Detection method of small floating targets on water surface based on improved YOLOv6 network model

Abstract

The present invention relates to a method for detecting small floating targets on the water surface based on an improved YOLOv6 network model. The method comprises: using a pre-trained improved YOLOv6 network model to detect images of floating targets on the water surface and outputting small surface target detection results. The improved YOLOv6 network model uses a detail information enhancement module to improve its ability to extract detail information of small targets, an adaptive noise suppression module to suppress noise, and a standardized Gaussian Wasserstein distance to calculate regression loss during model training. The present invention can accurately detect small floating targets on the water surface.

Description

Technical Field

[0001] The present invention relates to the field of target detection, and in particular to a method for detecting small targets floating on a water surface based on an improved YOLOv6 network model. Background Art

[0002] With the rapid development of the global economy and the increase in human activities, marine environmental pollution is becoming increasingly prominent, with floating debris being a particularly significant threat to the marine environment. This floating debris not only affects the balance and stability of marine ecosystems but also poses a serious threat to the development and utilization of marine resources. Therefore, effectively detecting and recovering floating debris has become a crucial step in addressing this environmental issue.

[0003] In the past, surface debris removal often relied on manual salvage, but this approach faced challenges such as high risk, low efficiency, and insufficient manpower. In recent years, with the rapid development of deep learning, surface debris detection technology has made substantial progress. By combining various sensors with highly efficient detection algorithms, rapid and accurate detection and salvage of surface debris can be achieved. The application of this technology is of great significance.

[0004] While existing object detection algorithms have made significant progress in both accuracy and inference speed in recent years, significant improvements have been made in small object detection tasks. Some models even achieve less than half the accuracy of large object detection for small objects. In floating object detection datasets, small objects account for more than half of the total number of objects. Therefore, improving the detection accuracy of small objects is crucial for this task.

[0005] The inventors of this application found that the main reasons for the low accuracy of small target detection are as follows: (1) insufficient detail information of small targets on the feature map; (2) the features of small targets are not obvious enough and are easily affected by background interference; (3) the loss function has limitations.

[0006] In response to the above-mentioned problems existing in the task of detecting floating targets on the water surface, the present invention provides a method for detecting small floating targets on the water surface. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide a detection method for small floating targets on the water surface based on an improved YOLOv6 network model, which can accurately detect small floating targets on the water surface.

[0008] In order to solve the above technical problems, the technical solution of the present invention is: a method for detecting small floating targets on the water surface based on an improved YOLOv6 network model, comprising:

[0009] The pre-trained improved YOLOv6 network model is used to detect the floating target image on the water surface and output the small target detection results on the water surface;

[0010] The improved YOLOv6 network model adopts a detail information enhancement module to improve the ability to extract detail information of small targets, adopts an adaptive noise suppression module to suppress noise, and uses a standardized Gaussian Wasserstein distance to calculate the regression loss in model training.

[0011] Furthermore, the working process of the detail information enhancement module includes:

[0012] The dilated convolution with different expansion rates is used to extract the detail feature maps on the low-level feature maps in parallel, and the extracted detail feature maps are fused in the feature dimension. cat Input channel attention module, generate feature map channel attention weight α; add channel attention weight α to feature map F cat Multiply, split the multiplied feature map sequentially in the channel dimension according to the ratio before fusion and sum them up. The formula is as follows:

[0013] D i =ψ(F cat ⊙α), (1)

[0014] Among them, ⊙ represents the vector dot product, ψ(·) represents the function of sequential splitting and element-wise summation, and D i Feature maps generated for the detail information enhancement module;

[0015] α=σ(W(P gavg (F cat ))), (2)

[0016] Among them, W∈R is the one-dimensional convolution weight with a step size of k, P gavg (·) is the global average pooling function, and σ(·) is the Sigmoid function.

[0017]

[0018] Where |t| odd represents the odd number closest to the absolute value of t, C is the number of feature map channels, γ = 2, b = 1.

[0019] Furthermore, the working process of the adaptive noise suppression module includes:

[0020] The feature map G is pooled using local average pooling. iDownsampling is performed, and then the downsampled feature map is feature mapped, and then the mapped feature map is subjected to the Sigmid function to generate the spatial attention weight feature map, and finally upsampling is performed to restore it to the feature map G i Size and multiply it into the feature map G i , the formula is as follows:

[0021] P i =G i ⊙S i , (4)

[0022] Where ⊙ represents the vector dot product. i The formula is:

[0023] S i =U(σ(W2·δ(BN(W1·P lavg (G i ))))), (5)

[0024] Among them, U(·) represents the upsampling function, σ(·) is the Sigmoid function, δ(·) is the Relu function, W1, W2∈R are the convolution weights with a step size of 1 and a convolution kernel size of 1, BN(·) is the BatchNorm2d function, P lavg (·) represents the local average pooling function.

[0025] Furthermore, the specific steps for using the standardized Gaussian Wasserstein distance to calculate the regression loss in model training include:

[0026] First, the Bounding Boxes are modeled as a Gaussian distribution, and then the standardized Wasserstein distance is used to measure the similarity between the two distributions;

[0027] For a horizontal bounding box B = (cx, cy, w, h), it is modeled as a 2D Gaussian distribution as follows:

[0028]

[0029] Where μ is the mean, ∑ is the variance, and the similarity between two bounding boxes is represented by the distance between the Gaussian distributions of the boxes;

[0030] For two 2D Gaussian distributions, m1 = N(μ1,∑1) and m2 = N(μ2,∑2), their second-order Wasserstein distance can be defined as:

[0031]

[0032] in Frobenius norm, for two bounding boxes A = (cx a ,cy a ,w a ,h a ) and B=(cx b ,cy b ,w b ,h b ), formula (7) is simplified as follows:

[0033]

[0034] W2 2 (N a ,N b ) is a distance metric and cannot be used directly as a similarity metric. Therefore, its exponential form is normalized to obtain a new metric called the normalized Wasserstein distance NWD:

[0035]

[0036] Where C is a constant that is closely related to the dataset. The bounding box loss function based on the NWD indicator is designed as follows:

[0037] L NWD =1-NWD(N p ,N g ), (10)

[0038] where N p is the Gaussian distribution model of the prediction box, N g It is the Gaussian distribution model of the true value box, and the NWD-based loss function is used as the bounding box loss function of our model. The total loss function is as follows:

[0039] L sum =L cls +λL NWD +βL obj , (11)

[0040] Among them L cls and L obj are the classification loss and target loss, λ and β are the bounding box loss weight and target loss weight.

[0041] Furthermore, the method also includes: before using the training set to train the improved YOLOv6 network model, performing data enhancement processing on the surface floating target images in the training set.

[0042] After adopting the above technical solution, the present invention uses convolutional building blocks to form a backbone-neck-detection head structure network. The present invention designs a detail information enhancement module and an adaptive noise suppression module within the network structure. The detail information enhancement module is designed to better capture the detailed information of small targets, thereby improving the detection accuracy of small targets, while the adaptive noise suppression module can adaptively suppress background noise interference, thereby highlighting the characteristics of small targets and further improving the detection capability of small targets. In addition, the present invention introduces a loss function based on the standardized Wasserstein distance metric, which can better measure the similarity between the small target prediction box and the target box, so that the model converges better during the training process. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a network model diagram of the improved YOLOv6 network model of the present invention;

[0044] Figure 2 This is a flow chart of a method for detecting small floating targets on the water surface based on an improved YOLOv6 network model of the present invention;

[0045] Figure 3 This is a structural diagram of the detail information enhancement module of the present invention;

[0046] Figure 4 is a structural diagram of the adaptive noise suppression module of the present invention;

[0047] Figure 5 PR curve of the Flow+-test results of the improved YOLOv6 network model in the present invention;

[0048] Figure 6 This figure shows the detection results of some images in the Flow+-test using the improved YOLOv6 network model in this invention. DETAILED DESCRIPTION

[0049] In order to make the contents of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments in conjunction with the accompanying drawings.

[0050] like Figure 1 、 2 As shown in FIG, a detection method for small floating targets on the water surface based on an improved YOLOv6 network model includes:

[0051] The pre-trained improved YOLOv6 network model is used to detect the floating target image on the water surface and output the small target detection results on the water surface;

[0052] The improved YOLOv6 network model adopts a detail information enhancement module to improve the ability to extract detail information of small targets, adopts an adaptive noise suppression module to suppress noise, and uses a standardized Gaussian Wasserstein distance to calculate the regression loss in model training.

[0053] The improved YOLOv6 network model in this embodiment enhances the extraction of small target detail feature information, suppresses the noise information that interferes with the target on the feature map, and uses a better regression loss function, thereby being able to accurately detect small floating targets on the water surface.

[0054] The technical solutions involved in the above embodiments are described in detail below in conjunction with specific embodiments.

[0055] like Figure 2 As shown in FIG, a detection method for small floating targets on the water surface based on an improved YOLOv6 network model includes:

[0056] S1, divide the surface floating target dataset into training set, validation set and test set in the ratio of 6:2:2, and perform data augmentation on the training set;

[0057] S2: To address the problem of insufficient detail information extraction of small targets, a detail information enhancement module is designed to improve the model's ability to extract detail information of small targets. In view of the phenomenon that water scenes have a lot of noise interference, an adaptive noise suppression module is designed to suppress noise. As the intersection-based measurement method is not friendly to small targets, a regression loss function based on the standardized Wasserstein distance measurement method is introduced;

[0058] S3: Build the improved YOLOv6 network model. The backbone network extracts image features. The neck network fuses the feature maps of different scales extracted by the backbone network. The detection head detects the fused feature maps.

[0059] S4, load the training set and validation set images of the water surface floating target dataset into the improved YOLOv6 network model for training and validation, and save the model with the best effect in the validation set;

[0060] S5: Use the saved result model to test the images in the test set, obtain the test results, and complete the recognition of floating targets on the water surface.

[0061] The Detail Information Enhancement Module (DIEM) is designed in the neck network to enhance the network’s extraction of small target detail features. The improved YOLOv6 network model in this embodiment is improved on the basis of YOLOv6-L. i In and C i-1 ,C i-2 Fusion of Ci-2 Performing a convolution operation with k=1 followed by downsampling is a simple operation that cannot extract sufficient texture features of small objects in the underlying feature map, nor can it obtain detailed information about small objects, which is not conducive to the model's recognition and positioning of small objects. Therefore, this embodiment proposes a detail information enhancement module.

[0062] In one embodiment, Figure 3 As shown in Figure 2, the detail information enhancement module combines dilated convolution and efficient channel attention module (ECA). Dilated convolution has a larger receptive field. The parallel combination of dilated convolution with different expansion rates allows the model to better capture the detailed features of small targets. The ECA module can retain important feature information by adaptively suppressing the unimportant feature information extracted by dilated convolution. The structure of the detail information enhancement module is shown in Figure 2. Figure 3 Specifically, in order to extract the input feature map C i The detailed features of the image are calculated in parallel using dilated convolutions with different dilation rates, where the dilation rate r is set to 1, 3, and 5. This dilation rate is a hyperparameter, and its setting is related to the dataset. Next, the extracted feature maps are fused in the channel dimension to generate a new feature map. This new feature map is then input into the ECA module to generate the channel attention weight value of this feature map, and the weight value is multiplied by each channel of the feature map. Finally, the feature map is sequentially split and element-wise summed according to the channel dimension, and then added pixel by pixel to generate the final feature map D i .

[0063] The formula is as follows:

[0064] D i =ψ(F cat ⊙α), (1)

[0065] Among them, F cat is the feature merged after the dilated convolution, α is F cat The channel attention weight generated after inputting the ECA module, ⊙ represents the vector dot product, ψ(·) represents the function of sequential splitting and element-by-element summation, D i It is the feature map after detail information enhancement.

[0066] α=σ(W(P gavg (F cat ))), (2)

[0067] Among them, W∈R is the one-dimensional convolution weight with a step size of k, P gavg (·) is the global average pooling function, and σ(·) is the Sigmoid function.

[0068]

[0069] Where |t| odd represents the odd number closest to the absolute value of t, C is the number of feature map channels, and γ is set to 2 and b = 1.

[0070] In the detection scenario of floating targets on the water surface, due to the influence of weather conditions, the captured targets are often easily interfered by factors such as reflections, glare and water waves, which causes the model to mistakenly detect the background as the target, which seriously affects the detection accuracy of the model. In addition, small targets occupy a major part in this scenario. These small targets occupy a small proportion of pixels in the image, which may be only a few pixels or less. They are similar to the visual features such as color and texture of the background, and the features are not obvious enough, making the model more easily confused. Therefore, it is crucial to highlight the target features by suppressing the interference of noise to improve the detection performance of floating garbage on the water surface. Therefore, this embodiment designs an adaptive noise suppression module (ANSM) in the neck network. By suppressing background noise, the difference between foreground and background features is increased to improve the model's feature recognition ability for the target, and then used to suppress the interference of noise on the feature map on the target.

[0071] In one embodiment, the structure of the adaptive noise suppression module is as follows: Figure 4 As shown, the local average pooling method is first used to pool the feature map G i Downsampling is performed, and then the downsampled feature map is feature mapped, and then the mapped feature map is subjected to the Sigmid function to generate the spatial attention weight feature map, and finally upsampling is performed to restore it to the feature map G i Size and multiply it into the feature map G i The adaptive noise suppression module can adaptively adjust the pixel weights at different positions on the feature map, thereby suppressing background noise features and highlighting foreground target features with rich details. The formula is expressed as follows:

[0072] P i =G i ⊙S i , (4)

[0073] Among them, G i is the input feature map, S i is the spatial weight of the feature map, and ⊙ represents the vector dot product. i The formula is:

[0074] S i =U(σ(W2·δ(BN(W1·P lavg (G i ))))), (5)

[0075] Among them, U(·) represents the upsampling function, σ(·) is the Sigmoid function, δ(·) is the Relu function, W1, W2∈R are the convolution weights with a step size of 1 and a convolution kernel size of 1, BN(·) is the BatchNorm2d function, P lavg (·) represents the local average pooling function, and Gi is the input feature map.

[0076] During training, a loss function based on the standardized Wasserstein distance metric is used to better measure the similarity between the predicted and target boxes of small objects, leading to better model convergence. Specifically, the loss function based on the standardized Wasserstein distance metric first models the ground truth and predicted boxes as Gaussian distributions. The similarity between these two distributions is then measured using the standardized Wasserstein distance instead of the Intersection over Union (IoU). This metric allows for similarity to be measured even when the two boxes have no overlap or very little overlap. Furthermore, NWD is insensitive to target scale and is more stable for small objects.

[0077] For a horizontal bounding box B = (cx, cy, w, h), it can be modeled as a 2D Gaussian distribution as follows:

[0078]

[0079] Where μ is the mean and ∑ is the variance. In this way, the similarity between two bounding boxes can be expressed as the distance between the Gaussian distributions of the boxes. The distance between two distributions is calculated using the Wasserstein distance. For two 2D Gaussian distributions, m1 = N(μ1,∑1) and m2 = N(μ2,∑2), their second-order Wasserstein distance can be defined as:

[0080]

[0081] in Frobenius norm. For two bounding boxes A = (cx a ,cy a ,w a ,h a ) and B=(cx b ,cy b ,w b ,h b ), formula (7) can be further simplified as follows:

[0082]

[0083] However, W2 2 (N a,N b ) is a distance metric and cannot be used directly as a similarity metric (i.e., a value between 0 and 1 as IoU). Therefore, we normalize it using its exponential form to obtain a new metric called normalized Wasserstein distance (NWD):

[0084]

[0085] Where C is a constant that is closely related to the dataset. The bounding box loss function based on the NWD indicator is designed as follows:

[0086] L NWD =1-NWD(N p ,N g ), (10)

[0087] where N p is the Gaussian distribution model of the prediction box, N g is the Gaussian distribution model of the true value box. We use the NWD-based loss function as the bounding box loss function of our model, and the total loss function is as follows:

[0088] L sum =L cls +λL NWD +βL obj , (11)

[0089] Among them L cls and L obj are the classification loss and the target loss. λ and β are the bounding box loss weight and the target loss weight.

[0090] The following is a detailed introduction to the experiments and verification data of the method for detecting small floating targets on the water surface based on the improved YOLOv6 network model of this embodiment.

[0091] 1. Dataset

[0092] The experiment used two datasets, the extended dataset Flow+ and the public dataset Pascal VOC2007.

[0093] The Flow+ dataset is based on the Flow dataset, with some images re-annotated and supplemented with our own dataset of floating debris. Analysis of the Flow dataset revealed that its sample categories were too narrow, consisting solely of bottles. Furthermore, we discovered that some images in the dataset were incorrectly labeled, such as cans and paper boxes being mistakenly labeled as bottles. This is unreasonable because cans and paper boxes have distinct features and textures that differ from bottles. Such mislabeling can cause the model to confuse the definitions of different categories, affecting detection accuracy.

[0094] Therefore, we corrected the incorrect annotations of the Flow dataset and collected and annotated the floating garbage data in inland waters. The re-made dataset is named Flow+, which includes a total of 3060 pictures, including 5 categories: bottles, paper boxes, cans, foam and plastic bags. 2000 of them are from the Flow dataset, and the other 1060 are self-made datasets. We divide the dataset into training set, validation set and test set in a ratio of 6:2:2, where the training set contains 1836 pictures, and the validation set and test set include 612 pictures respectively. The target pixel area in the dataset is less than 32 2 A small target, less than 96 2 For large targets, greater than 32 2 and less than 96 2 The number of medium-sized objects is large, with small objects comprising the largest proportion, totaling 3,823. The dataset has a long tail, so we selected 649 images from the training set and performed random augmentation, mixed augmentation, and mosaic augmentation on 1,298 images each. This augmentation achieved a more balanced training set compared to the previous one and also addressed the issue of insufficient training sets.

[0095] Voc2007 is a classic dataset for object detection. It contains 9,963 images, including 5,011 in the training and validation sets and 4,952 in the test set. The dataset covers 20 categories: airplane, bicycle, bird, boat, bottle, bus, car, cat, chair, cow, dining table, dog, horse, motorbike, person, potted plant, sheep, sofa, train, and TV monitor.

[0096] 2. Experimental details and evaluation indicators

[0097] The experiments were conducted using Python 3.7 and Pytorch 1.12.0+cu116, and the GPU used for training was an NVIDIA GeForce RTX 3090. The comparative experiments and ablation experiments between DENR-YOLOv6 and YOLOv6 were conducted with the same basic hyperparameter settings. Training used the SGD optimizer, with an initial learning rate of 0.01, a Cosine learning rate strategy, a momentum decay of 0.0005, and a momentum of 0.937. The batch size was set to 16, the number of epochs was set to 100, and image HSV color gamut enhancement, random affine transformation enhancement, image blending enhancement, and image mosaic enhancement were used during training. The hyperparameters in formula (9) were set to C = 1.6, and the hyperparameters in formula (11) were set to λ = 2.5 and β = 0.5. The hyperparameters of the other comparison models were the officially initialized hyperparameters of the models and were not modified.

[0098] Each image in an object detection task may contain objects of different categories, and the model's object classification and localization performance needs to be evaluated. In object detection, mean average precision (mAP) is the primary metric. mAP is the average of the average precision (AP) values ​​across multiple classification tasks, while average precision (AP) is the area under the precision-recall (PR) curve. Using precision as the horizontal axis and recall as the vertical axis, we can obtain a PR curve. The formulas for calculating precision and recall are as follows:

[0099]

[0100] Among them, TP means IoU>IoU threshold The number of detection boxes, FP means IoU≤IoU threshold FN represents the number of detection boxes that are not detected.

[0101] In this example, we use the MS COCO evaluation standard. 100 points are sampled on the PR curve for calculation. The mAP value is calculated every 0.05 with the IoU threshold ranging from 0.5 to 0.95, and the average of all results is taken as the final mAP result. S For pixel area less than 32 2 The mAP value of the target box, mAP M For pixel area greater than 32 2 and less than 96 2 The mAP value of the target box, mAP L For pixel area greater than 96 2 The mAP value of the target box.

[0102] 3. Comparative experiment

[0103] The improved YOLOv6 network model in this embodiment is evaluated on the Flow+-test along with several other one-stage detection algorithms, including YOLOX-L, YOLOv5-L, YOLOv6-L, and YOLOv7, the two-stage classic algorithm Faster-RCNN, and the Transformer-based detection algorithm Deformable-DETR. The experimental results are shown in Table 1.

[0104] Table 1 Experimental results of different detection algorithms in Flow+-test

[0105]

[0106] As can be seen from Table 1, compared with other comparison models, the improved YOLOv6 has better mAP and mAP S and mAP L The best performance was achieved in terms of indicators, reaching 54.2%, 40.6% and 79.1% accuracy respectively. M The performance of the improved YOLOv6 is slightly lower than that of the Faster-RCNN model (0.3%), which is the best in this metric, but is within an acceptable range. These results show that the improved YOLOv6 improves the detection accuracy of small and large objects while maintaining the average detection accuracy and medium target detection accuracy.

[0107] We draw a PR curve for the test results of the improved YOLOv6 on the Flow+-test, as shown in Figure 5 The area under the curve for each category is the AP value for that category. A higher AP value indicates better detection performance for that category. As can be seen from the figure, the improved YOLOv6's AP values ​​for bottles, paper boxes, and cans are slightly lower than the average AP value for all categories, while its AP values ​​for foam and plastic bags are higher than the average AP value for all categories.

[0108] Figure 6 To improve the detection results of some representative images of the YOLOv6 network model on Flow+-test.

[0109] From the detection results, it can be seen that the targets in the first column of pictures can be detected even when they are affected by lighting and the feature information is blurred; the targets in the second column of pictures are affected by reflections, but can still be accurately located by DENR-YOLOv6; the targets in the third and fourth columns of pictures are respectively disturbed by water waves and aquatic plants, and the improved YOLOv6 network model can also accurately identify and locate the targets; the fifth column of pictures shows some extremely small targets, and the improved YOLOv6 network model can also detect and identify these targets well.

[0110] In order to further verify the versatility of the improved YOLOv6 network model in this embodiment, we also conducted a comparative experiment on the Voc2007 dataset. The experimental results are shown in Table 2. Compared with other comparison models, the improved YOLOv6 network model (without NWD loss) has a higher mAP S The accuracy of the indicators reached 11.4%, achieving the highest accuracy in small target detection, and the mAP, mAP M and mAP L The indicators are also higher than other comparison models, reaching accuracies of 31.6%, 20.9% and 35.6% respectively.

[0111] Table 2 Experimental results of different detection algorithms in Voc2007-test

[0112]

[0113] In summary, the improved YOLOv6 network model in this embodiment achieved the best results compared to other comparison models. By combining dilated convolutions with an efficient channel attention mechanism, it enhances feature extraction from shallow feature maps in the backbone network. By using an adaptive noise suppression module to suppress noise in feature maps before input to the detection head, the accuracy of small object detection can be significantly improved.

[0114] With the above-described preferred embodiments of the present invention as a guide, and with reference to the above description, relevant personnel are fully capable of making various changes and modifications without departing from the technical scope of this invention. The technical scope of this invention is not limited to the contents of the specification and must be determined according to the scope of the claims.

Claims

1. A method for detecting small floating targets on the water surface based on an improved YOLOv6 network model, characterized in that: include: The pre-trained improved YOLOv6 network model is used to detect the floating target image on the water surface and output the small target detection results on the water surface; The improved YOLOv6 network model uses a detail information enhancement module to improve the ability to extract detail information of small objects, an adaptive noise suppression module to suppress noise, and a standardized Gaussian Wasserstein distance to calculate the regression loss in model training; The working process of the detail information enhancement module includes: The dilated convolution with different expansion rates is used to extract the detail feature maps on the low-level feature maps in parallel, and the extracted detail feature maps are fused in the feature dimension. cat Input channel attention module, generate feature map channel attention weight α; add channel attention weight α to feature map F cat Multiply, split the multiplied feature map sequentially in the channel dimension according to the ratio before fusion and sum them up. The formula is as follows: D i =ψ(F cat ⊙a), (1) Among them, ⊙ represents the vector dot product, ψ(·) represents the function of sequential splitting and element-wise summation, and D i Feature maps generated for the detail information enhancement module; α=σ(W(P gavg (F cat ))), (2) Among them, W∈R is the one-dimensional convolution weight with a step size of k, P gavg (·) is the global average pooling function, σ(·) is the Sigmoid function; Where |t| odd represents the odd number closest to the absolute value of t, C is the number of feature map channels, γ = 2, b = 1; The working process of the adaptive noise suppression module includes: The feature map G is pooled using local average pooling. i Downsampling is performed, and then the downsampled feature map is feature mapped, and then the feature map after feature mapping is passed through the Sigmid function to generate a spatial attention weight feature map, and finally upsampling is performed to restore it to the feature map G i The spatial size is multiplied into the feature map G i , the formula is as follows: P i =G i ⊙S i , (4) Among them, ⊙ represents the vector dot product, S i The formula is: S i =U(σ(W2·δ(BN(W1·P lavg (G i )))))), (5) Among them, U(·) represents the upsampling function, σ(·) is the Sigmoid function, δ(·) is the Relu function, W1, W2∈R are the convolution weights with a step size of 1 and a convolution kernel size of 1, BN(·) is the BatchNorm2d function, P lavg (·) represents the local average pooling function.

2. The method for detecting small floating targets on the water surface based on the improved YOLOv6 network model according to claim 1, characterized in that: The specific steps for calculating the regression loss in model training using the standardized Gaussian Wasserstein distance include: First, the Bounding Boxes are modeled as a Gaussian distribution, and then the standardized Wasserstein distance is used to measure the similarity between the two distributions; For a horizontal bounding box B = (cx, cy, w, h), it is modeled as a 2D Gaussian distribution as follows: Where μ is the mean, ∑ is the variance, and the similarity between two bounding boxes is represented by the distance between the Gaussian distributions of the boxes; For two 2D Gaussian distributions, m1 = N(μ1,∑1) and m2 = N(μ2,∑2), their second-order Wasserstein distance can be defined as: in Frobenius norm, for two bounding boxes A = (cx a ,cy a ,w a ,h a ) and B=(cx b ,cy b ,w b ,h b ), formula (7) is simplified as follows: Is a distance metric and cannot be used directly as a similarity metric. Therefore, it is normalized using its exponential form to obtain a new metric called normalized Wasserstein distance (NWD): Where C is a constant that is closely related to the dataset. The bounding box loss function based on the NWD indicator is designed as follows: L NWD =1-NWD(N p ,OF g ), (10) where N p is the Gaussian distribution model of the prediction box, N g It is the Gaussian distribution model of the true value box. The loss function based on NWD is used as the bounding box loss function of our model. The total loss function of the model is as follows: L sum =L cls +λL NWD +βL obj , (11) Among them L cls and L obj are the classification loss and target loss, λ and β are the bounding box loss weight and target loss weight.

3. The method for detecting small floating targets on the water surface based on the improved YOLOv6 network model according to claim 1, characterized in that: The method also includes: before using the training set to train the improved YOLOv6 network model, performing data enhancement processing on the surface floating target image in the training set.

Images