Water line identification system and method based on improved YOLOv8

Abstract

The invention discloses a water line identification system and method based on improved YOLOv8. The system comprises an image acquisition module, an image processing module, a model construction module and a water level detection module. The image acquisition module is used for acquiring historical water level images; the image processing module is used for marking a water level line in the historical water level image and constructing a training data set; the model construction module is used for constructing an improved YOLOv8 target recognition model and training the target recognition model based on the training data set to obtain a water level line recognition model; the water level detection module collects a real-time water level image and identifies the real-time water level image based on the water level line identification model to obtain a real-time water level line. The problem of inaccurate water level detection in the existing detection algorithm is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image recognition, and in particular relates to a water level line recognition system and method based on improved YOLOv8. Background Art

[0002] Water level measurement is a technology that studies the location of the interface between two different media. Its background and significance are far-reaching, covering multiple fields such as disaster prevention and mitigation, water resources management, and ecological protection. With the global climate change and the intensification of human activities, the importance of water level detection has become increasingly prominent. In terms of disaster prevention and mitigation, water level detection is an important part of the flood warning system. In terms of water resources management, the role of water level detection cannot be ignored. With population growth and economic development, the contradiction between water supply and demand has become increasingly prominent. Accurate water level data helps to scientifically and rationally dispatch water resources, optimize the management of reservoirs, rivers, and irrigation systems, and ensure the stability of agricultural water use and urban water supply. Through water level detection, the dynamic changes of water resources can be grasped in real time, and a more reasonable water resource utilization and allocation plan can be formulated. This not only helps to improve the efficiency of water resource utilization, but also effectively responds to extreme climate conditions such as droughts, and guarantees the water demand for agricultural production and residents.

[0003] At present, deep learning-based detection methods have almost replaced traditional detection methods, which are divided into two-stage and one-stage detectors. The two-stage detector is based on the proposed area, and then obtains the object class in the area box, such as R-CNN, Fast R-CNN and Faster R-CNN. The one-stage detector directly predicts the position and category to achieve real-time target detection, such as SSD300, CenterNet and YOLO series. The above aspects have the existence of waterline in the construction of the current target detection method. It is only for the improvement and optimization of common data sets such as PASCAL VOC and MS COCO, and lacks the problem of waterline detection. Summary of the invention

[0004] The present invention aims to solve the deficiencies of the prior art and provides the following solutions:

[0005] A water level recognition system based on improved YOLOv8, comprising: an image acquisition module, an image processing module, a model building module and a water level detection module;

[0006] The image acquisition module is used to acquire historical water level images;

[0007] The image processing module is used to mark the water level lines in the historical water level image and construct a training data set;

[0008] The model building module is used to build an improved YOLOv8 target recognition model, and train the target recognition model based on the training data set to obtain a water level recognition model;

[0009] The water level detection module is used to collect a real-time water level image, and recognize the real-time water level image based on the water level line recognition model to obtain a real-time water level line.

[0010] Preferably, the image processing module includes: a feature annotation unit and a data set construction unit;

[0011] The feature annotation unit is used to annotate the water level line in the historical water level image;

[0012] The data set construction unit is used to divide the annotated images into a training set and a test set according to a preset ratio to obtain the training data set.

[0013] Preferably, the model building module includes: a model building unit and a model training unit;

[0014] The model building unit is used to build a YOLOv8 model, and improve the YOLOv8 model based on the CA attention mechanism and the DDHA fusion attention mechanism decoupling detection head to obtain the target recognition model;

[0015] The model training unit is used to train the target recognition model based on the training data set to obtain the water level line recognition model.

[0016] Preferably, the target recognition model comprises: a backbone network, an improved head network and an improved detection head;

[0017] The backbone network includes: a CBS module, a C2F module and an SPPF module;

[0018] The improved head network is improved from the head network of the initial YOLOv8 model: the CA attention mechanism, the target detection layer and the DDHA fusion attention mechanism decoupling detection head are added to the head network of the initial YOLOv8 model, and the improved head network includes: a CA attention module, a CBS module, a C2F module, an Up-sample module and a DDHA fusion attention mechanism decoupling detection head;

[0019] The improved detection head is improved from the detection head of the initial YOLOv8 model: a new multi-target detection layer is added to the detection head of the initial YOLOv8 model.

[0020] Preferably, the DDHA fusion attention mechanism decoupling detection head is composed of a CBS module, an ECA module, an ESA module and a Concat connection layer;

[0021] The workflow of the DDHA fusion attention mechanism decoupling detection head includes:

[0022] F′=CBS(F)

[0023]

[0024] F″′=F+F″

[0025] Among them, CBS represents convolution with a kernel of 1 and a stride of 1; Concat represents tensor concatenation; ECA() represents a high-performance channel attention module; ESA() represents a high-performance spatial attention module; F represents the module input feature map; F' represents the CBS module output feature map; F" represents the Concat output feature map; F"' represents the output feature map of the DDHA fusion attention mechanism decoupled detection head.

[0026] The present invention also provides a water level line recognition method based on improved YOLOv8, and the recognition method is applied to any of the above-mentioned recognition systems, comprising the following steps:

[0027] S1. Collect historical water level images;

[0028] S2. marking the water level lines in the historical water level image to construct a training data set;

[0029] S3. Construct an improved YOLOv8 target recognition model, and train the target recognition model based on the training data set to obtain a water level recognition model;

[0030] S4. Collecting a real-time water level image, and identifying the real-time water level image based on the water level line recognition model to obtain a real-time water level line.

[0031] Preferably, S2 includes:

[0032] marking the water level line in the historical water level image;

[0033] The annotated images are divided into a training set and a test set according to a preset ratio to obtain the training data set.

[0034] Preferably, S3 includes:

[0035] Construct a YOLOv8 model, and improve the YOLOv8 model based on the CA attention mechanism and the DDHA fusion attention mechanism decoupling detection head to obtain the target recognition model;

[0036] The target recognition model is trained based on the training data set to obtain the water level line recognition model.

[0037] Preferably, the target recognition model comprises: a backbone network, an improved head network and an improved detection head;

[0038] The backbone network includes: a CBS module, a C2F module and an SPPF module;

[0039] The improved head network is improved from the head network of the initial YOLOv8 model: the CA attention mechanism, the target detection layer and the DDHA fusion attention mechanism decoupling detection head are added to the head network of the initial YOLOv8 model, and the improved head network includes: a CA attention module, a CBS module, a C2F module, an Up-sample module and a DDHA fusion attention mechanism decoupling detection head;

[0040] The improved detection head is improved from the detection head of the initial YOLOv8 model: a new multi-target detection layer is added to the detection head of the initial YOLOv8 model.

[0041] Preferably, the DDHA fusion attention mechanism decoupling detection head is composed of a CBS module, an ECA module, an ESA module and a Concat connection layer;

[0042] The workflow of the DDHA fusion attention mechanism decoupling detection head includes:

[0043] F′=CBS(F)

[0044]

[0045] F″′=F+F″

[0046] Among them, CBS represents convolution with a kernel of 1 and a stride of 1; Concat represents tensor concatenation; ECA() represents a high-performance channel attention module; ESA() represents a high-performance spatial attention module; F represents the module input feature map; F' represents the CBS module output feature map; F" represents the Concat output feature map; F"' represents the output feature map of the DDHA fusion attention mechanism decoupled detection head.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] The present invention is based on the YOLOv8 target recognition model, and its framework includes a backbone network and a head network. A DDHA detection head is designed. A CA module is used in the backbone network, and a DDHA module and a target detection layer are added to the head network to construct a water level line detection method, which solves the problem of inaccurate water level line detection in the existing detection algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0050] Figure 1 A schematic diagram of the system structure of an embodiment of the present invention;

[0051] Figure 2 A schematic diagram of the network structure of a water level recognition model according to an embodiment of the present invention;

[0052] Figure 3 A schematic diagram of the structure of a CBS module according to an embodiment of the present invention;

[0053] Figure 4 This is a schematic diagram of the structure of a C2F module according to an embodiment of the present invention;

[0054] Figure 5 It is a structural schematic diagram of the SPPF module of an embodiment of the present invention;

[0055] Figure 6 This is a structural schematic diagram of the DDHA fusion attention mechanism decoupling detection head in an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

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

[0058] Embodiment 1

[0059] In this embodiment, if Figure 1 As shown, a water level recognition system based on improved YOLOv8 includes: an image acquisition module, an image processing module, a model building module and a water level detection module.

[0060] The image acquisition module is used to collect historical water level images.

[0061] The image processing module is used to annotate the water level lines in the historical water level images and construct a training data set.

[0062] The image processing module includes: a feature annotation unit and a data set construction unit; the feature annotation unit is used to annotate the water level lines in the historical water level image; the data set construction unit is used to divide the annotated image into a training set and a test set according to a preset ratio to obtain a training data set. In this embodiment, the preset ratio is set to: training set: test set = 8:2.

[0063] The model building module is used to build an improved YOLOv8 target recognition model, and train the target recognition model based on the training data set to obtain a water level recognition model.

[0064] The model building module includes: a model building unit and a model training unit. The model building unit is used to build a YOLOv8 model, and improve the YOLOv8 model based on the CA attention mechanism and the DDHA fusion attention mechanism decoupled detection head to obtain a target recognition model; the model training unit is used to train the target recognition model based on the training data set to obtain a water level recognition model, such as Figure 2 shown.

[0065] The target recognition model includes: a backbone network, an improved head network and an improved detection head; the backbone network includes: a CBS module, a C2F module and an SPPF module; the improved head network is improved from the head network of the initial YOLOv8 model: a CA attention mechanism, a target detection layer and a DDHA fusion attention mechanism decoupled detection head are added to the head network of the initial YOLOv8 model, and the improved head network includes: a CA attention module, a CBS module, a C2F module, an Up-sample module and a DDHA fusion attention mechanism decoupled detection head; the improved detection head is improved from the detection head of the initial YOLOv8 model: a new multi-target detection layer is added to the detection head of the initial YOLOv8 model.

[0066] In this embodiment, the CBS module is composed of a convolution module, a normalization module and a SiLU activation function, and its structure is as follows: Figure 3 As shown. The C2F module is the main feature extraction module in YOLOv8, and its structure is as follows Figure 4 The SPPF module is a residual structure. Figure 5 As shown in the figure, features are extracted through three consecutive maximum pooling operations. The same 5×5 convolution kernel is used in each maximum pooling process. Finally, the results before and after pooling are concat-operated to retain the detailed information before pooling. The SPPF module gradually reduces the spatial size of the feature map and can better capture feature information at different scales. At the same time, by using three consecutive maximum pooling operations, each maximum pooling operation effectively reduces the size of the feature map, improves the computational efficiency, and realizes the fusion of multi-scale features.

[0067] Furthermore, the CA attention mechanism can improve the network's attention to feature areas. The use of the attention mechanism can improve the network's feature fusion capabilities. Here, the CA attention module is used to complete the addition of the CA attention mechanism. The CA attention module can obtain the location information of the target object and introduce its horizontal and vertical coordinate information into the attention of different channels, thereby improving the large-scale feature extraction capability of the target detection algorithm while reducing the computational burden. The CA attention module processes the input feature map in two main steps: coordinate information embedding and coordinate attention generation.

[0068] (1) Information embedding of coordinates means that the input feature map is average pooled in the X and Y directions. The channel outputs after average pooling with the pooling kernels (H, 1) and (1, W) are as follows. The output formula of the feature map at height h with output channel c is as follows:

[0069]

[0070] Similarly, the output formula of the feature map with width w and output channel c is as follows:

[0071]

[0072] Where W and H represent the width and height of the input image, x c (h, i) represents the value of the input image at the coordinate position (h, i), x c (j, w) represents the value of the input image at the coordinate position (j, w).

[0073] (2) The generation of coordinate attention is done after the coordinate information is embedded. The embedded information is concatenated and convolved, and then nonlinear operations are implemented through activation functions. The converted result generates more accurate coordinate position information. The formula of the entire operation process is as follows:

[0074] f=δ(F1[z h *,z w ])

[0075] Among them, [z h , z w ] represents the stitching work after the coordinate information is embedded, F1 represents the 1×1 convolution operation, and δ represents the activation function for feature map normalization and nonlinear operation (sigmoid is generally used as the activation function to reduce the complexity and calculation amount of the model). The final result is is the intermediate feature of spatial information mapping in the horizontal and vertical directions, and r represents the pooling step size. Afterwards, it is decomposed into two two-dimensional tensors according to the spatial dimension and The two-dimensional tensor is expanded to the same number of channels as the input through two 1×1 convolution operations, and the result is as follows:

[0076] g h =8((F h (f h ))

[0077] g w =δ(F w (f w ))

[0078] Among them, F h Indicates f h Channel function for 1×1 convolution operation, F w Indicates f w Channel function for 1×1 convolution operation, g h Indicates f h The result after nonlinear transformation, g w Indicates f w The result after nonlinear transformation. h and g w The attention weight matrix multiplication process is performed to obtain the final output of the CA attention module, which is calculated as follows:

[0079]

[0080] Among them, x c (i, j) represents the input feature of the image at coordinate (i, j), y c (i, j) represents the output feature of the image at coordinate (i, j), Represents the ordinate output result of the attention weight matrix multiplication, Represents the horizontal axis output result of the attention weight matrix multiplication.

[0081] Furthermore, the detection head of the original YOLOv8 model only provides three different scales of Head detection layers, which are used to detect small targets, medium targets and large targets respectively. After multiple downsampling, the target information in the feature map becomes very sparse, and the original three Head detection layers can no longer meet the detection requirements for the water level line, which easily leads to missed detections. In order to solve this problem and improve the network's detection capability and overall detection effect for small targets, this embodiment adds a new small target detection layer on the basis of the original three Head detection layers, and adds a new scale feature map on the basis of the original algorithm. The original image is downsampled by 4 times, 8 times, 16 times, and 32 times to obtain a 20×20 large target detection feature map, a 40×40 medium target feature detection map, an 80×80 small target feature detection map, and a 160×160 small target feature detection map, as shown in FIG. Figure 2As shown, the detection performance and reliability of the original model are improved.

[0082] In this embodiment, the DDHA fusion attention mechanism decoupling detection head is composed of a CBS module, an ECA module, an ESA module and a Concat connection layer. The structure is as follows: Figure 6 As shown in the figure. The workflow of DDHA fusion attention mechanism decoupling detection head includes:

[0083] F′=CBS(F)

[0084]

[0085] F″′=F+F″

[0086] Among them, CBS represents convolution with a kernel of 1 and a stride of 1; Concat represents tensor concatenation; ECA() represents a high-performance channel attention module; ESA() represents a high-performance spatial attention module; F represents the module input feature map; F' represents the CBS module output feature map; F" represents the Concat output feature map; F"' represents the output feature map of the DDHA fusion attention mechanism decoupled detection head.

[0087] After the water level recognition model is built, the accuracy (Precision), recall (Recall) and average precision (Average Precision, AP) are selected as evaluation indicators to compare the performance of this model with other models. The accuracy, recall and average precision are defined as follows:

[0088]

[0089] Among them, R i represents the recall rate, TP represents the number of correctly detected positive samples, FN represents the number of incorrectly detected negative samples, and P i It represents the accuracy, and FP represents the number of falsely detected positive samples.

[0090] The water level detection module is used to collect real-time water level images, and recognize the real-time water level images based on the water level line recognition model to obtain the real-time water level line. In this embodiment, a high-definition camera is used to collect real-time water level images. If there are more than two (including two) water level line detection targets in the image, the least squares method is used to fit the position coordinates of the identified water level lines to obtain the equation of the straight line, which is the real-time water level line.

[0091] Embodiment 2

[0092] In this embodiment, a water level recognition method based on improved YOLOv8 includes the following steps:

[0093] S1. Collect historical water level images.

[0094] S2. Label the water level lines in the historical water level images and construct a training dataset.

[0095] S2 includes: marking the water level lines in the historical water level image; dividing the marked image into a training set and a test set according to a preset ratio to obtain a training data set.

[0096] S3. Build an improved YOLOv8 target recognition model and train the target recognition model based on the training data set to obtain a water level recognition model.

[0097] S3 includes: building a YOLOv8 model, and improving the YOLOv8 model based on the CA attention mechanism and the DDHA fusion attention mechanism decoupling detection head to obtain a target recognition model; training the target recognition model based on the training data set to obtain a water level recognition model.

[0098] The target recognition model includes: a backbone network, an improved head network and an improved detection head; the backbone network includes: a CBS module, a C2F module and an SPPF module; the improved head network is improved from the head network of the initial YOLOv8 model: a CA attention mechanism, a target detection layer and a DDHA fusion attention mechanism decoupled detection head are added to the head network of the initial YOLOv8 model, and the improved head network includes: a CA attention module, a CBS module, a C2F module, an Up-sample module and a DDHA fusion attention mechanism decoupled detection head; the improved detection head is improved from the detection head of the initial YOLOv8 model: a new multi-target detection layer is added to the detection head of the initial YOLOv8 model.

[0099] The DDHA fusion attention mechanism decoupling detection head consists of a CBS module, an ECA module, an ESA module, and a Concat connection layer; the workflow of the DDHA fusion attention mechanism decoupling detection head includes:

[0100] F′=CBS(F)

[0101]

[0102] F″′=F+F″

[0103] Among them, CBS represents convolution with a kernel of 1 and a stride of 1; Concat represents tensor concatenation; ECA() represents a high-performance channel attention module; ESA() represents a high-performance spatial attention module; F represents the module input feature map; F' represents the CBS module output feature map; F" represents the Concat output feature map; F"' represents the output feature map of the DDHA fusion attention mechanism decoupled detection head.

[0104] S4. Collect a real-time water level image, and recognize the real-time water level image based on a water level line recognition model to obtain a real-time water level line.

[0105] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A water level recognition system based on improved YOLOv8, characterized in that: include: Image acquisition module, image processing module, model building module and water level detection module; The image acquisition module is used to acquire historical water level images; The image processing module is used to mark the water level lines in the historical water level image and construct a training data set; The model building module is used to build an improved YOLOv8 target recognition model, and train the target recognition model based on the training data set to obtain a water level recognition model; The water level detection module is used to collect a real-time water level image, and recognize the real-time water level image based on the water level line recognition model to obtain a real-time water level line.

2. According to claim 1, a water level recognition system based on improved YOLOv8 is characterized in that: The image processing module includes: a feature annotation unit and a data set construction unit; The feature annotation unit is used to annotate the water level line in the historical water level image; The data set construction unit is used to divide the annotated images into a training set and a test set according to a preset ratio to obtain the training data set.

3. According to claim 1, a water level recognition system based on improved YOLOv8 is characterized in that: The model building module includes: a model building unit and a model training unit; The model building unit is used to build a YOLOv8 model, and improve the YOLOv8 model based on the CA attention mechanism and the DDHA fusion attention mechanism decoupling detection head to obtain the target recognition model; The model training unit is used to train the target recognition model based on the training data set to obtain the water level line recognition model.

4. According to claim 3, a water level recognition system based on improved YOLOv8 is characterized in that: The target recognition model includes: a backbone network, an improved head network and an improved detection head; The backbone network includes: a CBS module, a C2F module and an SPPF module; The improved head network is improved from the head network of the initial YOLOv8 model: the CA attention mechanism, the target detection layer and the DDHA fusion attention mechanism decoupling detection head are added to the head network of the initial YOLOv8 model, and the improved head network includes: a CA attention module, a CBS module, a C2F module, an Up-sample module and a DDHA fusion attention mechanism decoupling detection head; The improved detection head is improved from the detection head of the initial YOLOv8 model: a new multi-target detection layer is added to the detection head of the initial YOLOv8 model.

5. According to claim 4, a water level recognition system based on improved YOLOv8 is characterized in that: The DDHA fusion attention mechanism decoupling detection head is composed of a CBS module, an ECA module, an ESA module and a Concat connection layer; The workflow of the DDHA fusion attention mechanism decoupling detection head includes: F′=CBS(F) F″′=F+F″ Among them, CBS represents convolution with a kernel of 1 and a stride of 1; Concat represents tensor concatenation; ECA() represents a high-performance channel attention module; ESA() represents a high-performance spatial attention module; F represents the module input feature map; F' represents the CBS module output feature map; F" represents the Concat output feature map; F"' represents the output feature map of the DDHA fusion attention mechanism decoupled detection head.

6. A water level recognition method based on improved YOLOv8, the recognition method is applied to the recognition system according to any one of claims 1 to 5, characterized in that: The following steps are involved: S1. Collect historical water level images; S2. marking the water level lines in the historical water level image to construct a training data set; S3. Construct an improved YOLOv8 target recognition model, and train the target recognition model based on the training data set to obtain a water level recognition model; S4. Collecting a real-time water level image, and identifying the real-time water level image based on the water level line recognition model to obtain a real-time water level line.

7. According to claim 6, a water level recognition method based on improved YOLOv8 is characterized in that: The S2 includes: marking the water level line in the historical water level image; The annotated images are divided into a training set and a test set according to a preset ratio to obtain the training data set.

8. According to claim 6, a water level recognition method based on improved YOLOv8 is characterized in that: The S3 includes: Construct a YOLOv8 model, and improve the YOLOv8 model based on the CA attention mechanism and the DDHA fusion attention mechanism decoupling detection head to obtain the target recognition model; The target recognition model is trained based on the training data set to obtain the water level line recognition model.

9. The water level recognition method based on improved YOLOv8 according to claim 8, characterized in that: The target recognition model includes: a backbone network, an improved head network and an improved detection head; The backbone network includes: a CBS module, a C2F module and an SPPF module; The improved head network is improved from the head network of the initial YOLOv8 model: the CA attention mechanism, the target detection layer and the DDHA fusion attention mechanism decoupling detection head are added to the head network of the initial YOLOv8 model, and the improved head network includes: a CA attention module, a CBS module, a C2F module, an Up-sample module and a DDHA fusion attention mechanism decoupling detection head; The improved detection head is improved from the detection head of the initial YOLOv8 model: a new multi-target detection layer is added to the detection head of the initial YOLOv8 model.

10. The water level recognition method based on improved YOLOv8 according to claim 9, characterized in that: The DDHA fusion attention mechanism decoupling detection head is composed of a CBS module, an ECA module, an ESA module and a Concat connection layer; The workflow of the DDHA fusion attention mechanism decoupling detection head includes: F′=CBS(F) F″′=F+F″ Among them, CBS represents convolution with a kernel of 1 and a stride of 1; Concat represents tensor concatenation; ECA() represents a high-performance channel attention module; ESA() represents a high-performance spatial attention module; F represents the module input feature map; F' represents the CBS module output feature map; F" represents the Concat output feature map; F"' represents the output feature map of the DDHA fusion attention mechanism decoupled detection head.

Images