Underwater low-resolution small-target biological detection method based on improved pyramid

Abstract

The invention discloses an underwater low-resolution small-target biological detection method based on an improved pyramid. The underwater low-resolution small-target biological detection method comprises the following steps: firstly, dividing a public underwater image data set into a training set, a verification set and a test set in proportion; then constructing a GPBS-YOLOv8 model, and replacing a C2f structure with a PPA module in a backbone network to enhance multi-scale features; introducing a GSConv module at the neck to reduce parameter quantity and optimize feature representation; in the feature fusion stage, a BiFPN capable of learning weight is adopted, and a P2 small-scale detection layer is newly added; shape-IoU loss is used to improve shape regression accuracy. And finally, training the model on a training set, and verifying the real-time and high-precision detection performance of the model on a low-resolution small target in a complex underwater environment on a test set.

Claims

1. A method for underwater low-resolution small-target biological detection based on an improved pyramid, characterized by: The following steps are involved: S1: Acquire underwater target image data, and randomly divide the underwater target image data into a training set, a validation set, and a test set with consistent data distribution; S2: Build an improved GPBS-YOLOv8 target detection model, which includes a backbone network, a neck network, and a target detection head. The specific construction method includes the following steps: S2-1: Backbone network optimization: Using the image data of the training set as input, the PPA module is used to reconstruct the Bottleneck layer of the C2f module in YOLOv8 to form an improved C2f_PPA module, which specifically includes: The input feature map is divided into local and global spaces, and the spatial weights of the features are determined by the feedforward network to obtain weighted features; Local and global features are obtained through channel selection, rearrangement and bilinear interpolation, and deep features are obtained by fusing sequential convolution branches; Enhance feature expression using adaptive channels and spatial attention mechanisms; S2-2: Neck network optimization: The output of the backbone network is used as input, and the GSConv module is used to replace the original Conv module of YOLOv8. The feature maps output by standard convolution and depthwise separable convolution are concatenated, and the channel order is rearranged using the shuffle operation to reduce the number of network parameters and optimize feature representation. S2-3: Feature Fusion Network Optimization: Using the neck network output as input, the BiFPN structure replaces the original FPN+PAN structure of YOLOv8. This constructs a feature fusion mechanism with bidirectional cross-scale connections and learnable weights. A new P2 small-scale target detection layer is added to improve the detection capability of small underwater targets. S2-4: Detection head optimization: The output features of the feature fusion network are used as input, and the Shape-IoU loss function is used instead of the CIoU loss function to accurately measure the differences in target shape and scale, improving the detection accuracy of the model under the conditions of target occlusion and morphological changes in underwater environments; S3: Input the training set described in step S1 into the GPBS-YOLOv8 model constructed in step S2 for training to obtain a trained underwater target detection model, and use the test set described in step S1 to evaluate the target detection performance of the GPBS-YOLOv8 model and output the underwater target detection result.

2. The underwater low-resolution small target biological detection method based on improved pyramid according to claim 1, characterized in that: The feedforward network in step S2-1 determines the spatial weight of the feature using the following steps: In local and global branches, the Patch size parameter p is introduced to distinguish different branches; Unfold and Reshape the input feature map Segment into a set of spatially contiguous patches of size (p×p,H′ / p,W′ / p,C′); Perform an averaging operation on each patch along the channel dimension to generate a channel average feature of size ((p×p,H′ / p,W′ / p,C′)); The channel average feature is input into the feedforward network FFN, which is used to perform linear transformation on the channel average feature and generate a probability distribution in the spatial dimension through an activation function; According to this probability distribution, the feature weights of the original patch are adjusted.

3. The underwater low-resolution small target biological detection method based on improved pyramid according to claim 1, characterized in that: The calculation of deep features in step S2-1 is as follows: The following steps are involved: First, the weighted features are sequentially processed through channel selection, Reshape and bilinear interpolation operations to output local features F local ∈R H′×W′×C′ and the global feature F global ∈R H′×W′×C′ ; In the sequential convolution branch, the same weighted features are sequentially input into three consecutive 3×3 convolution operations, and F is calculated in sequence. conv1 ∈R H′×W′×C′ 、F conv2 ∈R H′×W′×C′ and F conv3 ∈R H′×W′×C′ , and the F conv1 、F conv2 and F conv3 Add together to generate the final deep feature F conv ∈R H′×W′×C′ .

4. The underwater low-resolution small target biological detection method based on improved pyramid according to claim 1, characterized in that: Step S2-1 uses the adaptive channel and spatial attention mechanism to enhance feature expression, which specifically includes the following steps: First, F′∈R H′×W′×C′ Input one-dimensional channel attention module M C ∈R 1×1×C′ , the channel attention coefficient is calculated by global pooling and channel-by-channel weight generation operations, and multiplied by F′ according to the channel dimension to enhance the key channel features; then, the channel-enhanced features are input into the two-dimensional spatial attention module M S ∈R H′×W′×1 , a spatial attention map is generated through convolution and activation operations, and multiplied with the channel enhanced features according to the spatial dimension to further highlight the salient spatial areas, thereby obtaining the final enhanced feature output.

5. The underwater low-resolution small target biological detection method based on improved pyramid according to claim 1, characterized in that: In step S2-2, the GSConv module is used to replace the Conv module in the original YOLOv8 network, which specifically includes the following steps: S2-2-1: The feature map with the number of channels C1 output from the backbone network is input into the standard convolution layer, and its channel number is compressed to C2 / 2. The compressed feature map is then input into the depthwise separable convolution layer, which outputs two sets of feature maps. The feature map output by the standard convolution and the feature map output by the depthwise separable convolution are concatenated in the channel dimension, and the channel order is rearranged through the shuffle operation to break the local correlation between channels and integrate important semantic information. The final output is a feature map with the number of channels C2. S2-2-2: When the feature map with the output channel number C2 enters the neck network, the GSConv module is used instead of the ordinary Conv module to process the cascade feature map; S2-2-3: Use the GSConv module only in the neck network, not in the backbone network or detection head, to perform local optimization on neck feature processing.

6. The underwater low-resolution small target biological detection method based on improved pyramid according to claim 5, characterized in that: The GSConv module processes features by including a standard convolution branch and a depth-separable convolution branch: Standard convolution branch: The number of channels is mapped from C1 to C2 / 2 through 3×3 standard convolution. The calculation formula of the output feature is: Among them, H represents the spatial height of the feature map, W represents the spatial width of the feature map, Conv 3×3 represents a 3×3 standard convolution, X std It is the standard convolution output feature; Depthwise separable convolution branch: First, perform depthwise convolution on X, then perform 1×1 pointwise convolution, compressing the number of channels to C2 / 2. The calculation formula for the output feature is: Among them, Depthwise 3×3 Indicates that each channel performs a 3×3 convolution operation on the depth of the convolution, Pointwise indicates a point-by-point convolution operation of size 1×1, X dw Represents the depth-wise separable convolution output features.

7. The underwater low-resolution small target biological detection method based on improved pyramid according to claim 1, characterized in that: In the feature fusion network optimization in step S2-3, the BiFPN structure introduces learnable weight parameters to perform weighted fusion on the input feature layers from each scale, and adaptively allocates the fusion weights of each input feature layer through fast normalization to optimize the cross-scale feature fusion process.

8. The underwater low-resolution small target biological detection method based on improved pyramid according to claim 7, characterized in that: In the BiFPN structure, single input nodes that do not participate in feature fusion are deleted, and only fusion nodes with at least two inputs are retained; and jump connections are added between input nodes and output nodes at the same scale to construct a residual flow path, thereby achieving broader feature interaction and multiple repeated fusion between different scales.

9. The underwater low-resolution small target biological detection method based on improved pyramid according to claim 8, characterized in that: Based on the BiFPN structure, an additional P2 small-scale object detection layer is added, and the channel parameters of all nodes in BiFPN are uniformly set to 256.

10. The underwater low-resolution small target biological detection method based on improved pyramid according to claim 1, characterized in that: The training of the GPBS-YOLOv8 target detection model uses the Shape-IoU loss function instead of the original CIoU loss function. The Shape-IoU loss function is defined as follows: Calculate the IoU between the predicted box and the real box. The specific calculation formula is: Among them, B represents the predicted bounding box, B gt represents the ground-truth bounding box; The shape normalization coefficient is calculated based on the width and height of the real box. gt and height h gt Introducing the same shape scaling parameter scale, the specific calculation formula is: Among them, (w gt ) scale and (h gt ) scale are the width and height of the real frame after power operation; ww and hh are the normalized shape coefficients in the width and height directions respectively; Calculate the shape distance of the center point offset. The specific calculation formula is: Among them, x c ,y c is the center coordinate value of the prediction box, is the center coordinate value of the real box, and c represents the diagonal length of the minimum circumscribed rectangle covering the predicted box and the real box; Calculate the shape inconsistency of width and height offsets. The specific calculation formula is: Among them, |w - w gt | and |h - h gt | are the absolute differences in width and height between the predicted bounding box and the ground truth bounding box, respectively; Calculate the shape inconsistency penalty term. The specific calculation formula is: Among them, W t is the width offset; Summarizing the above items into the final loss function, the specific calculation formula is: L shape-IoU =1-IOU+distance shape +0.5×Ω shape Among them, 1-IOU is the loss of overlap between the predicted box and the real box; distance shape is the center point shape distance, 0.5×Ω shape is a high shape inconsistency penalty.

Images