Phytoplankton detection method, system and electronic device

Abstract

The application discloses a kind of plankton detection method, system and electronic equipment, it is related to computer vision technical field.The present application can be based on the image data of different regions, different river reaches of the river to be detected by using river plankton detection model, accurately, quickly get the plankton detection result of each pixel.And, multiple domain self-adaptive feature extraction modules are arranged in the river plankton detection model of the application, the distinguishing ability of target and background difference in image data can be improved, multiple boundary alignment modules can enhance the target feature extraction capability, the accuracy of plankton detection can be significantly improved, and the economic and ecological cost of plankton management can be effectively reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, in particular to a method and system for detecting planktonic algae and an electronic device. BACKGROUND

[0002] In recent years, with the acceleration of urbanization, a large amount of industrial, agricultural and living waste is discharged into rivers, leading to serious water eutrophication, and a large number of planktonic algae reproduce in the river, consume oxygen in the water and squeeze the living space of other aquatic plants and animals, eventually forming black and smelly water bodies, and the ecological environment is damaged.

[0003] The detection of planktonic algae can effectively prevent and warn the occurrence of the phenomenon. The existing method for detecting planktonic algae in the river uses video monitoring on the data source, which has high equipment cost and is difficult to popularize and arrange in a large number of small rivers. Traditional methods such as frame difference method and background difference method are used for image interpretation, which has low detection accuracy and slow speed, and cannot be applied to large-scale detection in cities. SUMMARY

[0004] To solve the above problems existing in the prior art, the present application provides a method and system for detecting planktonic algae and an electronic device.

[0005] To achieve the above purpose, the present application provides the following solutions:

[0006] A method for detecting planktonic algae, comprising:

[0007] Obtaining image data of different regions and different river sections of a river to be detected, and preprocessing the image data to obtain target image data;

[0008] Constructing a river planktonic algae detection model; the river planktonic algae detection model comprises a plurality of domain adaptive feature extraction modules, a plurality of boundary alignment modules, a plurality of deconvolution modules, a plurality of feature accumulation modules and an activation function layer;

[0009] Inputting the target image data into the river planktonic algae detection model to obtain a pixel-by-pixel planktonic algae detection result.

[0010] Optionally, the target image data is input into the river planktonic algae detection model to obtain a pixel-by-pixel planktonic algae detection result, specifically comprising:

[0011] The target image data is input into a first domain adaptive feature extraction module, the features output by the first domain adaptive feature extraction module are input into a second domain adaptive feature extraction module and a first boundary alignment module respectively, the features output by the second domain adaptive feature extraction module are input into a third domain adaptive feature extraction module and a second boundary alignment module respectively, the features output by the third domain adaptive feature extraction module are input into a fourth domain adaptive feature extraction module and a third boundary alignment module respectively, the features output by the fourth domain adaptive feature extraction module are input into a fifth domain adaptive feature extraction module and a fourth boundary alignment module respectively, the features output by the fifth domain adaptive feature extraction module are input into a fifth boundary alignment module, the features output by the fifth boundary alignment module are input into a fifth deconvolution module, the features output by the fifth deconvolution module and the features output by the fourth boundary alignment module are both input into a fourth feature accumulation module, the features output by the fourth feature accumulation module are input into a ninth boundary alignment module, the features output by the ninth boundary alignment module are input into a fourth deconvolution module, the features output by the fourth deconvolution module and the features output by the third boundary alignment module are both input into a third feature accumulation module, the features output by the third feature accumulation module are input into an eighth boundary alignment module, the features output by the eighth boundary alignment module are input into a third deconvolution module, the features output by the third deconvolution module and the features output by the second boundary alignment module are both input into a second feature accumulation module, the features output by the second feature accumulation module are input into a seventh boundary alignment module, the features output by the seventh boundary alignment module are input into a second deconvolution module, the features output by the second deconvolution module and the features output by the first boundary alignment module are both input into a first feature accumulation module, the features output by the first feature accumulation module are input into a sixth boundary alignment module, the features output by the sixth boundary alignment module are input into a first deconvolution module, the features output by the first deconvolution module are input into an activation function layer, and the data output by the activation function layer is taken as a phytoplankton detection result.

[0012] Optionally, the domain adaptive feature extraction module comprises, in sequence, a first feature extraction unit, a second feature extraction unit, a third feature extraction unit, a feature accumulation unit and a feature processing unit.

[0013] The feature accumulation unit is configured to fuse the features input into the first feature extraction unit and the features output by the third feature extraction unit.

[0014] Optionally, the first feature extraction unit comprises, in sequence, a 1×1 convolution layer, a BN layer and a GELU activation function layer; the second feature extraction unit comprises, in sequence, a 3×3 convolution layer, a BN layer and a GELU activation function layer; the third feature extraction unit comprises, in sequence, a 1×1 convolution layer and a BN layer; and the feature processing unit comprises, in sequence, an IN layer and a GELU activation function layer.

[0015] Optionally, the channel number of the 1x1 convolution layer in the first feature extraction unit and the channel number of the 3x3 convolution layer in the second feature extraction unit are both 96; and the channel number of the 1x1 convolution layer in the third feature extraction unit is 32.

[0016] Optionally, the boundary alignment module comprises neighborhood correction of the feature points and feature value correction.

[0017] Optionally, the river plankton algae detection model is constructed by:

[0018] obtaining a deep learning training sample set;

[0019] constructing an initial detection model;

[0020] training the initial detection model by using the deep learning training sample set to obtain a trained model; and taking the trained model as the river plankton algae detection model.

[0021] Optionally, a device NVIDIA A40 is used to implement the training of the initial detection model on a python-based deep learning framework Pytorch 1.11.

[0022] According to the specific embodiments of the present application, the following technical effects are disclosed:

[0023] The plankton algae detection method provided by the present application can obtain pixel-by-pixel plankton algae detection results accurately and quickly based on image data of different regions and different river sections of a river to be detected by using the river plankton algae detection model. In addition, the present application sets multiple domain adaptive feature extraction modules in the river plankton algae detection model, which can improve the ability to distinguish the differences between the target and the background in the image data, and sets multiple boundary alignment modules, which can enhance the target feature extraction capability and significantly improve the accuracy of plankton algae detection, thereby effectively reducing the economic and ecological cost of plankton algae treatment.

[0024] In addition, the present application also provides the following two implementation structures:

[0025] A plankton algae detection system applied to the plankton algae detection method provided above, wherein the system comprises:

[0026] a data acquisition module configured to acquire image data of different regions and different river sections of a river to be detected, and to pre-process the image data to obtain target image data;

[0027] A model construction module is configured to construct a river plankton algae detection model, wherein the river plankton algae detection model comprises a plurality of domain adaptive feature extraction modules, a plurality of boundary alignment modules, a plurality of deconvolution modules, a plurality of feature accumulation modules, and an activation function layer.

[0028] A result detection module is configured to input the target image data into the river plankton algae detection model to obtain a pixel-by-pixel plankton algae detection result.

[0029] An electronic device comprises:

[0030] A memory is configured to store a computer program.

[0031] A processor is connected to the memory and is configured to call and execute the computer program to implement the plankton algae detection method provided above.

[0032] The technical effects achieved by the two implementation structures provided by the present application are the same as the technical effects achieved by the plankton algae detection method provided above, and therefore will not be described here. BRIEF DESCRIPTION OF DRAWINGS

[0033] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0034] Figure 1 A flowchart of the plankton algae detection method provided by the present application;

[0035] Figure 2 A structure schematic diagram of the river plankton algae detection model provided by the present application;

[0036] Figure 3 A structure schematic diagram of the domain adaptive feature extraction module provided by the present application;

[0037] Figure 4 A structure schematic diagram of the boundary alignment module provided by the present application;

[0038] Figure 5 A comparison diagram of plankton algae detection results of different detection methods in region one provided by the present application, wherein, Figure 5 (a) of the comparison diagram is an image diagram of region one, Figure 5 (b) of the comparison diagram is a label schematic diagram of the image diagram of region one, Figure 5 (c) of the comparison diagram is a plankton algae detection result diagram of region one using DeeplabV3+, Figure 5Fig. 4 is a detection result diagram of the phytoplankton in region one by using DeeplabV3+, Figure 5 Fig. 5 is a detection result diagram of the phytoplankton in region one by using ERFNet, Figure 5 Fig. 6 is a detection result diagram of the phytoplankton in region one by using GADM;

[0039] Figure 6 Fig. 7 is a comparison diagram of the detection results of the phytoplankton in region two by using different detection methods provided by the present application; wherein, Figure 6 Fig. 7(a) is an image diagram of region two, Figure 6 Fig. 7(b) is a mark indication diagram of the image diagram of region two, Figure 6 Fig. 7(c) is a detection result diagram of the phytoplankton in region two by using DeeplabV3+, Figure 6 Fig. 7(d) is a detection result diagram of the phytoplankton in region two by using NestedUNet, Figure 6 Fig. 7(e) is a detection result diagram of the phytoplankton in region two by using ERFNet, Figure 6 Fig. 7(f) is a detection result diagram of the phytoplankton in region two by using GADM;

[0040] Figure 7 Fig. 8 is a comparison diagram of the detection results of the phytoplankton in region three by using different detection methods provided by the present application; wherein, Figure 7 Fig. 8(a) is an image diagram of region three, Figure 7 Fig. 8(b) is a mark indication diagram of the image diagram of region three, Figure 7 Fig. 8(c) is a detection result diagram of the phytoplankton in region three by using DeeplabV3+, Figure 7 Fig. 8(d) is a detection result diagram of the phytoplankton in region three by using NestedUNet, Figure 7 Fig. 8(e) is a detection result diagram of the phytoplankton in region three by using ERFNet, Figure 7 Fig. 8(f) is a detection result diagram of the phytoplankton in region three by using GADM. DETAILED DESCRIPTION

[0041] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0042] The present application aims to provide a phytoplankton detection method, system and electronic device, which can solve the problems of high equipment cost, low detection precision and slow speed in the prior art, and can be applied to large-scale detection in cities.

[0043] In order to make the above objectives, characteristics and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0044] As shown in Figure 1 The method for detecting planktonic algae provided by the present application comprises the following steps.

[0045] Step 100: Obtain image data of different regions and different reaches of a river to be detected, and pre-process the image data to obtain target image data.

[0046] Step 101: Construct a river planktonic algae detection model. As shown in Figure 2 The constructed river planktonic algae detection model (Green Algae Detection Model, GADM) comprises a plurality of domain adaptive feature extraction modules (Domain Adaptation Module, DAM), a plurality of boundary alignment modules (Boundary Alignment Module, BAM), a plurality of deconvolutional modules (Deconvolutional, DeConv), a plurality of feature accumulation modules and a Sigmoid activation function layer.

[0047] Step 102: Input the target image data into the river planktonic algae detection model to obtain a pixel-by-pixel planktonic algae detection result. For example, based on the river planktonic algae detection model constructed in step 101, first pass through a plurality of serially connected domain adaptive feature extraction modules to obtain a multi-dimensional feature map of the target image data with an input dimension of 320 × 320 × 3, which is divided into 5 different levels according to the feature map scale, Figure 2 wherein the feature map size and the channel number of each level are 16, 32, 64, 128 and 128 respectively, and the length and width of each level feature map is half of that of the previous level feature map. The fifth level extraction feature is optimized by the boundary alignment module, and the feature map is restored to be consistent with the length and width of the previous level feature map by deconvolution operation. The extraction features of other levels are first optimized by the boundary alignment module, then accumulated with the deconvolution restored features, and the accumulated features are again boundary aligned. Finally, the feature map size is enlarged by deconvolution, and the pixel-by-pixel planktonic algae detection result is output by the activation function layer.

[0048] In detail, take the third domain adaptive feature extraction module DAM-3 as an example to describe the data processing process in the domain adaptive feature extraction module. Figure 2 Figure 2 ​In the specific implementation, the first domain adaptive feature extraction module is DAM-1, the second domain adaptive feature extraction module is DAM-2, and the like. The structure of the domain adaptive feature extraction module is as shown in Figure 3 The domain adaptive feature extraction module includes, in sequence, a first feature extraction unit, a second feature extraction unit, a third feature extraction unit, a feature accumulation unit, and a feature processing unit.

[0049] The feature accumulation unit is configured to fuse the features input into the first feature extraction unit and the features output by the third feature extraction unit. The first feature extraction unit includes, in sequence, a 1x1 convolution layer with 96 channels, a BN layer, and a GELU activation function layer. The second feature extraction unit includes, in sequence, a 3x3 convolution layer with 96 channels, a BN layer, and a GELU activation function layer. The third feature extraction unit includes, in sequence, a 1x1 convolution layer with 32 channels and a BN layer. The feature processing unit includes, in sequence, an IN layer and a GELU activation function layer.

[0050] Based on the structure of the domain adaptive feature extraction module, the convolution operation is used to change the feature channel number and establish neighborhood correlation. For example, 1x1 convolution is performed on the input feature (Input) with a dimension of 80x80x32 to expand the channel number to 3 times that of the input feature, to obtain a feature with a dimension of 80x80x96, and then 3x3 convolution is performed to extract the feature, and then 1x1 convolution is performed to restore the channel to the output feature consistent with the input feature. The output feature is superimposed with the input feature with a dimension of 80x80x32. In this way, the network can have sufficient information for feature extraction, and the overall parameter amount of the network can be reduced. The feature superposition mechanism can obtain the residual error between the learned feature value and the true value, and improve the feature extraction capability. The interval of the convolution operation is regularized by BatchNorm (BN layer) to convert the feature distribution interval to the range of [0, 1], to speed up the process of fitting the optimal model by gradient descent, and the GELU activation function layer is added for normalization to enhance the nonlinearity of the river plankton detection model. Finally, the superimposed feature is input into the IN layer (InstanceNorm) for standardization processing, so that the single-channel feature of a single image can be individually regularized, which can effectively correct the influence of image tone difference. Among them, InstanceNorm normalizes the feature distribution to a distribution with a mean of 0 and a variance of 1, and normalization usually normalizes the distribution to [0, 1]. The difference between BatchNorm and InstanceNorm is that BatchNorm performs standardization between different samples in a training batch, and InstanceNorm performs standardization within a sample. Compared with the former, the network training speed can be effectively accelerated, and the image tone difference can be corrected.

[0051] The specific structure of the boundary alignment module (BAM) adopted is as shown in Figure 4As shown, the boundary alignment module is mainly used for optimizing the feature map boundary, which includes two parts of neighborhood correction of feature points and feature value correction. Among them, the neighborhood correction of feature points is realized by combination of conventional convolution and asymmetric convolution. First, 8-neighborhood weight of a pixel point is obtained through 3*3 convolution, and then 1*7 convolution and 7*1 are used to extend the feature correlation from the vertical and horizontal directions respectively. The convolution kernel in the vertical direction highlights the horizontal ground object, and the convolution kernel in the horizontal direction highlights the vertical ground object. Combined with the spatial context relationship, appropriate correlation weight coefficients are obtained. The feature value correction is realized by point convolution. After neighborhood correction, there is still a certain error between the feature value of each position and the true value. Through 1*1 convolution, the information integration of multiple channels can be realized, and the feature difference of each position pixel can also be obtained, and the feature scale is not changed in the process.

[0052] Further in order to improve the detection precision and efficiency of the river plankton algae detection model, the following steps need to be implemented in the process of constructing the river plankton algae detection model:

[0053] Step 1: Obtain a deep learning training sample set. For example, obtain target images by a drone, patrol and shoot unmanned aerial vehicle images of different regions and different river sections along the river, perform preprocessing such as splicing, cropping, radiation calibration and geometric correction on the target images, select images containing plankton algae, manually mark the plankton algae targets and backgrounds according to expert knowledge, and make a deep learning training sample set. The target image is a two-dimensional visible light band color image.

[0054] For example, a data set of suspended algae is made by aerial photography of the river by a drone. The data set has a total of 36877 image pairs, and the image size is 320*320 pixels. After random allocation according to 6:1:1, the training set, test set and validation set are 27657, 4609 and 4609 image pairs respectively. In the process of inputting the image into the model, the image is randomly subjected to conventional data enhancement such as cropping, scaling and mirroring.

[0055] Step 2: Construct an initial detection model. The structure of the initial detection model constructed is the same as that of the river plankton algae detection model finally constructed, except that the model structure parameters are different.

[0056] Step 3: Train the initial detection model using the deep learning training sample set to obtain a trained model. The trained model is used as the river plankton algae detection model.

[0057] The advantages of the above-mentioned provided scheme of the present application are illustrated by comparing and verifying the detection results of the river plankton algae detection model provided by the present application with the detection results of different existing models.

[0058] In this example, the training of all network models of the present application is implemented on the python-based deep learning framework Pytorch 1.11, and the device used is NVIDIA A40. To ensure the fairness of model comparison, the hyperparameters of all network models are uniformly set as follows: training batch (epoch) is 100, learning rate is 0.01, single training sample number (batchsize) is 8, and loss function is BCEloss.

[0059] Among them, the network structures similar to DeeplabV3+, NestedUNet and ERFNet are used as comparison models. For model detection performance evaluation, the present application takes the pixel as the minimum evaluation unit, and takes the F1 value, accuracy, recall rate and overall accuracy as the evaluation index. The calculation formula of each evaluation index is as follows:

[0060]

[0061]

[0062]

[0063]

[0064] In the formula, TP represents the number of pixels whose label is algae and detection result is also algae. TN represents the number of pixels whose label is algae and detection result is background. FP represents the number of pixels whose label is background and detection result is algae. FN represents the number of pixels whose label is background and detection result is also background.

[0065] After quantization of the experimental results, as shown in Table 1, the best performance among the three comparison models is ERFNet, and the four indicators are the highest among the three models. Comparing the model proposed by the present application with the model, the F1 value, accuracy, recall rate and total accuracy of the GADM model proposed by the present application are increased by 1.92%, 0.63%, 3.1% and 1.6% respectively. The high accuracy indicates less over-detection, and the significantly improved recall rate indicates that the missed detection is significantly less than the comparison model.

[0066] Table 1 Comparison of unmanned aerial vehicle suspended algae detection accuracy (%)

[0067] Model F1 value Accuracy Recall Total accuracy DeeplabV3+ 86.830 90.241 83.668 88.084 NestedUnet 79.609 89.059 71.972 82.689 ERFNet 89.249 91.122 87.453 90.108 GADM 91.163 91.757 90.576 91.755

[0068] Region one is large-area green algae extraction in a simple scene, and region two is small-area green algae extraction in a complex scene. Figure 5It can be seen that although there is a certain false detection on the edge of each model, the distribution of green algae can be roughly identified, which shows that the unmanned aerial vehicle as a data source, combined with computer vision information interpretation technology, can be applied to intelligent green algae recognition and extraction. Region two is scattered distribution, different size and different types of algae, Figure 6 It can be seen that compared with the models DeeplabV3+, NestedUnet, there is a small amount of missing extraction, ERFNet exists over-extraction, and two independent distributed suspended algae are connected together, and the GADM of the application extracts two pieces of algae separately, which is consistent with the conclusion that the quantitative comparison recall rate is higher than that of the comparison model. Region three has a large number of shore grasses, and the types of planktonic algae are different, mainly comparing the extraction ability of the model in a complex scene. From Figure 7 It can be seen that compared with the models DeeplabV3+, NestedUnet, there is a small amount of missing extraction, ERFNet exists over-extraction, and two independent distributed suspended algae are connected together, and the GADM of the application extracts two pieces of algae separately, which is consistent with the conclusion that the quantitative comparison recall rate is higher than that of the comparison model. Region three has a large number of shore grasses, and the types of planktonic algae are different, mainly comparing the extraction ability of the model in a complex scene. From

[0069] Based on this, compared with the prior art, the application has the following advantages:

[0070] (1) In view of the characteristics of wide distribution of river planktonic algae, complex background and large image tone difference, the domain adaptive feature extraction module is used to improve the method to distinguish the difference between the target and the background.

[0071] (2) In view of the characteristics of the river planktonic algae, the boundary alignment module is designed to enhance the target feature extraction capability.

[0072] (3) The application provides a river planktonic algae detection method based on unmanned aerial vehicle image, which has the advantages of fast speed and high precision compared with the prior art, and can effectively reduce the economic and ecological cost of planktonic algae treatment.

[0073] Further, the application also provides the following implementation structure:

[0074] A planktonic algae detection system applied to the above-provided planktonic algae detection method. The system comprises:

[0075] A data acquisition module for acquiring image data of different regions and different river sections of the river to be detected, and preprocessing the image data to obtain target image data.

[0076] The model construction module is configured to construct a river algae detection model.

[0077] The result detection module is configured to input the target image data into the river algae detection model to obtain a pixel-by-pixel algae detection result.

[0078] An electronic device comprises:

[0079] A memory is configured to store a computer program.

[0080] A processor is connected to the memory and configured to call and execute the computer program to implement the algae detection method provided above.

[0081] In addition, the computer program in the memory described above is implemented in the form of a software function unit and sold or used as an independent product. Based on this understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory, a random access memory, a magnetic disk or an optical disk, and various media that can store program codes.

[0082] In summary, the principle of the present application is to fly over the river by a UAV and use an artificial intelligence vision algorithm to intelligently interpret algae. A GADM (Gadm Detection Model) is constructed and pre-trained, a boundary alignment module is designed to improve the ability of the feature to distinguish the difference between the target and the background, and a domain adaptive feature extraction module is designed to improve the multi-scale object recognition ability. The river is regularly inspected by a UAV, the image is input into the model, and the pixel-by-pixel algae detection result is output, which can be effectively applied to urban river management and pollution prevention.

[0083] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method part.

[0084] The principles and implementation manners of the present application are described by using specific examples in the present application, and the above examples are only used to help understand the method of the present application and its core idea; meanwhile, for the general technical personnel in the art, the specific implementation manners and application ranges will be changed according to the idea of the present application. In conclusion, the content of the present specification should not be understood as the limitation of the present application.

Claims

1. A method for detecting phytoplankton, characterized in that: include: Acquire image data of different areas and sections of the river to be detected, and preprocess the image data to obtain target image data; Construct a model for detecting phytoplankton in rivers; The river phytoplankton detection model includes: multiple domain adaptive feature extraction modules, multiple boundary alignment modules, multiple deconvolution modules, multiple feature accumulation modules and an activation function layer; wherein the domain adaptive feature extraction module includes a first feature extraction unit, a second feature extraction unit, a third feature extraction unit, a feature accumulation unit and a feature processing unit cascaded in sequence; the feature accumulation unit is used to fuse the features input to the first feature extraction unit and the features output by the third feature extraction unit; the first feature extraction unit includes a 1×1 convolution layer, a BN layer and a GELU activation function layer cascaded in sequence; the second feature extraction unit includes a 3×3 convolution layer, a BN layer and a GELU activation function layer cascaded in sequence; the third feature extraction unit includes a 1×1 convolution layer and a BN layer cascaded in sequence; the feature processing unit includes an IN layer and a GELU activation function layer cascaded in sequence; Inputting the target image data into the river phytoplankton detection model to obtain pixel-by-pixel phytoplankton detection results includes: The target image data is input into the first domain adaptive feature extraction module, the features output by the first domain adaptive feature extraction module are respectively input into the second domain adaptive feature extraction module and the first boundary alignment module, the features output by the second domain adaptive feature extraction module are respectively input into the third domain adaptive feature extraction module and the second boundary alignment module, the features output by the third domain adaptive feature extraction module are respectively input into the fourth domain adaptive feature extraction module and the third boundary alignment module, the features output by the fourth domain adaptive feature extraction module are respectively input into the fifth domain adaptive feature extraction module and the fourth boundary alignment module, the features output by the fifth domain adaptive feature extraction module are input into the fifth boundary alignment module, the features output by the fifth boundary alignment module are input into the fifth deconvolution module, the features output by the fifth deconvolution module and the features output by the fourth boundary alignment module are both input into the fourth feature accumulation module, the features output by the fourth feature accumulation module are input into the ninth boundary alignment module, and the ninth boundary alignment module The output features are input into the fourth deconvolution module, the features output by the fourth deconvolution module and the features output by the third boundary alignment module are both input into the third feature accumulation module, the features output by the third feature accumulation module are input into the eighth boundary alignment module, the features output by the eighth boundary alignment module are input into the third deconvolution module, the features output by the third deconvolution module and the features output by the second boundary alignment module are both input into the second feature accumulation module, the features output by the second feature accumulation module are input into the seventh boundary alignment module, the features output by the seventh boundary alignment module are input into the second deconvolution module, the features output by the second deconvolution module and the features output by the first boundary alignment module are both input into the first feature accumulation module, the features output by the first feature accumulation module are input into the sixth boundary alignment module, the features output by the sixth boundary alignment module are input into the first deconvolution module, the features output by the first deconvolution module are input into the activation function layer, and the data output by the activation function layer is used as the phytoplankton detection result.

2. The method for detecting phytoplankton according to claim 1, wherein: The number of channels of the 1×1 convolutional layer in the first feature extraction unit and the number of channels of the 3×3 convolutional layer in the second feature extraction unit are both 96; the number of channels of the 1×1 convolutional layer in the third feature extraction is 32.

3. The method for detecting phytoplankton according to claim 1, wherein: The boundary alignment module includes neighborhood correction and eigenvalue correction of feature points.

4. The method for detecting phytoplankton according to claim 1, wherein: Construct a river phytoplankton detection model, including: Obtain deep learning training sample sets; Build an initial detection model; The deep learning training sample set is used to train the initial detection model to obtain a trained model; and the trained model is used as the river phytoplankton detection model.

5. The method for detecting phytoplankton according to claim 4, wherein: The initial detection model was trained using NVIDIA A40 and a Python-based deep learning framework, Pytorch 1.

11.

6. A phytoplankton detection system, characterized in that: The method for detecting phytoplankton according to any one of claims 1 to 5, wherein the system comprises: A data acquisition module is used to acquire image data of different areas and sections of the river to be detected, and pre-process the image data to obtain target image data; A model building module for building a river phytoplankton detection model; the river phytoplankton detection model includes: multiple domain adaptive feature extraction modules, multiple boundary alignment modules, multiple deconvolution modules, multiple feature accumulation modules and an activation function layer; The result detection module is used to input the target image data into the river phytoplankton detection model to obtain pixel-by-pixel phytoplankton detection results.

7. An electronic device, characterized in that: include: Memory for storing computer programs; A processor, connected to the memory, is used to retrieve and execute the computer program to implement the phytoplankton detection method according to any one of claims 1 to 5.

Images