A deep learning-based underwater small biological target detection method and system

Abstract

The present application relates to computer vision and underwater detection technology, provide a kind of underwater small biological target detection method and system based on deep learning, the WLDA-YOLOv8 deep learning model constructed includes main network, neck network and detection head, main and neck embed light multi-scale module, detection head introduces dynamic detection head and adaptive sample selection strategy;Training stage uses improved WIoUv3s loss to carry out boundary box regression, the model is trained using underwater biological data set, finally realizes the accurate detection of small target in underwater image, enhances the capture and fusion ability to small target subtle features.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer vision and underwater detection technology, and particularly relates to an underwater small biological target detection method and system based on deep learning. BACKGROUND

[0002] At present, underwater small biological detection mainly adopts a target detection algorithm based on deep learning, wherein YOLOv8 becomes one of the commonly used schemes due to its balance between speed and accuracy, mainly includes a backbone, a neck and a detection head, and adopts a C2f module, an FPN+PAN structure and a decoupling head; when YOLOv8 is directly applied to an underwater environment, the following significant defects exist:

[0003] Weak small target feature extraction capability: the original C2f module has limited perception ability for multi-scale features, underwater small targets (such as shrimps and small fish) have a small pixel ratio, and subtle features are easily lost in deep network, resulting in a high missing detection rate;

[0004] Rigid sample allocation strategy: the original model adopts a fixed threshold label allocation strategy, which is difficult to adapt to the characteristics of variable underwater target size and dense distribution, resulting in unbalanced positive and negative sample allocation, affecting model convergence and accuracy;

[0005] Insufficient boundary box regression accuracy: the CIoU loss function lacks a distinguishing mechanism for difficult and easy samples and angle sensitivity, and has a large positioning deviation for underwater targets with special shapes and blurred edges;

[0006] High model calculation complexity: the original model still has a large parameter quantity and calculation quantity (GFLOPs), which is not conducive to deployment in underwater embedded devices with limited computing resources. SUMMARY

[0007] The present application aims to solve the above problems, and designs an underwater small biological target detection method and system based on deep learning.

[0008] The present application provides an underwater small biological target detection method based on deep learning in a first aspect, which comprises the following steps:

[0009] A WLDA-YOLOv8 deep learning model is constructed, wherein the WLDA-YOLOv8 deep learning model includes a backbone, a neck and a detection head, a lightweight and efficient multi-scale module is adopted in the backbone and the neck, a dynamic detection head module is introduced in the detection head, and the dynamic detection head module works cooperatively with a dynamic adaptive training sample selection strategy;

[0010] In the training stage of the WLDA-YOLOv8 deep learning model, an improved WIoUv3s loss function is used for boundary box regression;

[0011] The WLDA-YOLOv8 deep learning model is trained by using the underwater small biological data set, and a trained detection model is obtained.

[0012] The underwater image to be detected is input into the trained detection model, and a detection result of the underwater small biological is output.

[0013] Optionally, in the first implementation manner of the first aspect, the working process of the lightweight and efficient multi-scale module includes:

[0014] The input feature map is evenly divided into four parts in the channel dimension and is subjected to multi-branch convolution processing, wherein the four parts are Part1, Part2, Part3 and Part4 respectively.

[0015] The processed four part features are spliced in the channel dimension, the spliced features are subjected to channel attention mechanism, the channel weight is generated through global average pooling, a fully connected layer and a Sigmoid activation function, the channel number is adjusted through 1x1 convolution, and the feature map is output.

[0016] Optionally, in the second implementation manner of the first aspect, Part1 and Part4 adopt identity mapping, Part2 adopts 3x3 depth separable convolution, and Part3 adopts 5x5 depth separable convolution.

[0017] Optionally, in the third implementation manner of the first aspect, the implementation process of the dynamic adaptive training sample selection strategy is:

[0018] The IoU value of each real target frame and all candidate anchor frames is calculated, the sum of the mean value and the standard deviation of the IoU value is calculated, and is used as the dynamic IoU threshold of the real target frame.

[0019] The candidate anchor frame with the IoU value greater than the dynamic IoU threshold is listed as a positive sample, and the rest is listed as a negative sample.

[0020] Optionally, in the fourth implementation manner of the first aspect, the dynamic detection head module is composed of a plurality of serial DyHead Blocks, and each DyHead Block sequentially performs scale perception attention, spatial perception attention and channel perception attention.

[0021] Optionally, in the fifth implementation manner of the first aspect, the scale perception attention learns different scale weights through global average pooling and a fully connected layer, the spatial perception attention focuses on the spatial key area of the feature map through deformable convolution, and the channel perception attention dynamically calibrates the channel dimension feature response through the lightweight channel attention module.

[0022] Optionally, in a sixth implementation of the first aspect of the present invention, the calculation formula for the improved WIoUv3s loss function is as follows:

[0023] ;

[0024] ;

[0025] ;

[0026] ;

[0027]

[0028] in w , h These represent the width and height of the prediction box, respectively. , These represent the width and height of the actual bounding box, respectively. , The width and height of the smallest bounding rectangle that simultaneously contains both the predicted bounding box and the ground truth bounding box. These represent the center coordinates of the predicted bounding box and the ground truth bounding box, respectively. , This indicates the calculation of the Euclidean distance between two points. This indicator measures the quality of the anchor frame and increases as IoU decreases; and It is an adjustable hyperparameter; Indicates the results obtained during the current training. The optimal value of loss, This represents the actual IoU loss value obtained during the current training. This represents the basis of the WIoU loss function. This represents the final loss value. The function p() is the Euclidean distance function, and p² mentioned in the formula is the square of the distance.

[0029] A second aspect of the present invention provides a deep learning-based underwater small biological target detection system, the system comprising:

[0030] A building module is used to build a WLDA-YOLOv8 deep learning model, wherein the WLDA-YOLOv8 deep learning model includes a backbone network, a neck network, and a detection head. The backbone network and the neck network employ lightweight and efficient multi-scale modules, and the detection head introduces a dynamic detection head module that works in conjunction with a dynamic adaptive training sample selection strategy.

[0031] The bounding box regression module is used to perform bounding box regression using the improved WIoUv3s loss function during the training phase of the WLDA-YOLOv8 deep learning model.

[0032] The training module is used to train the WLDA-YOLOv8 deep learning model using a small underwater biological dataset to obtain a trained detection model.

[0033] The output module is used to input the underwater image to be detected into the trained detection model and output the detection results of small underwater organisms.

[0034] A third aspect of the present invention provides a deep learning-based underwater small biological target detection device, the deep learning-based underwater small biological target detection device comprising a memory and at least one processor, the memory storing instructions; the at least one processor invokes the instructions in the memory to cause the deep learning-based underwater small biological target detection device to perform the various steps of the deep learning-based underwater small biological target detection method as described in any of the preceding claims.

[0035] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of the deep learning-based underwater small biological target detection method as described in any of the preceding claims.

[0036] The technical solution provided by this invention replaces the original C2f module with a lightweight and efficient multi-scale module (LEM). The LEM module uses multi-branch convolutional kernels of different sizes to extract features in parallel and introduces a channel attention mechanism. This significantly reduces the number of parameters and computational complexity while enhancing the ability to capture and fuse subtle features of small underwater targets. The number of parameters in the model (WLDA-YOLOv8) of this invention is reduced from 11.1M in the original YOLOv8s to 7.9M, and the computational cost (GFLOPs) is reduced to 15.9. This effectively overcomes the shortcomings of insufficient feature extraction of small targets, high false negative rate and high model complexity in existing technologies. Attached Figure Description

[0037] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0038] Figure 1 This is a diagram of the overall network structure of WLDA-YOLOv8 provided in an embodiment of the present invention;

[0039] Figure 2A LEMConv2d operation structure diagram provided for an embodiment of the present application is provided;

[0040] Figure 3 A LEM module overall structure diagram provided for an embodiment of the present application is provided;

[0041] Figure 4 A DyHead structure diagram provided for an embodiment of the present application is provided;

[0042] Figure 5 A loss function parameter schematic diagram provided for an embodiment of the present application is provided;

[0043] Figure 6 An F1 curve and a recall rate curve provided for an embodiment of the present application are provided;

[0044] Figure 7 A first WLDA-YOLOv8 target category detection effect diagram provided for an embodiment of the present application is provided;

[0045] Figure 8 A second WLDA-YOLOv8 target category detection effect diagram provided for an embodiment of the present application is provided;

[0046] Figure 9 A third WLDA-YOLOv8 target category detection effect diagram provided for an embodiment of the present application is provided;

[0047] Figure 10 A detection result comparison diagram of different methods provided for an embodiment of the present application is provided;

[0048] Figure 11 A structure schematic diagram of a deep learning-based underwater small biological target detection system provided for an embodiment of the present application is provided;

[0049] Figure 12 A structure schematic diagram of a deep learning-based underwater small biological target detection device provided for an embodiment of the present application is provided. DETAILED DESCRIPTION

[0050] The terms "first", "second", "third", "fourth" and the like in the description, claims, and above drawings of the present application (if any) are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order. It is to be understood that the use of the terms so termed herein is interchangeable under appropriate circumstances such that the embodiments described herein are, for example, capable of orderly implementation in sequences other than those illustrated or other than those described herein. Moreover, the terms "comprise" or "have" and any variations thereof are intended to cover a non-exclusive inclusion, for example, a process, method, article, or apparatus that comprises a list of steps or units not necessarily limited to those specifically listed, but can include other steps or units not expressly listed or inherent to such process, method, article, or apparatus.

[0051] For the convenience of understanding, the specific process of the embodiments of the present application is described below, and the present application provides a waterborne small biological target detection method based on improved YOLOv8. The core of the method is to construct a deep learning model named WLDA-YOLOv8 (Waterborne Lightweight Detection with Dynamic Attention YOLOv8). The model works collaboratively through three core improvement modules, which significantly improves the detection performance of underwater small targets. Figure 1 The overall network architecture of the WLDA-YOLOv8 deep learning model of the present application is shown.

[0052] As shown in Figure 1 , the WLDA-YOLOv8 deep learning model of the present application also consists of three parts: backbone network, neck network and detection head. The innovation lies in the replacement and enhancement of the following parts of the original YOLOv8s model:

[0053] In the backbone network and the neck network, the original C2f module is replaced by the lightweight and efficient multi-scale module (LEM) designed by the present application.

[0054] In the detection head part, the dynamic head (DyHead) module is introduced and works collaboratively with the adaptive training sample selection (ATSS) strategy.

[0055] In the boundary box regression task in the model training stage, the improved design of the WIoU v3s (Wise Intersection over Union version 3 star) loss function is used to replace the original CIoU loss function.

[0056] The specific structure, function and coordination of each improved part are described in detail as follows:

[0057] 1.1 Lightweight Efficient Multi-scale Module (LEM, Lightweight Efficient Multi-scale module)

[0058] Function: This module aims to replace the original C2f module with lower computational cost (parameter quantity and FLOPs), enhance the network's ability to extract and fuse multi-scale features, especially the features of small targets, and solve the problem of easy loss of underwater small target features.

[0059] Specific structure and coordination: The core of the LEM module is the LEMConv2d operation inside it. Its structure is as shown in Figure 2 andFigure 3 As shown.

[0060] Input feature map division: LEMConv2d first divides the input feature map into 4 parts (Part1, Part2, Part3, Part4) uniformly in the channel dimension.

[0061] Multi-branch convolution processing:

[0062] Branch one (Part1): Identity, preserving the original feature information.

[0063] Branch two (Part2): Perform 3x3 depthwise separable convolution (DWConv, Depthwise SeparableConvolution) operation, extract medium receptive field features.

[0064] Branch three (Part3): Perform 5x5 depthwise separable convolution (DWConv) operation, extract larger receptive field features.

[0065] Branch four (Part4): Identity.

[0066] Feature fusion: Concatenate the features processed by the four branches in the channel dimension.

[0067] Channel attention optimization: Apply channel attention mechanism (SE Attention) to the concatenated features, i.e. generate channel weights through global average pooling (GAP), fully connected layer (FC) and Sigmoid activation function, enhance important feature channels and suppress redundant channels.

[0068] Output: Finally adjust the number of channels through a layer of 1x1 convolution (Conv) to form the output feature map.

[0069] Through the above design, the LEM module uses different sizes of convolution kernels to extract multi-scale features in parallel, and dynamically optimizes feature weights through attention mechanism, which significantly reduces the parameter amount while realizing more rich feature representation.

[0070] 1.2 Dynamic adaptive training sample selection (Dynamic ATSS) and dynamic detection head (DyHead)

[0071] Function: ATSS dynamically calculates adaptive IoU thresholds for each ground truth (GT), enabling more intelligent positive and negative sample assignment and addressing the sample assignment imbalance caused by the varying scales of underwater targets. The DyHead module enhances the detection head's ability to express features of different forms and sizes by fusing scale, spatial, and channel attention mechanisms. The combination of the two improves the model's convergence speed and detection accuracy.

[0072] Specific structure and cooperation:

[0073] Dynamic ATSS: During training, for each ground truth box, calculate the IoU value with all candidate anchor boxes, and calculate the sum of the mean and standard deviation of these IoU values as the dynamic threshold for the GT. Only candidate boxes with an IoU greater than this threshold are considered positive samples. This process is independent for each target, allowing it to adapt to targets of different sizes and densities.

[0074] DyHead: Its structure is shown in Figure 4 It is a plug-and-play module embedded before the detection head. It is composed of multiple serial DyHead Blocks, each of which performs the following functions in sequence:

[0075] Scale-aware Attention: Learn different scale weights through global average pooling and fully connected layers to enhance the fusion of multi-scale features.

[0076] Spatial-aware Attention: Use deformable convolution (DeformableConvolution) mechanism to allow the convolution kernel to adaptively focus on the spatial key areas of the feature map.

[0077] Channel-aware Attention: Use a lightweight channel attention module (similar to SEBlock but more efficient) to dynamically calibrate the feature response in the channel dimension.

[0078] 1.3 Improved bounding box regression loss function (WIoU v3s)

[0079] Function: The WIoU v3s loss function replaces the original model's CIoU Loss by introducing a dynamic non-monotonic focusing mechanism and an angle penalty factor, aiming to more effectively evaluate anchor box quality and optimize loss distribution strategies, thereby significantly improving the model's bounding box regression accuracy, especially for small targets and complex-shaped targets.

[0080] The calculation of WIoU v3s is based on the following formula:

[0081] ;

[0082] ;

[0083] ;

[0084] ;

[0085]

[0086] where w , h are the width and height of the prediction box respectively, , are the width and height of the ground truth box respectively, , are the width and height of the minimum bounding box containing both the prediction and ground truth box, are the center coordinates of the prediction and ground truth box respectively, , denotes the Euclidean distance between two points, denotes the indicator measuring the quality of anchor box, which increases as IoU decreases; and are adjustable hyperparameters; denotes the optimal value of the loss obtained by the current training, denotes the actual IoU loss value obtained by the current training. By constructing a non-monotonic focusing factor based on and applying it to the WIoU v3s, the model can dynamically balance the loss weight distribution of high and low quality samples—reduce the gradient interference of high quality samples, while appropriately improving the attention to low quality samples; denotes the basis of the WIoU loss function, which considers the IoU loss of the center point distance and the aspect ratio. The purpose is to obtain more stable and effective gradient than the traditional IoU loss, so as to provide better regression effect, denotes the final loss value, which reweights the basic loss by dynamic weight , and the function p() is the Euclidean distance function, and the p² mentioned in the formula is the square of the distance.

[0087] Dynamic non-monotonic focusing mechanism: WIoU v3s draws on the idea of Focal-EIoU, using a dynamic non-monotonic gradient focusing factor to adaptively adjust the contribution of different quality anchor boxes to the loss. For anchor boxes with accurate predictions and high IoU, a smaller gradient gain is assigned to reduce their weight in the total loss; for anchor boxes of difficult samples with low IoU, a larger gradient gain is assigned to encourage the model to pay more attention to these error-prone samples.

[0088] Angle penalty term (inherited from SIoU): The angle between the predicted box and the center line of the true box is considered in the loss calculation, which reduces the degree of freedom of regression, speeds up the convergence speed, and improves the positioning accuracy. The related parameter diagram is shown in Figure 5 .

[0089] 2. Alternative:

[0090] To reflect the wide applicability of the present application, the following lists alternative options for each improvement point:

[0091] Alternative of LEM module:

[0092] The multi-branch structure can use group convolution (Group Convolution) or shuffle operation (Shuffle Operation) to further reduce the computational load.

[0093] The channel attention mechanism (SE) can be replaced by ECA-Net (Efficient Channel Attention), CBAM (Convolutional Block Attention Module) and other attention mechanisms.

[0094] The 3x3 and 5x5 convolution can use dilated convolution (Dilated Convolution) to further expand the receptive field without increasing the parameter amount.

[0095] Alternative of sample assignment and detection head:

[0096] The ATSS strategy can be replaced by other dynamic sample assignment strategies such as OTA (Optimal Transport Assignment) or SimOTA.

[0097] The DyHead module can be replaced by other advanced detection head structures, such as variants of Decoupled Head, or Transformer-based detection heads (such as the Head in DETR).

[0098] Alternative of loss function:

[0099] WIoU v3s can be replaced by other loss function variants that consider sample quality, such as Focal-EIoU, Alpha-IoU, or its earlier versions, WLIOU v1, WLIOU v2. The core is to replace the static loss function with a regression loss with a dynamic focusing mechanism.

[0100] Table 1 Comparison of detection results of different methods on the UMDD (Underwater Marine Debris Detection Dataset) dataset

[0101]

[0102] To further explore the contribution of each improved module to the performance of the WLDA-YOLOv8 deep learning model, an ablation experiment was conducted on the UMDD dataset with YOLOv8s as the benchmark. The performance improvement of each module is shown in Table 2.

[0103] After introducing the WLIOU v3 module, the boundary box regression accuracy of the model is significantly improved. WLIOU v3 focuses on optimizing the regression loss of the boundary box, especially for complex shapes or fuzzy boundaries. For example, for jellyfish, which is difficult to detect, mAP@50:95 is improved from 47.7% to 51.8%, and the performance on small target classes such as small fish is more obvious, with an increase of 13.1%. The model's mAP@50:95 is improved by 2.4%, which clearly shows that the WLIOU v3 module plays an important role in accurately positioning the boundary box and improving the detection accuracy of small targets.

[0104] By adding the Dynamic ATSS module, the model obtains a more intelligent sample selection mechanism. Dynamic ATSS effectively handles the diversity of complex scenes and target shapes by dynamically adjusting the sample allocation weight. It performs significantly better on fish and starfish classes, with mAP@50:95 increasing to 79.1% and 77.3%, respectively.

[0105] By introducing the LEM module, the model's feature extraction capability is further optimized, especially in small target detection. For example, the mAP@50:95 of the small fish class is improved from 56.5% to 59.4%, demonstrating the effectiveness of the LEM module in capturing small target details and background information. In addition, the LEM module performs very well on other classes such as jellyfish and shrimp, with mAP@50:95 increasing to 53.4% and 79.8%, respectively. The model's mAP@50:95 is improved by 3.5%,

[0106] Figure 7The performance of the WLDA-YOLOv8 deep learning model in target detection in real water body scenes is shown. As can be seen from the figure, all categories of targets can be accurately identified, especially in complex environmental backgrounds and target boundary conditions, showing stable and efficient detection capability. Figure 6 The F1 curve and recall rate curve in the ablation experiment are shown. The addition of each module has different effects on the performance of the model. The Wiou v3 module significantly improves the F1 score, reflecting the advantages of precise positioning. The DynamicATSS module mainly improves the precision of the model in the high confidence area. The LEM module significantly improves the recall rate in the low confidence area, reflecting its ability to capture detailed features.

[0107] Table 2 Ablation experiment results of WLDA-YOLOv8

[0108]

[0109] Figure 8 The comparison of WLDA-YOLOv8 with WSSOD_YOLOv8 and 2PCC_YOLOv8 in mAP@50 and Loss curve is shown. From the mAP change curve, WLDA-YOLOv8 always maintains a faster precision growth throughout the training process, especially in the early stage of training, indicating that the model can quickly learn effective features. After about 25 epochs, the mAP gradually stabilizes, eventually reaching 77.9 %, significantly higher than WSSOD_YOLOv8 and 2PCC_YOLOv8. Before about 25 epochs, the loss value of WLDA-YOLOv8 decreases faster, reaching a lower Loss value in a shorter time, indicating good training stability and faster convergence speed.

[0110] From Figure 9 It can be seen that WLDA-YOLOv8 has significantly higher mAP@50:95 than other models under different delays. This indicates that the model has achieved a better balance between detection accuracy and inference speed. When the delay is 1.0 ms, its mAP@50:95 is close to 50 %, which is much higher than other comparison models. When the delay is increased to 3.0 ms, the mAP of WLDA-YOLOv8 further rises to more than 52 %, still maintaining the highest performance. This trend proves its higher computational efficiency and detection performance. Although models such as DSW-YOLOv8n

[25] and YOLOv5 perform well at low delay, their mAP growth slows down significantly when the delay exceeds 2.5 ms.

[0111] To qualitatively compare the detection results, this section provides some 3-group signal conditions that are poor or complex images with similar target and background colors, including small and overlapping targets, for model performance testing. Blue markers represent missed detections, and red markers represent false detections. Figure 10 The results of the three methods are predicted. WLDA-YOLOv8 performs well in detecting small targets and complex overlapping targets, especially in detecting small underwater organisms such as fish and crabs. WLDA-YOLOv8 significantly reduces missed and false detections. Blue markers represent missed detections, and red markers represent false detections. Compared with 2PCC_YOLOv8 and WSSOD_YOLOv8, WLDA-YOLOv8 better handles complex environments with poor signal conditions or similar target and background colors, reducing the occurrence of missed and false detections.

[0112] In summary, WLDA-YOLOv8 not only improves the detection accuracy of small targets and complex backgrounds but also effectively solves the limitations of 2PCC_YOLOv8 and WSSOD_YOLOv8 in detecting overlapping targets. These experimental results verify the performance advantages of the model in underwater target detection. Whether it is a small target, an overlapping target, or a complex background environment, the improved model has shown excellent detection results, proving its applicability in practical applications.

[0113] In this embodiment, the training stage is based on the constructed underwater small organism dataset, which is preferably the UMDD underwater organism dataset. It contains typical underwater small organism samples such as shrimp, small fish, jellyfish, starfish, and crab. Through data enhancement strategies, sample diversity is improved. Random horizontal / vertical flipping, brightness dynamic adjustment, contrast optimization, and Gaussian noise injection are performed on the original images to simulate underwater light changes, water turbidity, and other complex environmental disturbances. During training, the preprocessed dataset is divided into training and validation sets in an 8:2 ratio. The Adam optimizer is used for parameter iteration, with 100 training rounds and a batch size of 16. The learning rate is dynamically decayed through a cosine annealing strategy. During the training process, the model dynamically calculates the IoU threshold for each real target box to distinguish between positive and negative samples through the DynamicATSS strategy, solving the sample allocation imbalance problem caused by the variable size of underwater targets. With the dynamic focusing mechanism and angle penalty term of the Wiouv3s loss function, the boundary box regression accuracy of low-quality samples is strengthened. At the same time, the LEM modules of the backbone network and the neck network continuously optimize the feature extraction capability through multi-branch convolution and attention mechanism. The DyHead module of the detection head calibrates the feature expression through scale, space, and channel triple attention.

[0114] The underwater image to be detected is preprocessed: the original image is uniformly scaled to 640*640 resolution, converted to RGB format and normalized to the range [0, 1] to match the input requirements of the WLDA-YOLOv8 deep learning model; after the preprocessed image is input into the trained model, the backbone network gradually extracts multi-scale features through multiple cascaded LEM modules, the shallow features retain the edge, texture and other detail information of small underwater organisms, and the deep features capture the semantic information of the target; the neck network fuses the multi-scale features from top to bottom and from bottom to top through the FPN+PAN structure, and strengthens the feature expression of small targets; the fused features are processed by the DyHead module of the detection head, which adapts to targets of different sizes through scale perception attention, focuses on the key areas of the target through spatial perception attention, and suppresses background noise through channel perception attention, and finally outputs the class probability, confidence and bounding box coordinates of the target by the decoupling head; in order to remove redundant detection boxes, non-maximum suppression processing is performed on the output results to retain the bounding box with the highest confidence; the final output detection result includes: the specific category of small underwater organisms, the confidence score of the category and the bounding box coordinates of the target in the original image.

[0115] Referring to Figure 11 The underwater small organism target detection system based on deep learning provided by the embodiment of the application has the advantages that:

[0116] The WLDA-YOLOv8 deep learning model includes a backbone network, a neck network and a detection head, a lightweight and efficient multi-scale module is used in the backbone network and the neck network, a dynamic detection head module is introduced in the detection head, and the dynamic detection head module works cooperatively with a dynamic adaptive training sample selection strategy;

[0117] The boundary box regression module is used for boundary box regression by using an improved WIoUv3s loss function in the training stage of the WLDA-YOLOv8 deep learning model;

[0118] The training module is used for training the WLDA-YOLOv8 deep learning model by using an underwater small organism data set to obtain a trained detection model;

[0119] The output module is used for inputting the underwater image to be detected into the trained detection model, and outputting the detection result of the small underwater organism.

[0120] Figure 12is a structural schematic diagram of a deep learning-based underwater small biological target detection device provided by an embodiment of the present application. The deep learning-based underwater small biological target detection device 300 can have relatively large differences due to different configurations or performances, and can include one or more central processing units (CPUs) 310 (for example, one or more processors) and a memory 320, a storage medium 330 (for example, one or more mass storage devices). The memory 320 and the storage medium 330 can be temporary storage or persistent storage. The program stored in the storage medium 330 can include one or more modules (not shown in the figure), and each module can include a series of instruction operations in the deep learning-based underwater small biological target detection device 300. Further, the processor 310 can be configured to communicate with the storage medium 330 and execute the series of instruction operations in the storage medium 330 on the deep learning-based underwater small biological target detection device 300 to implement the method provided by the above embodiment.

[0121] The deep learning-based underwater small biological target detection device 300 can also include one or more power supplies 340, one or more wired or wireless network interfaces 350, and one or more input / output interfaces 360. Those skilled in the art can understand that the deep learning-based underwater small biological target detection device 300 can also include other components that are not shown in the figure, such as a bus, an antenna, a display, a keyboard, a mouse, and the like. Figure 12 The deep learning-based underwater small biological target detection device structure shown does not constitute a limitation on the computer device provided by the present application, and can include more or fewer components than shown, or combine certain components, or different component arrangements.

[0122] The present application also provides a computer readable storage medium, which can be a non-volatile computer readable storage medium or a volatile computer readable storage medium. The computer readable storage medium has instructions stored therein, and when the instructions are run on a computer, the computer executes the steps of the deep learning-based underwater small biological target detection method provided by the above embodiments.

[0123] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described device or apparatus, unit can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0124] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or say the part that contributes to the prior art or the whole 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 method described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0125] The basic principles, main features and advantages of the present application are shown and described above. Those skilled in the art should understand that the present application is not limited by the above embodiments, and the above embodiments and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based method for detecting small underwater biological targets, characterized in that, The method includes the following steps: A WLDA-YOLOv8 deep learning model is constructed, wherein the WLDA-YOLOv8 deep learning model includes a backbone network, a neck network, and a detection head. The backbone network and the neck network employ lightweight and efficient multi-scale modules, and the detection head introduces a dynamic detection head module that works in conjunction with a dynamic adaptive training sample selection strategy. During the training phase of the WLDA-YOLOv8 deep learning model, the improved WIoUv3s loss function is used for bounding box regression. The WLDA-YOLOv8 deep learning model was trained using an underwater small biological dataset to obtain a trained detection model. The underwater image to be detected is input into the trained detection model, which outputs the detection results of small underwater organisms. The working process of the lightweight, high-efficiency, multi-scale module includes: The input feature map is evenly divided into 4 parts along the channel dimension and multi-branch convolution is performed. The 4 parts are Part1, Part2, Part3 and Part4. The four processed features are concatenated along the channel dimension. A channel attention mechanism is applied to the concatenated features. Channel weights are generated through global average pooling, fully connected layers, and the Sigmoid activation function. The number of channels is adjusted through 1×1 convolution, and the feature map is output. Part 1 and Part 4 use identity mapping, Part 2 uses 3×3 depthwise separable convolution, and Part 3 uses 5×5 depthwise separable convolution. The implementation process of the dynamic adaptive training sample selection strategy is as follows: Calculate the IoU value between each ground truth bounding box and all candidate anchor boxes, calculate the sum of the mean and standard deviation of the IoU values, and use it as the dynamic IoU threshold for the ground truth bounding box; Candidate anchor boxes with IoU values ​​greater than the dynamic IoU threshold are listed as positive samples, and the remaining columns are listed as negative samples. The dynamic detection head module consists of multiple serial DyHead Blocks, each of which sequentially performs scale-aware attention, spatial-aware attention, and channel-aware attention. The scale-aware attention module learns the weight distribution of features at different scales through global average pooling and fully connected layers. The spatial-aware attention module uses deformable convolution to adaptively focus on key spatial regions in the feature map. The channel-aware attention module dynamically calibrates the feature response intensity in the channel dimension through a lightweight channel attention mechanism. The improved WIoUv3s loss function is calculated using the following formula: ; ; ; ； in w , h These represent the width and height of the prediction box, respectively. , These represent the width and height of the actual bounding box, respectively. , This represents the width and height of the smallest bounding rectangle containing the predicted bounding box and the ground truth bounding box. These represent the center coordinates of the predicted bounding box and the ground truth bounding box, respectively. x and y represent the coordinates of the center point. This indicates the calculation of the Euclidean distance between two points. This indicator measures the quality of the anchor frame and increases as IoU decreases; and It is an adjustable hyperparameter; Indicates the results obtained during the current training. The optimal value of loss, This represents the actual IoU loss value obtained during the current training. This represents the basis of the WIoU loss function. This represents the final loss value. The function p() is the Euclidean distance function, and p² mentioned in the formula is the square of the distance.

2. A deep learning-based underwater small biological target detection system, characterized in that, The system includes: The building module is used to construct a WLDA-YOLOv8 deep learning model, which includes a backbone network, a neck network, and a detection head. The backbone and neck networks employ lightweight and efficient multi-scale modules. The detection head incorporates a dynamic detection head module that works in conjunction with a dynamic adaptive training sample selection strategy: the input feature map is uniformly divided into four parts along the channel dimension and subjected to multi-branch convolution processing, where the four parts are Part1, Part2, Part3, and Part4; the processed four feature parts are concatenated along the channel dimension, and a channel attention mechanism is applied to the concatenated features. The system generates channel weights through global average pooling, fully connected layers, and a sigmoid activation function, adjusts the number of channels through 1×1 convolutions, and outputs a feature map. Parts 1 and 4 use identity mapping, Part 2 uses 3×3 depthwise separable convolutions, and Part 3 uses 5×5 depthwise separable convolutions. The IoU value between each ground truth bounding box and all candidate anchor boxes is calculated, and the sum of the mean and standard deviation of the IoU values ​​is used as the dynamic IoU threshold for the ground truth bounding box. Candidate anchor boxes with IoU values ​​greater than the dynamic IoU threshold are listed as positive samples, and the rest are listed as negative samples. The dynamic detection head module consists of multiple serial DyHead Blocks, each of which sequentially executes scale-aware attention, spatial-aware attention, and channel-aware attention. The scale-aware attention module learns the weight distribution of features at different scales through global average pooling and fully connected layers. The spatial-aware attention module uses deformable convolutions to adaptively focus on key spatial regions in the feature map. The channel-aware attention module dynamically calibrates the feature response intensity in the channel dimension through a lightweight channel attention mechanism. The bounding box regression module is used for bounding box regression during the training phase of the WLDA-YOLOv8 deep learning model, employing an improved WIoUv3s loss function. The calculation formula for the improved WIoUv3s loss function is as follows: ; ; ; ； in w , h These represent the width and height of the prediction box, respectively. , These represent the width and height of the actual bounding box, respectively. , This represents the width and height of the smallest bounding rectangle containing the predicted bounding box and the ground truth bounding box. These represent the center coordinates of the predicted bounding box and the ground truth bounding box, respectively. x and y represent the coordinates of the center point. This indicates the calculation of the Euclidean distance between two points. This indicator measures the quality of the anchor frame and increases as IoU decreases; and It is an adjustable hyperparameter; Indicates the results obtained during the current training. The optimal value of loss, This represents the actual IoU loss value obtained during the current training. This represents the basis of the WIoU loss function. This represents the final loss value. The function p() is the Euclidean distance function, and p² mentioned in the formula is the square of the distance. The training module is used to train the WLDA-YOLOv8 deep learning model using a dataset of small underwater organisms to obtain a trained detection model. The output module is used to input the underwater image to be detected into the trained detection model and output the detection results of small underwater organisms.

3. A deep learning-based underwater small biological target detection device, characterized in that, The deep learning-based underwater small biological target detection device includes a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the deep learning-based underwater small biological target detection device to perform the various steps of the deep learning-based underwater small biological target detection method as described in claim 1.

4. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the various steps of the deep learning-based underwater small biological target detection method as described in claim 1.

Images