned compressed imaging method, the present invention proposes an adaptive compressed feature selection method, which utilizes image feature learning, sets cutoff parameters during statistical calculations, quantifies the importance of each feature, can learn and select a corresponding number of key features, and maximize the inter-class differences of plankton images; in the process of dimensionality reduction of in-situ plankton images, this method improves the intra-class correlation of the selection results and the inter-class differences of the selection results, thereby effectively improving the accuracy of plankton monitoring and recognition.

[0042] In addition, in some embodiments, the following beneficial effects are also achieved:

[0043] The present invention provides a recognition algorithm suitable for compressed imaging data, which gradually extracts composite high-level features from low-level features, thereby improving the recognition accuracy, thereby improving and ensuring the accuracy of plankton detection.

[0044] Other beneficial effects of the embodiments of the present invention will be further described below. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a flow chart of a compressed imaging recognition method for plankton monitoring in an embodiment of the present invention;

[0046] Figure 2 This is a schematic diagram of the overall framework of a compressed imaging recognition system according to an embodiment of the present invention;

[0047] Figure 3 is a schematic diagram of a microscopic imaging system according to an embodiment of the present invention;

[0048] Figure 4 3 is a schematic diagram of the recognition accuracy of the plankton in situ image dataset under different signal-to-noise ratios in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] Glossary:

[0050] Measurement matrix: The measurement matrix is ​​a matrix used to reduce the dimensionality of measurement data in compressed sensing. The Hadamard inner product between the measurement data and the measurement matrix is ​​performed to achieve the purpose of dimensionality reduction. In the hardware system, the measurement matrix is ​​binarized and loaded into the DMD to control the flipping shape of the DMD array. The meaning of an orthogonal deterministic measurement matrix is ​​that the matrix is ​​mutually orthogonal and each element is deterministic.

[0051] The most critical compression feature refers to the characteristics of the imaging target. In some embodiments of the present invention, this mainly refers to the lines, contours and other features of plankton. Frequency domain feature information is also used for quantification during measurement.

[0052] Image inter-class differences: refers to the categories of plankton. In some embodiments of the present invention, plankton are divided into 12 categories during classification.

[0053] Cutoff parameter: refers to the number of feature points required for measurement, which is directly related to the sampling rate and also indicates the number of measurement modes selected.

[0054] Key features: refers to the spectral distribution selected when extracting image features during measurement (that is, transforming the image into the frequency domain and then selecting several points in the frequency domain for measurement). It is essentially a frequency domain feature.

[0055] Identification model: The identification model is a network model for classifying plankton.

[0056] Bernoulli probability model: The Bernoulli probability model is the probability distribution when selecting the measurement matrix.

[0057] Feature selection model: The feature selection model is a combination of measurement matrices used in compression imaging, loaded in the DMD hardware system, and used for image measurement.

[0058] Before introducing the specific embodiments of the present invention, the idea of ​​the present invention is first introduced as follows:

[0059] The main idea of ​​the present invention is to use the discrete cosine transform basis to calculate and generate an orthogonal deterministic measurement matrix, improve the information entropy aggregation characteristics, and through adaptive feature selection technology, sample the most critical compression features within limited data to maximize the differences between plankton image classes, and finally complete the recognition task through a classification network.

[0060] The present invention will be further described below with reference to the accompanying drawings and in combination with preferred embodiments. It should be noted that, unless there is a conflict, the embodiments and features in the embodiments of the present application can be combined with each other.

[0061] In addition, in the description of the embodiments of the present invention, “a plurality of” means two or more than two, unless otherwise clearly and specifically defined.

[0062] Example

[0063] The embodiment of the present invention provides a compressed imaging recognition method for a plankton monitor, the process of which is as follows: Figure 1 As shown, the following steps are included:

[0064] S1, parameter optimization using a plankton in situ image dataset;

[0065] S2. Loading the optimized optimal measurement mode combination onto the digital micro-mirror of the hardware system in sequence, optically modulating the image of the current scene, and obtaining a measurement result;

[0066] S3. Complete the recognition task through the classification network.

[0067] The principles of the method of the embodiment of the present invention are described in detail below: (1) Based on the theory of compressed sensing, a deterministic compressed imaging method is designed using the discrete cosine transform basis as the measurement mode to reduce redundant information in the image and obtain compressed data directly in the sampling stage, so as to reduce the amount of sampled data of the plankton monitor and improve the sampling efficiency. The discrete cosine transform has the characteristic of aggregating the image energy in the spectrum. The result of each measurement is the spectrum coefficient of the image, and each basis has specific frequency information. Therefore, it is only necessary to measure these coefficients in a targeted manner to complete the compressed imaging reconstruction, which increases the prior information of the compressed measurement and improves the imaging quality and speed. (2) Based on the above-mentioned compressed imaging method, an adaptive compressed feature selection method is proposed. This method utilizes image feature learning, sets the cutoff parameter during statistical calculation, quantifies the importance of each feature, and can learn and select a corresponding number of key features, thereby maximizing the inter-class difference of the plankton image. In the process of dimensionality reduction of the in-situ plankton image, this method improves the intra-class correlation of the selection results and improves the inter-class difference of the selection results. (3) Based on the above-mentioned compression imaging method and adaptive compression feature selection method, a recognition algorithm suitable for compressed imaging data was designed. Composite high-level features were gradually extracted from low-level features, and the recognition accuracy was improved to improve and ensure the accuracy of plankton detection.

[0068] The embodiment of the present invention also provides a compression imaging recognition system, the overall framework of the system is as follows Figure 2As shown in the figure, the overall framework is mainly divided into three workflows: back-end optimization, front-end measurement, and back-end processing: (1) First, the in-situ image dataset of plankton is used for parameter optimization, including adaptive selection of compressed features and optimization of recognition model parameters. The optimization results include two parts, namely the recognition model that can process compressed measurement data and the optimal measurement mode combination. These measurement modes correspond one-to-one to the frequency features in the DCT (Discrete Cosine Transform) basis and are orthogonal to each other, which can maximize the difference between classes. In the optimization results, different information clusters are sampled with higher priority, while reducing the amount of raw data, capturing effective feature information and filtering noise data, which can solve the problem of ill-conditioned solution of raw information in the measurement data. (2) Secondly, the optimal measurement mode combination obtained by optimization is sequentially loaded onto the DMD (Digital Micromirror Device) of the hardware system to optically modulate the image of the current scene. The hardware system constitutes a passive (image plane conjugate) single-pixel camera module, which realizes online compressed sampling. In the compressed sampling framework, the central control system sends synchronous timing control signals to control the operation of the DMD and single-photon detector to obtain the measurement results. (3) Finally, the modulation results measured by the hardware system are directly input into the optimized recognition model to obtain the final classification results.

[0069] In some embodiments, step S1 further includes: establishing a probability model and optimizing the parameters in the probability model, and selecting a measurement matrix through probability distribution. The measurement pattern distribution is arranged from low frequency to high frequency, forming a feature sampling probability vector X. The in situ plankton dataset is then input into the network for parameter optimization, and the optimization target is the probability (importance) of each frequency feature being selected. Each back propagation will update the probability of the i-th frequency feature being selected until the frequency features with the highest probability values ​​are finally obtained, which are the selected ones (each frequency feature has a corresponding measurement matrix one-to-one correspondence).

[0070] The following is a detailed description of the parameter optimization method design of the embodiment of the present invention:

[0071] The optimization process is carried out on the back end and involves the optimization of two models, namely the Bernoulli probability model and the plankton recognition model designed based on the SwinTransformer framework.

[0072] The embodiment of the present invention is used to perform compression imaging of plankton. In the image, low-frequency features represent the overall contour information of the image, and high-frequency features represent the edge texture information of the image. According to the target characteristics of plankton, it is necessary to obtain both the overall contour of the plankton and its edge texture features. Therefore, it is necessary to find the "optimal frequency combination" during compression measurement, that is, the frequency features unique to each plankton. In order to quantify the importance of each frequency feature, a Bernoulli probability model is established. Specifically, a feature sampling probability vector X is generated to associate the feature with the importance scale. Assume that the probability of the i-th frequency feature being selected is Pr[X i =1]=p i , according to p i Determine the random variable X i The value of X i There are only two states, namely {selected, not selected}, which are represented by {X i =1,X i =0}, and finally X is the sum of all random variables X i However, in order to continuously perform chain derivation during the back propagation process of optimizing the probability parameters, the probability p i The value of is determined by Gumbel-Softmax (Gumbel, Gumbel distribution; Softmax, a function) sampling, which is as follows: a vector [η is assigned to each frequency feature i1 ,η i2 ], by comparing η i1 and η i2 The value of X is determined i The value of is as shown in the formula:

[0073]

[0074] Then use the Softmax operation to adjust η i1 and η i2 Normalize it so that it can represent probability, and then use the Gumbel sampling method to make the sampling differentiable and closest to the true distribution. Finally, the complete Gumbel-Softmax process is shown in the formula:

[0075] p i =softmax{[ln(η i1 )+G i ] / temperature} (2)

[0076] where η i1 is a randomly given initial value, G iIt is the sampling result of independent and identical distribution and obeys Gumbel distribution. Temperature is a hyperparameter greater than 0. The larger the value of temperature, the higher the p i The distribution of the results is smoother; the smaller the value of temperature, the better the p i The distribution of the results is closer to the discrete one-hot code. i1 is a trainable parameter, so during the optimization process, if the i-th feature contains valid information, then η i1 Will move towards i Converge in the direction of maximum possible, otherwise, η i1 The convergence direction of p i As small as possible.

[0077] The Gumbel-Softmax process is equivalent to a reparameterization process, using a standard variable to represent a complex variable. Therefore, this probability model can be used in the subsequent compressed feature adaptive selection algorithm to select the most critical plankton in situ image feature information in the measurement process.

[0078] After the sampling method is determined, the parameters in the Bernoulli probability model are optimized. The parameters that need to be optimized in the model are divided into two categories: the probability value p that represents a single compression feature i and the actual number of features selected l(x). Both are iteratively optimized through continuous chain derivation and back propagation. As shown in the formula and formula:

[0079]

[0080] in, and are the probability parameters p in the jth and j+1th rounds respectively. i The value of α is A positive learning rate, To solve the partial derivative operation, L cla is the loss function resulting from the classification error.

[0081]

[0082] Among them, num f is the number of frequency features that should be sampled at the current sampling rate, l(x) is the number of "1" in X, and is also the number of frequency features actually selected, l j (x) and l j+1 (x) are the values ​​of the number of features l(x) selected in the jth and j+1th rounds, respectively.

[0083] The parameters of the recognition model were then optimized, and an effective network model was selected based on the characteristics of the in-situ plankton images. This model was then improved to be suitable for compressed measurement data. In the recognition of plankton, for categories with large differences in overall contours, global feature extraction can achieve relatively good resolution, which also corresponds to the low-frequency portion of the compressed measurement. For categories with relatively similar overall contours, however, they can only be distinguished by comparing the texture information of local features, which also corresponds to the high-frequency portion of the compressed measurement. Therefore, for the recognition of plankton, the selected baseline network needs to take into account the ability to extract both local and global features. The "window offset" operation in the Swin Transformer can transfer information between different windows, extracting both local and global features, which meets the recognition requirements of in-situ plankton images.

[0084] Figure 2 The recognition model M shown θ It is designed based on the Swin Transformer framework, where θ represents the model parameters and S represents the input after source decoding. Since the output in the classification task can be described as the inference category, that is, c = M θ (S), where c is the predicted category of the output. The model is optimized using a deep learning algorithm to minimize the error loss function generated by the classification model and the feature selection model. The loss function is shown in the formula:

[0085]

[0086] Where n is the number of feature patterns to be selected, l(X) is the number of feature patterns actually selected, l(X) is the value of l(x) after several iterations, and argmin represents the variable value when the objective function f(x) takes the minimum value. is the loss function generated by the classification model, is the cross entropy loss function used to calculate the classification loss, as shown in the formula; c is the predicted category, l is the actual category label; L num It is the error caused by the number of selected features; λ is a hyperparameter used to balance the classification error and feature selection error. The larger the value of λ, the more the model focuses on optimizing the number of selected features and the accuracy of plankton image acquisition, which directly affects the setting of the undersampling rate; the smaller the value of λ, the more the model focuses on optimizing the classification model parameters.

[0087]

[0088] Where ln is the natural logarithm.

[0089] Through the Bernoulli probability model, several measurement matrices with the largest selection probability are obtained to form a feature selection model.

[0090] The embodiment of the present invention also provides a compressed sampling system, according to the framework of the passive (image plane conjugate) single pixel camera used, according to the overall design scheme, the schematic diagram is as shown in FIG. Figure 3 As shown, the positional relationship of each component is fixed to ensure good measurement results. In this diagram, the microscope lens is used to image the target. Next, the focusing system, consisting of condenser lenses A and B, is located in front of the APD (Avalanche Photo Diode) detector. This system ensures that all modulated light intensity is collected by the detector. The image plane conjugate single-pixel camera framework provided in this embodiment of the present invention optically modulates the target image, achieving real-time online compressed sampling.

[0091] Verification Example

[0092] 1. Plankton in-situ image simulation results

[0093] The simulation accuracy of different models for the plankton in situ image dataset at different sampling rates is shown in Table 1, which shows the accuracy results of the plankton in situ image dataset at different sampling rates. Three models are compared: the full sampling model of Swin Transformer, the "low-frequency feature model" that prioritizes sampling low-frequency features, and the "feature selection model" proposed in the embodiment of the present invention. The simulation accuracy is rounded to two decimal places.

[0094] Table 1

[0095]

[0096] It can be seen that the "feature selection model" proposed in the embodiment of the present invention has the best overall performance compared to the "low-frequency feature model". Even under the extreme compression sampling rate of 0.01, the recognition accuracy is still 88.58%; and when compared with the full sampling model, the 0.25 sampling rate reduces the amount of data by 75% compared with the Swin Transformer full sampling baseline model, so the accuracy drop of 1.06% is acceptable.

[0097] 2. Noise image simulation results

[0098] like Figure 4 The following table shows the recognition accuracy of the plankton in situ image dataset under different signal-to-noise ratios. The accuracy data is rounded to two decimal places. The horizontal axis is the signal-to-noise ratio (dB) and the vertical axis is the accuracy (%).

[0099] from Figure 4As can be seen from the figure, when the signal-to-noise ratio is 30 dB, the image is covered by a large amount of noise, but the recognition accuracy reaches 91.51%. Even in the extreme case of a signal-to-noise ratio of 20 dB, the accuracy is still 88.83%. Therefore, the "feature selection model" provided by the embodiment of the present invention has sufficient noise resistance and can be used in actual measurement scenarios.

[0100] 3. Deformation image simulation results

[0101] The accuracy of the plankton in-situ image deformation simulation is shown in Table 2, which shows the recognition accuracy of the plankton deformation image. The accuracy data is rounded to two decimal places.

[0102] Table 2

[0103]

[0104] By comparing the accuracy rates in Table 2 with those in Table 1, we can see that the incompleteness of plankton targets does have a negative impact on identification. The accuracy rates of all models have declined to some extent, but the performance of the "feature selection model" is still the best overall compared to the "low-frequency feature model", indicating that low-frequency features are not the most suitable features for plankton identification. In addition, at a sampling rate of 0.25, the accuracy of the "feature selection model" is 91.52%, which meets the expected target.

[0105] After experimental verification, the method of the embodiment of the present invention can achieve the following indicators when the compression sampling rate is 0.25 and the data volume is reduced by 75%: (1) The recognition accuracy of the in-situ plankton image with a pixel resolution of 224×224 can reach 92.04%. (2) The recognition accuracy of the in-situ plankton image with a signal-to-noise ratio of 30dB can reach 91.51%. (3) The recognition accuracy of the in-situ plankton image with deformed and incomplete targets can reach 91.52%. (4) When the compression sampling rate is adjusted to 0.007, the recognition accuracy of the plankton image in the actual desktop experiment can reach 83.33%.

[0106] Application Scenario

[0107] 1. It can be used for real-time monitoring of plankton in marine areas and prevent marine ecological disasters. In marine applications, the method proposed in this embodiment can be used to capture in-situ images of plankton on the seabed, providing a raw image dataset for subsequent scientific research on plankton. The method proposed in this embodiment can be applied to the collection of marine information, enabling accurate acquisition of marine information and building smart ocean projects that deeply integrate industrialization and informatization.

[0108] 2. Detect the types and numbers of plankton in natural waters and assess the ecosystem of the waters. In water monitoring applications, the methods proposed in this embodiment of the present invention can also be used in detection instruments, reducing the amount of data generated by the monitor while ensuring accuracy and improving detection performance. The methods proposed in this embodiment of the present invention can be applied to plankton monitors to enable real-time monitoring of plankton populations and numbers in waters, helping to prevent ecological disasters in waters.

[0109] 3. Scenarios requiring real-time big data monitoring, such as forest fire prevention, monitoring industries, etc. The method proposed in the embodiment of the present invention can reduce the amount of acquired data while ensuring target recognition accuracy and reducing the pressure on data storage and transmission bandwidth.

[0110] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0111] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0112] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0113] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0114] The above description further details the present invention in conjunction with specific / preferred embodiments, and the specific implementation of the present invention should not be construed as being limited to these descriptions. Persons skilled in the art will appreciate that, without departing from the spirit of the present invention, they may make various substitutions or modifications to the described embodiments, and such substitutions or modifications should be considered within the scope of the present invention. Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "preferred embodiments," "examples," "specific examples," or "some examples" indicates that the specific features, structures, materials, or characteristics described in conjunction with such embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of such terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Persons skilled in the art may combine and assemble the different embodiments or examples described in this specification, as well as features from different embodiments or examples, without conflicting opinions. Although the embodiments of the present invention and their advantages have been described in detail, it should be understood that various changes, substitutions, and modifications may be made herein without departing from the scope of the present invention.

Claims

1. A compressed imaging recognition method for plankton monitoring, characterized in that: The steps include: S1, parameter optimization using a plankton in situ image dataset; S2. Loading the optimized optimal measurement mode combination onto the digital micro-mirror of the hardware system in sequence, optically modulating the image of the current scene, and obtaining a measurement result; S3, complete the recognition task through the classification network; Wherein, step S1 includes: S11. Using the discrete cosine transform basis, calculate and generate the orthogonal deterministic measurement matrix to improve the information entropy aggregation characteristics; S12, through the adaptive feature selection method, the most critical compression features are sampled within the limited data to maximize the difference between plankton image classes; In S12, in the adaptive feature selection method, image feature learning is used to set the cutoff parameters in statistical calculation, quantify the importance of each feature, learn and select the corresponding number of key features, and optimize the optimal measurement mode combination; Step S1 also includes: establishing a probability model and optimizing the parameters in the probability model, and selecting a measurement matrix through probability distribution; the probability model is a Bernoulli probability model, and the method for establishing the Bernoulli probability model includes: generating a feature sampling probability vector , associating features with importance scales; the method for establishing the Bernoulli probability model also includes: assuming that The probability of a frequency feature being selected is ,according to Determine the random variable The value of There are only two states, selected and unselected, which are represented by { },final are all random variables The set of probability The value of is determined by Gumbel-Softmax sampling; a vector is assigned to each frequency feature , by comparison and The value of The value of is as follows: Then use Softmax operation to and Normalize it so that it can represent probability, and then use the Gumbel sampling method to make the sampling differentiable and closest to the true distribution. Finally, the complete Gumbel-Softmax process is shown in the following formula (2.2): in is a randomly given initial value, is the sampling result of independent and identical distribution and obeys Gumbel distribution, is a hyperparameter greater than 0.

2. The compressed imaging recognition method for plankton monitoring according to claim 1, wherein: The parameters that need to be optimized in the probability model include: the probability value that characterizes a single compression feature and the actual number of features selected ; Both are iteratively optimized through continuous chain derivation backpropagation.

3. The compressed imaging recognition method for plankton monitoring according to claim 2, wherein: The optimized calculation formula in the probability model is as follows: in, and They are Wheel and Probability parameters during the round The value of Is a A positive learning rate, To solve the partial derivative operation, is the loss function generated by classification error; in, is the number of frequency features that should be sampled at the current sampling rate, yes The number of 1s in is also the number of frequency features actually selected. and They are Wheel and The number of features selected during the round The value of .

4. The compressed imaging recognition method for plankton monitoring according to claim 1, wherein: The classification network is a benchmark network model established based on Swin Transformer, that is, a recognition model; step S3 specifically includes: inputting the measurement results into the optimized recognition model to obtain the final classification result.

5. The compressed imaging recognition method for plankton monitoring according to claim 4, characterized in that: The optimization of the recognition model specifically includes: The recognition model is optimized using a deep learning algorithm to minimize the error loss function generated by the classification model and the feature selection model. The loss function is shown in the following formula (2.5): in, represents the recognition model, represents the model parameters, represents the feature sampling probability vector, represents the input after source decoding, is the number of characteristic patterns to be selected, is the number of feature patterns actually selected, ,in is the predicted category of the output, argmin represents the variable value when the objective function f(x) takes the minimum value, is the loss function generated by the classification model, is the cross entropy loss function used to calculate the classification loss; is the predicted category, is the true category label; It is the error caused by the number of selected features; It is a hyperparameter used to balance classification error and feature selection error; Here, ln is the natural logarithm.

6. A compressed imaging recognition system for plankton monitoring, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method according to any one of claims 1 to 5 when executing the computer program.

7. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images