Method for detecting bird target by using YOLOv5s improved model

Abstract

The invention relates to the technical field of target detection, and particularly provides a method for detecting a bird target by using a YOLOv5s improved model, which comprises the following steps of: performing the following optimization on the YOLOv5s model: replacing two serially connected Conv convolutions of a Bottleneck module in the YOLOv5s model by using a Ghost convolution module, inserting a CA attention mechanism module in front of an SPP layer of the YOLOv5s model, then replacing a CIOU loss function of the YOLOv5s model with a WIOUv3 loss function, and finally, inserting a CIOU attention mechanism module in front of the SPP layer of the YOLOv5s model into the WIOUv3 loss function. And obtaining a YOLOv5s improved model, training the YOLOv5s improved model, and detecting a bird target by using the YOLOv5s improved model. According to the method, the calculated amount of the bird target detection process is reduced, the detection efficiency is improved, the detection precision is higher, and the application range is wider.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, specifically providing a method for detecting bird targets using an improved YOLOv5s model. Background Technology

[0002] In recent years, bird target detection technology has been widely applied in resource-constrained fields such as intelligent security monitoring, power transmission line safety operation and maintenance, and wildlife biodiversity monitoring and protection due to its significant value and practical effectiveness. Relying on cutting-edge algorithms and advanced equipment, this technology enables real-time and accurate bird detection, providing robust technical support for various industries.

[0003] The development of bird target detection methods can be divided into several stages. In the early stages, bird target detection methods mainly relied on traditional image processing techniques and pattern recognition methods. Specifically, edge detection, color histograms, and texture analysis were used to extract bird target features, and then template matching and rule-based classification were used to achieve bird detection and recognition. In relatively simple scenes, this traditional method could perform preliminary detection of birds with significant features. However, given the wide variety of bird morphologies and the complexity of their environments, traditional detection methods have significant shortcomings in generalization ability, and their detection results are poor for bird images with similar morphologies or complex backgrounds.

[0004] With the rapid development of machine learning algorithms, methods such as Support Vector Machines (SVMs) and Decision Trees have gradually emerged and been applied in the field of bird target detection. In practice, manually designed feature extraction algorithms, such as SIFT and SURF, are first used to accurately extract local features of bird targets. These extracted features are then input into the corresponding model for training and classification. Compared with earlier detection methods, these machine learning-based methods have achieved significant improvements in both detection accuracy and generalization ability. However, because the feature extraction process is highly dependent on manual operation, any bias in the selected features may lead the model to make unreasonable decisions.

[0005] In deep learning, Convolutional Neural Networks (CNNs) have brought significant breakthroughs to the field of object detection. Classic CNN models such as AlexNet and VGGNet have been widely applied to object detection tasks. CNN models, through multi-layer convolution and pooling operations, can automatically learn the features of objects, thus enabling them to process complex images. From an algorithmic architecture perspective, object detection algorithms are mainly divided into two-stage (such as R-CNN and its improved version Faster R-CNN) and one-stage (such as YOLO and SSD). Two-stage algorithms have high detection accuracy and can effectively handle the detection of small objects and complex scenes. However, their computational cost is enormous, and their detection speed is relatively slow, making it difficult to meet the stringent requirements of real-time detection. One-stage algorithms eliminate the step of generating candidate regions and directly detect the object. Among them, YOLO and SSD are outstanding representatives of one-stage algorithms. In particular, the YOLO model can process data quickly, has strong adaptability to changes in bird pose and size, and achieves a good balance between detection speed and accuracy, thus being widely used in real-time detection scenarios.

[0006] Throughout the iterative optimization of the YOLO model, different versions have each possessed unique characteristics and significant implications. The YOLOv3 model achieved a harmonious balance between speed and accuracy, boasting high detection precision and excellent real-time performance. However, its numerous parameters and large size made it difficult to run on resource-constrained devices. The YOLOv4 model improved upon YOLOv3, enhancing accuracy and adaptability to some extent, but its higher computational resource requirements further limited its application in certain scenarios. The YOLOv5 model has evolved towards a lighter design, encompassing multiple versions such as YOLOv5s, YOLOv5m, YOLOv5l, and YOLOv5x. Among these, the lightest, YOLOv5s, has relatively fewer parameters and lower computational cost, making it suitable for some resource-constrained scenarios. However, even with these advancements, the application of the YOLOv5s model remains somewhat limited on devices with even more limited resources. Therefore, how to further lightweight the YOLOv5s model while improving its operating efficiency and detection accuracy to broaden its practical application scenarios remains an important research direction in the field of target detection. Summary of the Invention

[0007] To address the aforementioned problems, this invention provides a method for detecting bird targets using an improved YOLOv5s model. By optimizing the YOLOv5s model, the computational load and number of model parameters are reduced, and the model size is decreased. This results in a more efficient bird target detection method that remains applicable even under resource constraints, thereby improving the detection accuracy of bird targets.

[0008] The method for detecting bird targets using an improved YOLOv5s model provided by this invention includes: S1: Optimize the YOLOv5s model to obtain an improved YOLOv5s model: Replace the two cascaded Conv convolutions in the bottleneck module of the YOLOv5s model with the Ghost convolution module; Insert a CA attention mechanism module before the SPP layer of the YOLOv5s model; Replace the CIOU loss function of the YOLOv5s model with the WIOUv3 loss function; S2: Collect multiple target images, label the target information in the target images, and the labeled target images form a dataset; S3: Improve the model by training YOLOv5s using the dataset; S4: Input the image to be detected into the trained YOLOv5s improved model to detect bird targets in the image.

[0009] Preferably, the intrinsic map and Ghost feature map of the target image are generated by the Ghost convolution module; The intrinsic map is generated by partial convolution kernels, and the calculation formula is as follows: ; in, Representing the intrinsic graph, This indicates that the convolutional filter of the YOLOv5s model was replaced by the Ghost convolutional module after replacing the two cascaded Conv convolutions. This represents the convolution operation. Indicates the deviation term; The intrinsic image is transformed linearly to obtain the Ghost feature map. The linear transformation formula is as follows: ; in, Indicates the number of intrinsic maps. Represents the Ghost feature map. Indicates the first The eigenmap is used for the first time Linear operations. This indicates the number of linear transformations.

[0010] Preferably, coordinate information is embedded in the feature map of the target image and coordinate attention is generated through the CA attention mechanism module; The process of embedding coordinate information is as follows: the two-dimensional RGB channel attention of the CA attention mechanism module is decomposed into two one-dimensional RGB channel attention encoding processes, and the features in the feature map of the target image are aggregated along the vertical and horizontal directions respectively; The process of generating coordinate attention is as follows: the feature maps in the vertical direction and the feature maps in the horizontal direction of the target image are concatenated, and then dimensionality reduction and activation operations are performed to obtain coordinate attention.

[0011] Preferably, the attention encoding process for the two one-dimensional RGB channels is as follows: For each channel and each height location of the target image, average along the width dimension: ; For each channel and each width position of the target image, average along the height dimension: ; in, This represents the number of pixels of the target in the width dimension of the target image. Represents the first [value] of the target image The channel is at a height of Averaged along the width dimension. This indicates the position of the target image in the height dimension. The first part represents the target image. A set of pixel values ​​for each channel. This represents the position index of the target image in the width dimension. This represents the number of pixels in the height dimension of the target image. Represents the first [value] of the target image Each channel is located at the width position. Averaged along the height dimension. This represents the position index of the target image in the height dimension.

[0012] Preferably, the generated coordinate attention is: ; in, This indicates that after coordinate attention processing, the target image of the [missing information] is [missing information]. Each channel, location is The output value at that location, Represents the target image. Each channel, location is The original pixel value at that location, Indicates the first image relative to the target image Each channel, width position Related height-oriented attention weights, Indicates the first image relative to the target image Each channel, height position Related width-direction attention weights.

[0013] Preferably, the expression for the WIOUv3 loss function is: ; ; in, This represents the WIOUv3 loss function. Indicates intersection, union, and ratio. This represents the width of the area where the predicted bounding box overlaps with the ground truth bounding box. This indicates the height of the area where the predicted bounding box overlaps with the ground truth bounding box. This indicates the computational load for the relevant region. Indicates the width of the prediction box. Indicates the height of the predicted bounding box. This represents the width of the actual bounding box. This represents the height of the actual bounding box.

[0014] Preferably, the target information includes target features.

[0015] Preferably, the target features in the target image are labeled using the minimum bounding rectangle annotation method.

[0016] Preferably, multiple aircraft photos are collected, multiple aircraft images are extracted from the multiple aircraft photos, and the multiple aircraft images are scaled according to a predetermined size, rotated by a predetermined angle, and then randomly added to different target images, with the aircraft images used as negative samples.

[0017] Compared with the prior art, the present invention can achieve the following beneficial effects: This invention replaces the two cascaded Conv convolutions in the bottleneck module of the YOLOv5s model with the Ghost convolution module, significantly reducing the computational load and number of model parameters of the YOLOv5s model, lowering computational complexity, and significantly reducing the resource and time costs required by the YOLOv5s model during computation. This improves the detection speed of bird targets and also reduces the size of the YOLOv5s model, allowing it to occupy less storage space. This makes the method applicable to devices with limited computing power and memory resources.

[0018] Introducing a CA attention mechanism module before the SPP layer in the backbone network of the YOLOv5s model enables the YOLOv5s model to focus more on long-range spatial features in the feature map, capturing long-range spatial dependencies and enhancing its ability to extract target features. By decomposing 2D channel attention into two 1D encoding processes, aggregating features along the vertical and horizontal directions respectively, the YOLOv5s model better understands the relationships between different positions within a channel, avoiding a large computational burden while retaining key positional information. This improves the accuracy of YOLOv5s model detection and localization, helping to preserve precise feature location information and providing a reliable basis for subsequent feature localization, achieving more accurate feature localization. The introduction of the CA attention mechanism module is significant for the YOLOv5s model in understanding complex spatial structures such as the shape, size, and orientation of targets, enhancing its ability to express features and thus improving its performance in target detection, particularly increasing the detection accuracy of bird targets.

[0019] To address the issues with the bounding box loss function in the YOLOv5s model, this invention introduces the WIOUv3 loss function to replace the CIOU loss function. By employing a dynamic non-monotonic allocation mechanism to dynamically adjust the gradient gain, the YOLOv5s model can focus on anchor boxes of average quality, while preventing low-quality anchor boxes from generating large harmful gradients, thus improving the generalization ability of bird target detection.

[0020] This invention adds negative samples of aircraft to the target image to help the YOLOv5s model learn bird recognition better, enhances the YOLOv5s model's ability to distinguish between aircraft and birds, improves the ability of bird target detection to distinguish non-targets, and thus improves the detection accuracy of bird targets. Attached Figure Description

[0021] Figure 1 This is a flowchart of a method for detecting bird targets using an improved YOLOv5s model according to an embodiment of the present invention; Figure 2 This is a graph of two concatenated Conv convolution operations in a YOLOv5s model; Figure 3 This is an operation diagram of the Ghost convolution module in the improved YOLOv5s model provided according to an embodiment of the present invention; Figure 4 This is an algorithm structure diagram of the improved YOLOv5s model provided according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the CA attention mechanism module provided in an embodiment of the present invention; Figure 6This is a diagram of the WIOUv3 loss function structure provided in an embodiment of the present invention; Figure 7 This is a diagram showing the specific parameter settings for the simulation experiment provided in the embodiments of the present invention; Figure 8 This is a parameter diagram of various parameters corresponding to the improved YOLOv5s model after training on the training set, provided according to an embodiment of the present invention. Figure 9 This is a parameter diagram of various parameters corresponding to the YOLOv5s improved model after testing with a test set, provided according to an embodiment of the present invention; Figure 10 This is a precision-recall curve corresponding to the original YOLOv5s model provided in the embodiments of the present invention; Figure 11 This is a precision-recall curve corresponding to the improved YOLOv5s model provided in the embodiments of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and do not constitute a limitation thereof. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of the invention. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, some operations related to the invention are not shown or described in the specification. This is to avoid obscuring the core parts of the invention with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0023] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined to form various implementations. Furthermore, the order of the steps or actions in the method description can be changed or adjusted in a manner readily apparent to those skilled in the art. Therefore, the various orders in the specification and drawings are merely for the clear description of a particular embodiment and do not imply a mandatory order, unless otherwise stated that a particular order must be followed.

[0024] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation on this invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0025] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0026] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0027] like Figure 1 As shown, this embodiment of the invention provides a method for detecting bird targets using an improved YOLOv5s model, comprising the following steps: S1: Optimize the YOLOv5s model to obtain an improved YOLOv5s model: Replace the two cascaded Conv convolutions in the bottleneck module of the YOLOv5s model with the Ghost convolution module, insert the CA attention mechanism module before the SPP layer of the YOLOv5s model, and replace the CIOU loss function of the YOLOv5s model with the WIOUv3 loss function.

[0028] The YOLOv5 model includes a bottleneck module, characterized by a narrow middle and wide ends. This means the input and output ends of the bottleneck module are relatively wide, while the lower-dimensional layers in the middle are narrow. This narrow middle layer acts as a bottleneck, leading to a significant accumulation of computational costs in the YOLOv5 model. Therefore, the middle lower-dimensional layers are the region with the highest concentration of computational costs and model parameters in the YOLOv5 model. Thus, lightweighting the middle lower-dimensional layers can significantly reduce the overall number of model parameters and computational costs in the YOLOv5 model.

[0029] like Figure 2 As shown (the features in the figure are the feature maps), the bottleneck module in the YOLOv5s model contains two cascaded Conv convolutions. In a conventional convolution operation, the target image is input into the first Conv convolution, which calculates and outputs a feature map. This output feature map is then input into the second Conv convolution for calculation to obtain the feature map of the target image. This process generates a large number of similar redundant feature maps, leading to increased computation. In this embodiment of the invention, the two cascaded Conv convolutions in the bottleneck module are replaced with a Ghost convolution module. The Ghost convolution module generates the feature map of the target image as follows: First, a small number of convolution kernels are used to generate a portion of the feature map, i.e., the intrinsic map, from the input target image. Then, the intrinsic map is used to generate more similar feature maps through a cheap operation (i.e., a simple and easy-to-implement operation; in this embodiment, the cheap operation is a linear transformation). In this embodiment, these similar feature maps are called Ghost feature maps. Finally, the intrinsic map and the Ghost feature map are concatenated to form the final feature map of the target image.

[0030] The core objective of the Ghost convolution module is to generate feature maps of a target image using a series of inexpensive operations. By deeply mining the rich feature information contained in the feature maps, the Ghost convolution module constructs a high-performance convolutional neural network. In practical applications, the target image only needs to be processed by the Ghost convolution module to successfully obtain the corresponding feature map, effectively reducing the computational load and number of parameters in the YOLOv5s model, significantly lowering the computational cost of obtaining feature maps, and providing strong support for the lightweight and efficient operation of the model.

[0031] like Figure 2 The image shown illustrates the process by which the bottleneck module of the YOLOv5s model generates feature maps based on the target image: Given input data That is, the target image. ,in, This indicates the number of channels for the input data; in this embodiment of the invention, it is RGB three channels. This represents the number of pixels in the height dimension of the input target image. This represents the number of pixels in the width dimension of the input target image. Represents the real number field. It is generated after two concatenated Conv convolutions. Feature map : , in, This indicates the result of two concatenated Conv convolutions. There are feature maps, and the height of each feature map is . Width is , This represents the convolution operation. Indicates the deviation term. Indicates that the convolution kernel is Convolutional filters.

[0032] The number of floating-point operations per second corresponding to two concatenated Conv convolutions and model parameter count It can be represented as: , , in, This indicates the size of the convolution filter, and the number of floating-point operations is an indicator of the computational complexity of the model.

[0033] like Figure 3 As shown, Figure 3 This demonstrates the process of generating feature maps by replacing two concatenated Conv convolutions with the Ghost convolution module. The specific details of this process are as follows: Suppose an intrinsic map is generated by a small number of convolutional kernels. have indivual, ,but: , in, This indicates that the convolutional filter in the YOLOv5s model was replaced with the Ghost convolutional module after two cascaded Conv convolutions. The corresponding number of floating-point operations per second is shown below. and model parameter count It can be represented as: , .

[0034] Remove the computation generated by a small number of convolutional kernels. Individual eigenmaps The rest Each feature map (i.e., the Ghost feature map) needs to be... Linear transformation of eigenmap This was obtained the following time, among which... , The "1" subtracted in the sublinear transformation is... Figure 3 China The "1"th order of the linear transformation of the eigenmap.

[0035] The linear transformation corresponding to the following formula can be used to generate Ghost feature maps: .

[0036] Through the intrinsic map Each basic feature is performed Linear operations still require computation. Then, it is combined with the data generated by a small number of convolution kernels. Individual eigenmaps directly spelled Each feature map (this process incurs no computational cost). Indicating the intrinsic map The first in Each intrinsic map, Indicates the first The eigenmap is used for the first time Linear transformation The first element generated by the linear transformation process is represented by the second element. A Ghost feature map, Indicates the degree of the linear transformation.

[0037] Let the kernel size of the linear transformation be... The number of floating-point operations per second generated by the above linear transformation process and model parameter count It can be represented as: , .

[0038] Obtained from calculation , , , , , It can be calculated that: The computational cost ratio of the Ghost convolution module to two cascaded Conv convolutions. for: .

[0039] The ratio of the number of parameters in a Ghost convolutional module to that in two cascaded Conv convolutions. for: .

[0040] Pick and Same size, and The ratio of computational workload can be increased. and the ratio of model parameters The expression is further simplified to: .

[0041] As can be seen, when the two cascaded Conv convolutions in the bottleneck module are replaced with the Ghost convolution module, the computational cost and number of parameters of the resulting YOLOv5s model are approximately the same as those of the original YOLOv5s model. That is, the amount of computation and the number of model parameters are reduced by approximately Times. Therefore, as Figure 4 As shown, in this embodiment of the invention, by replacing the two cascaded Conv convolutions in the bottleneck module with the Ghost convolution module, it is possible to ensure that the feature map redundancy does not affect the detection accuracy of the YOLOv5s model, while also reducing the overall convolutional computation and model parameter count of the YOLOv5s model, thereby improving the detection speed of the YOLOv5s model.

[0042] Figure 4 The explanation of the design is as follows: input: input Backbone: backbone network Neck: Neck (feature fusion network) Head: head (detection head) Concat: concatenate. up sample: upsampling Conv: Convolution, CBL: A module consisting of Conv (convolution), BN (Batch Normalization), and Leaky ReLU (linear unit activation function with leakage correction). CSP1_X: A CSP (Cross Stage Partial) structural module containing X bottleneck components. CSP2_1: A CSP structure module containing one Bottleneck component. CA: Coordinate Attention BN: Batch Normalization Leaky ReLU: Linear unit (activation function) with leakage correction. Bottleneck: bottleneck (module) add: addition (operation, feature addition).

[0043] The YOLOv5s model's backbone network includes an SPP layer, and the feature maps preceding the SPP layer contain rich feature information. To enhance the YOLOv5s model's feature extraction capability, this embodiment of the invention inserts a CA attention mechanism module before the SPP layer in the YOLOv5s model's backbone network. The CA attention mechanism module can fully mine and utilize this rich feature information, thereby enhancing the model's feature representation capability. Since the SPP layer is relatively far from the target image input location, and the spatial information between the SPP layer and the target image input location is rich, inserting the CA attention mechanism module can capture long-distance spatial dependencies, fully utilize the spatial information between the SPP layer and the target image input location, strengthen the YOLOv5s model's ability to extract target features, and improve the accuracy of the YOLOv5s model in detecting bird targets. This is of great significance for understanding complex spatial structures such as the shape, size, and orientation of targets.

[0044] like Figure 5 As shown, the CA attention mechanism module specifically functions as coordinate information embedding and coordinate attention generation. Specifically, it embeds coordinate information into the feature map and generates coordinate attention. Coordinate information embedding refers to adding location labels to the target features in the feature map; the location refers to the position of the labeled target feature within the target image. Coordinate attention generation involves generating coordinate attention for the labeled target features, reflecting the importance of the corresponding target features. Based on the generated coordinate attention, the importance of different target features in the target image is distinguished.

[0045] Figure 5In YOLOv5, Residual refers to a residual connection, a structural design in deep learning (especially ResNet). It allows input features to "skip" some layers and be added to the features processed by those layers. This alleviates the vanishing gradient problem during deep network training, making the network easier to train and extracting features more effectively. Re-weighting refers to adjusting feature weights to make the YOLOv5 improved model focus more on important features. Sigmoid is an activation function that maps input values ​​to between 0 and 1, generating weight values ​​for subsequent re-weighting operations, helping the YOLOv5 improved model assign different levels of importance to different features.

[0046] The specific process of embedding coordinate information (embedding coordinate information in the feature map) is as follows: To avoid compressing all spatial information between the SPP layer and the target image input location into the RGB channels, and to capture long-distance spatial interactions with precise location information (interaction refers to the fusion of different features, such as color features, texture features, etc.), this embodiment of the invention decomposes the two-dimensional RGB channel attention of the CA attention mechanism module into two one-dimensional RGB channel attention encoding processes, which aggregate features in the feature maps along the vertical and horizontal directions respectively, avoiding a large amount of computational burden while retaining key location information.

[0047] Global average pooling is decomposed into two one-dimensional average pooling operations: horizontal position information embedding and vertical position information embedding.

[0048] The horizontal position information is embedded using a size of The average pooling kernel, along the horizontal direction, applies to the input. ( The feature map (representing the dimension, i.e., channels) is subjected to channel-by-channel (RGB channels) pooling operation to obtain... The feature map. Its calculation method is to perform a feature map on each channel. (RGB channels) and each height position Averaged over the width dimension: , in, This represents the number of pixels of the target in the width dimension of the image, i.e., the length in the width direction. Represents the first [value] of the target image The channel is at a height of Averaged along the width dimension. This indicates the position of the target image in the height dimension. The first part represents the target image. A set of pixel values ​​for each channel. This represents the position index in the width dimension of the target image, used to determine the specific width position.

[0049] Use size The average pooling kernel, along the vertical direction, affects the input. The feature map is pooled channel-by-channel (RGB channels) to obtain The feature map is calculated for each channel. (RGB channels) and each width position Averaging over the height dimension: , in, This represents the number of pixels in the height dimension of the target image. Represents the first [value] of the target image Each channel is located at the width position. Averaged along the height dimension. This represents the position index of the target image in the height dimension, used to determine the specific height location (used when averaging in the height dimension). The first part represents the target image. A set of pixel values ​​for each channel.

[0050] By decomposing a set of (vertical and horizontal) feature maps generated by global average pooling, on the one hand, the CA attention mechanism module can establish long-distance dependencies within a single channel (RGB channel), thereby better understanding the association between different positions within the channel; on the other hand, these feature maps help to retain accurate positional information, providing a reliable basis for feature localization and achieving more accurate feature localization.

[0051] The specific process of generating coordinate attention is as follows: Feature maps are arranged according to spatial dimensions. and feature map Perform the concat operation (concat operation), the formula is as follows: , in, This represents a non-linear activation function, such as the ReLU function. and These represent a set of feature maps (vertical and horizontal). This represents the intermediate feature map obtained after splicing and non-linear operations.

[0052] splicing feature map and feature map Then, using 1×1 convolution... The convolutional transformation function reduces the dimensionality of the concatenated feature maps and then activates them using a non-linear activation function to generate new feature maps. ,in, The reduction rate is used to control the degree of dimensionality reduction; feature map. The dimension is , length is 1, width is Then, examine the feature map. Normalization and nonlinear mapping are performed without changing the feature map size. The feature map is then processed along the spatial dimension. Perform a split operation (slicing operation) to divide into... and Then use each and Dimensionality reduction is performed using a convolutional transformation function (1×1 convolution), followed by activation using a sigmoid activation function, to obtain... and Coordinate attention: , .

[0053] in, This represents the attention weights generated along the height direction. This represents the attention weights generated along the width direction. This represents the Sigmoid activation function, used to map values ​​to the [0,1] interval to obtain attention weights. This represents the transformation of the input features along the width direction. This represents the transformation of the input features along the height direction.

[0054] Finally, the formula for generating the corresponding output from coordinate attention can be expressed as: ; in, This indicates that after coordinate attention processing, the target image of the [missing information] is [missing information]. Each channel, location is The output value at that location, Represents the target image. Each channel, location is The original pixel value at that location, Indicates the first image relative to the target image Each channel, height position Related height-oriented attention weights, Indicates the first image relative to the target image Each channel, width position Related width-direction attention weights, This indicates the position of the target image in the width dimension. This indicates the position of the target image in the height dimension.

[0055] The output of the YOLOv5s model includes a Head layer containing the CIOU loss function. In this embodiment of the invention, the CIOU loss function of the YOLOv5s model is replaced with the WIOUv3 loss function.

[0056] The WIOU loss function introduces weighting factors to weight the anchor boxes, weighting the overlapping and non-overlapping regions of the predicted box (the anchor box closest to the ground truth box is considered the predicted box) and the ground truth box (the bounding box of the target area). It also handles non-overlapping cases (where the predicted and ground truth boxes have no spatial overlap), making the WIOU loss function more widely applicable in object detection tasks. The WIOUv3 loss function is an improved version of the WIOU loss function. It introduces a dynamic non-monotonic focusing mechanism, dynamically adjusting the loss weights based on the quality of the predicted box, strengthening the focus on low-quality anchor boxes. This mechanism allows the optimized YOLOv5s model to focus more on low-quality anchor boxes during training, thereby improving overall detection performance. Replacing the CIOU loss function of the YOLOv5s model with the WIOUv3 loss function can also accelerate YOLOv5s model convergence and improve object detection efficiency.

[0057] like Figure 6 The diagram shown illustrates the structure of the WIOUv3 loss function, where the coordinates of the predicted bounding box are... The coordinates of the ground truth bounding box are In the WIOUv3 loss function, the Intersection over Union (IOU) is used to measure the degree of overlap between predicted and ground truth bounding boxes in object detection tasks. The IOU represents the ratio of the intersection area to the union area of ​​the predicted and ground truth bounding boxes, and is defined as: , , in, This represents the WIOUv3 loss function. This represents the width of the area where the predicted bounding box overlaps with the ground truth bounding box. This indicates the height of the area where the predicted bounding box overlaps with the ground truth bounding box. This indicates the computational cost for the relevant area.

[0058] The Intersection over Union (IOU) ratio has a flaw, which can be reflected by the following formula: , in, This indicates gradient calculation. Since there's a difference between the predicted and ground truth boxes, the loss function measures this difference. A larger loss function results in a smaller intersection-over-union ratio (IoU), meaning less overlap between the predicted and ground truth boxes, indicating a greater difference between them. The goal of the calculation is to minimize the difference between the predicted and ground truth boxes, making them closer. Therefore, backpropagation of the loss function is necessary, which involves calculating the partial derivative. During this process, the partial derivative may vanish, meaning it disappears when there is no overlap between the two anchor boxes. or ), WIOUv3 loss function The vanishing gradient during backpropagation causes the width of the overlap between the predicted and ground truth bounding boxes to decrease. Unable to update while training YOLOv5s model.

[0059] The YOLOv5s model uses the CIOU loss function as the bounding box loss function. This loss function has the following problems: the penalty term is always 0 when the aspect ratios of the predicted and ground truth boxes are the same; the width and height of the predicted box cannot be increased or decreased simultaneously; and the penalty is increased for low-quality training samples caused by geometric factors. This invention addresses this by constructing a new penalty term... To address this issue, the bounding box loss of YOLOv5s models is based on additive loss and follows this paradigm: . in, This represents the bounding box loss of the YOLOv5s model after the loss function is replaced (after the introduction of the penalty term).

[0060] To address the issues with the CIOU loss function in the YOLOv5s model, this invention introduces the WIOUv3 loss function to replace the CIOU loss function as the bounding box loss function for the improved YOLOv5s model. The WIOUv3 loss function is defined as follows: , , , in, Indicates outlier degree. Indicates the monotonic focusing coefficient. The moving average of momentum. Indicates the non-monotonic focusing coefficient. and Both represent hyperparameters. and outlier Constitutes non-monotonic focusing coefficient , This represents the loss corresponding to the intersection-union ratio (IOU, a metric that measures the degree of overlap between the predicted and ground truth boxes). This indicates a separation operation.

[0061] The loss function for WIOUv1 is defined as follows: , , , in, This represents the coordinates of the center point of the true bounding box. Indicates the coordinates of the center point of the prediction box. This represents the width of the smallest bounding box that simultaneously contains both the predicted and ground truth boxes. The smallest bounding box is the smallest box that can cover both the predicted and ground truth boxes. This represents the height of the smallest bounding box that contains both the predicted and ground truth boxes. The separation operation involves defining outliers in the WIOUv3 loss function to describe anchor frame quality, then constructing non-monotonic focusing coefficients using the outliers, and finally combining these non-monotonic focusing coefficients with a distance-based attention loss function. Combining these, we finally obtain the WIOUv3 loss function. .

[0062] S2: Collect multiple target images, label the target information in the target images, and the labeled target images form a dataset.

[0063] Multiple target images are acquired. These images can be self-taken bird photos or existing bird photos selected directly. In this embodiment, target images are selected from the bird dataset CUB-200-2011. The selected target images contain birds with varying postures and species, and each target image must contain at least one bird. When a target image contains multiple birds, the birds can be of the same species or different species. This embodiment selects 800 target images.

[0064] The target information in the target image is labeled, including target features and the corresponding bounding boxes. Target features should be representative and identifiable features of the detected target. For example, in this embodiment, the target is a bird; therefore, the bird's head, beak, and tail are selected as target features, and the heads, beaks, and tails of birds in multiple selected target images are labeled. This embodiment uses LabelIImg software and the minimum bounding rectangle annotation method to label the bounding boxes of the bird's head, beak, and tail in the target image. The bounding box refers to the boundary of the bird's head, beak, and tail. Since the bounding box is the true boundary of the labeled target features, it is also called the true bounding box. The location refers to which part of the head, beak, or tail the labeled image belongs to. The labeled target images constitute a dataset.

[0065] When detecting bird targets, airplanes are the objects most easily confused with birds. This invention adds airplane images as negative samples to the target image to help the improved YOLOv5s model better learn bird recognition and enhance its ability to distinguish between airplanes and birds. The method for adding negative samples is as follows: various airplane photos are randomly selected or taken, the airplane images are extracted, and preprocessed. The preprocessing includes scaling the airplane image to a predetermined size and rotating it by a predetermined angle. Scaling the airplane image to a size similar to that of a bird, then rotating it by different angles, and randomly adding it to different target images completes the addition of negative samples.

[0066] S3: Use the dataset to train YOLOv5s to improve the model.

[0067] Before using the YOLOv5s improved model to detect bird targets, the YOLOv5s improved model needs to be trained using a dataset.

[0068] Before training an improved YOLOv5s model using target images from the dataset, the parameters of the target images in the dataset need to be set first. These parameters include the target image size, momentum term, learning rate, batch size, number of training iterations, and weight decay regularization. For example... Figure 7As shown, the specific parameters of the target image in this embodiment of the invention are as follows: the image size of the input target image is uniformly adjusted to 640×640 pixels to accelerate the training speed of the improved YOLOv5s model; the stochastic gradient descent method with a momentum term of 0.937 is used during training; in the first 50 rounds, the learning rate in the training of the improved YOLOv5s model is set to 0.001, and as the number of iteration rounds increases, the learning rate in the last 50 rounds of training is reduced to 0.0001; the batch size is set to 100, and the batch size is 8; to prevent the improved YOLOv5s model from overfitting, the weight decay regularization term is set to 0.0005. After 100 rounds of iterative training, the test results of the improved YOLOv5s model are as follows. Figure 8 , Figure 9 , Figure 10 and Figure 11 As shown.

[0069] Figure 7 In the table, birds0.9 represents the identified birds with a confidence level of 0.9, head0.8 represents the identified bird heads with a confidence level of 0.8, tail0.8 represents the identified bird tails with a confidence level of 0.8, and beak0.9 represents the identified bird beaks with a confidence level of 0.9. Figure 8 , Figure 9 The table presents relevant metrics for the improved YOLOv5s model at different stages of training (e.g., number of iterations, training steps). The horizontal axis represents the number of iterations, and the vertical axis represents the numerical values. Results indicate the results of the improved YOLOv5s model during training, based on the current parameters, and reflect its performance during training. The classification loss function measures the error of the improved YOLOv5s model in predicting target categories. It effectively evaluates the model's ability to correctly classify targets (e.g., distinguishing between "tail" and "beak"), making the model more accurate in category judgment. The target loss function comprehensively considers the errors of the improved YOLOv5s model in target detection across multiple aspects (e.g., category, bounding box). It is typically a composite loss function combining classification loss and bounding box regression loss, ensuring that the improved YOLOv5s model not only correctly classifies targets but also accurately locates them (determines their position in the image). Figure 10 , Figure 11In this context, precision refers to the proportion of actual positive instances among those predicted as positive by the YOLOv5s improved model (e.g., detecting "bird"). Recall refers to the proportion of all actual positive targets that were successfully predicted by the YOLOv5s improved model. "birds0.937" means that for the category "birds," the average precision (AP) of the YOLOv5s improved model is 0.937. Average precision is a measure of the detection performance for this category, considering both precision and recall; a higher value indicates better detection performance for "birds." "all classes" means all categories. "@0.5" represents the mean average precision, which is the average precision of all classes. "@0.5" indicates that the Intersection over Union (IOU) threshold used in the calculation is 0.5. "all classes 0.937 mAP@0.5" means that the mean average precision of all classes at an IOU threshold of 0.5 is 0.937, reflecting the good overall detection performance of the improved YOLOv5s model across all classes.

[0070] Training the improved YOLOv5s model using a dataset yields the weight file of the YOLOv5s model with the best convergence. This weight file contains standard parameters, which are used to determine whether a detected target is a bird. Before training, the weight file of the improved YOLOv5s model is empty. The weight parameters with the highest accuracy during training are saved and used as the standard parameters for the improved YOLOv5s model's weight file.

[0071] For the dataset, this invention provides another embodiment, dividing the dataset into a training set, a test set, and a validation set. After obtaining the improved YOLOv5s model, the improved YOLOv5s model is trained using the training set to obtain the optimal improved YOLOv5s model. Then, the improved YOLOv5s model is tested and validated using the test set and validation set: Target images from the test set are input into the improved YOLOv5s model, allowing it to identify targets in the target images within the test set. Performance metrics of the improved YOLOv5s model during the testing process are collected, including average precision, number of floating-point operations, and number of model parameters. The performance of the improved YOLOv5s model is evaluated based on these performance metrics. Target images from the test set are input into the improved YOLOv5s model, allowing it to identify targets in the target images within the validation set, thereby verifying the effectiveness of the improved YOLOv5s model in identifying targets.

[0072] If the dataset is divided into training, testing, and validation sets, the proportions need to be set according to requirements. For example, in this embodiment of the invention, the improved YOLOv5s model needs to be trained, tested, and validated. The focus of the improved YOLOv5s model is to complete the model training so that the model can learn better. Therefore, the training set accounts for a larger proportion of the dataset, while the testing and validation sets account for a smaller proportion. In this embodiment of the invention, the proportions of the training, testing, and validation sets are 8:1:1, that is, the training set accounts for 80% of the dataset and contains 640 target images, the testing set accounts for 10% of the dataset and contains 80 target images, and the validation set accounts for 10% of the dataset and also contains 80 target images.

[0073] S4: Input the image to be detected into the trained YOLOv5s improved model to detect bird targets in the image.

[0074] The trained YOLOv5s improved model is capable of identifying bird targets in the image to be detected. The image to be detected is input into the trained YOLOv5s improved model, and the YOLOv5s improved model outputs the bird information contained in the image to be detected.

[0075] To objectively and accurately evaluate the performance of the improved YOLOv5s model, this embodiment of the invention conducts a comparative experiment between the improved YOLOv5s model and the standard YOLOv5s model. The same dataset and hyperparameters were used to train the YOLOv5s model identically to ensure consistency of experimental conditions and avoid interference from other factors. The training process strictly followed the standard procedure for training the improved YOLOv5s model until the pre-set total number of rounds was completed. After training, the same test set was used to test the YOLOv5s model, obtaining its average precision, computational complexity (reflected by the number of floating-point operations), and number of model parameters, which were then compared with the test results of the improved YOLOv5s model. The specific steps are as follows: Floating-point operations (FLOPS) are used to measure the computational complexity of a model. A higher FLOPS indicates a greater computational load and higher demands on hardware computing power. A comparison of the floating-point operations of the improved YOLOv5s model and the standard YOLOv5s model is provided: Calculate the number of floating-point operations for two concatenated Conv convolutions in a YOLOv5s model. : , Both multiplication and addition operations are considered as two floating-point operations. and These represent the number of pixels in the height dimension and the number of pixels in the width dimension of the feature map, respectively. and These represent the width and height of the convolution kernel, respectively. This represents the number of channels in the input feature map. This indicates the number of channels in the output feature map.

[0076] In the feature maps output by the improved YOLOv5s model, a small number of convolutional kernels are first used to generate a portion of the feature maps (i.e., intrinsic maps) from the input target image. Then, these intrinsic maps are used to generate more Ghost feature maps through inexpensive operations. Therefore, the number of floating-point operations in the Ghost convolution module of the improved YOLOv5s model is significantly reduced. It consists of two parts: the number of floating-point operations for a small number of convolutional kernels and the number of floating-point operations for inexpensive operations. Calculate the number of floating-point operations for the small number of convolutional kernels in this case. and the number of floating-point operations for cheap operations They are respectively: , , In this case, both multiplication and addition operations are considered as two floating-point operations.

[0077] Floating-point operation count of the CA attention mechanism module Including the pooling operation required The number of floating-point operations required for a convolutional layer The number of floating-point operations required for output convolution. The calculation formulas are as follows: , , , in, Indicates the number of channels. This indicates the channel reduction rate.

[0078] The number of model parameters refers to the total number of parameters that need to be learned in the model, including the weights and biases of convolutional layers. The more parameters a model has, the stronger its expressive power, but it also requires more training data and computational resources to avoid overfitting.

[0079] The number of model parameters corresponding to two cascaded Conv convolutions in the YOLOv5s model. for: , in, and These represent the width and height of the convolution kernel, respectively. The model parameters of the Ghost convolution module in the improved YOLOv5s model also consist of two parts: the model parameters of a small number of convolution kernels. And the number of model parameters for inexpensive operation Calculate the number of model parameters with a small number of convolutional kernels. And the number of model parameters for inexpensive operation They are respectively: , , Model parameter count of the CA attention mechanism module Including the number of parameters in the convolutional layer model and the number of parameters of the output convolution model The calculation formulas are as follows: , , Mean Precision The formula for calculation is: , in, Indicates the number of categories in the dataset. The sequence number representing the dataset category. Indicates the first Precision values ​​for each category.

[0080] The number of model parameters, number of floating-point operations, and average precision were obtained through statistical calculations. The YOLOv5s model and the improved YOLOv5s model were compared in terms of these three indicators, and the results are shown in Table 1.

[0081] Table 1: Comparison of YOLOv5s Model and Improved YOLOv5s Model

[0082] In the table, @0.5 / % indicates that the threshold for the intersection-union ratio (IUGR) is 0.5, and the mean accuracy (mAP) is ultimately presented as a percentage. B represents the number of floating-point operations in Billions, and MB represents the number of model parameters in Million Bytes. As can be seen from the table, the improved YOLOv5s model proposed in this embodiment exhibits significant advantages compared to the standard YOLOv5s model. Specifically, in terms of the key performance indicator of mean accuracy (mAP), the improved YOLOv5s model proposed in this embodiment achieves a 2.6% improvement, effectively enhancing the accuracy of object detection. Regarding model lightweighting, the improved YOLOv5s model reduces the number of model parameters by 29%, significantly reducing the hardware resource requirements for storage and deployment. The improved YOLOv5s model also reduces computational complexity by 34% compared to the standard YOLOv5s model. This means that the improved YOLOv5s model proposed in this embodiment requires significantly fewer computational resources during operation, enabling more efficient execution of object detection tasks. In summary, the improved YOLOv5s model proposed in this embodiment not only surpasses the YOLOv5s model in detection accuracy, but also successfully achieves lightweighting of the improved YOLOv5s model by reducing the number of model parameters and floating-point operations, providing a more competitive solution for practical application scenarios.

[0083] This invention targets bird targets. The dataset consists of bird images, and the annotations on the target images are bird features. The improved YOLOv5s model can also be used to detect other types of targets. Simply replace the dataset with images of the corresponding target to be detected, replace the annotated features with the features of the corresponding target to be detected, and train the improved YOLOv5s model to recognize other targets.

[0084] Although embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

[0085] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for detecting bird targets using an improved YOLOv5s model, characterized in that, include: S1: Optimize the YOLOv5s model to obtain an improved YOLOv5s model: Replace the two cascaded Conv convolutions in the bottleneck module of the YOLOv5s model with the Ghost convolution module; Insert a CA attention mechanism module before the SPP layer of the YOLOv5s model; Replace the CIOU loss function of the YOLOv5s model with the WIOUv3 loss function; S2: Collect multiple target images, label the target information in the target images, and the labeled target images form a dataset; S3: Train the improved YOLOv5s model using the dataset; S4: Input the image to be detected into the trained YOLOv5s improved model to detect bird targets in the image.

2. The method for detecting bird targets using an improved YOLOv5s model as described in claim 1, characterized in that, The intrinsic map and Ghost feature map of the target image are generated through the Ghost convolution module; The intrinsic map is generated by partial convolution kernels, and the calculation formula is: ； in, This represents the intrinsic map. This indicates that the convolutional filter of the YOLOv5s model was replaced by the Ghost convolutional module after replacing two cascaded Conv convolutions. This represents the convolution operation. Indicates the deviation term; The intrinsic map is subjected to a linear transformation to obtain the Ghost feature map. The linear transformation formula is as follows: ； in, Indicates the number of intrinsic maps. Represents the Ghost feature map. Indicates the first The intrinsic map described above is used for the first... Linear operations. This indicates the number of linear transformations.

3. The method for detecting bird targets using an improved YOLOv5s model as described in claim 1, characterized in that, The CA attention mechanism module embeds coordinate information into the feature map of the target image and generates coordinate attention. The process of embedding coordinate information is as follows: the two-dimensional RGB channel attention of the CA attention mechanism module is decomposed into two one-dimensional RGB channel attention encoding processes, and the features in the feature map of the target image are aggregated along the vertical and horizontal directions respectively; The process of generating coordinate attention is as follows: the feature map in the vertical direction and the feature map in the horizontal direction of the target image are concatenated, and then dimensionality reduction and activation operations are performed to obtain coordinate attention.

4. The method for detecting bird targets using the improved YOLOv5s model as described in claim 3, characterized in that, The attention encoding process for two one-dimensional RGB channels is as follows: For each channel and each height position of the target image, average along the width dimension: ； For each channel and each width position of the target image, average along the height dimension: ； in, This represents the number of pixels of the target in the width dimension of the target image. Represents the first time for the target image The channel is at a height of Averaged along the width dimension. This indicates the position of the target image in the height dimension. The first part of the target image represents the... A set of pixel values ​​for each channel. This represents the position index of the target image in the width dimension. This indicates the number of pixels in the height dimension of the target image. Represents the first time for the target image Each channel is located at the width position. Averaged along the height dimension. This represents the position index of the target image in the height dimension.

5. The method for detecting bird targets using an improved YOLOv5s model as described in claim 3, characterized in that, The generated coordinate attention is: ； in, This indicates that after coordinate attention processing, the target image of the [missing information] is [missing information]. Each channel, location is The output value at that location, The target image is represented by the first... Each channel, location is The original pixel value at that location, Indicates the first image of the target image Each channel, width position Related height-oriented attention weights, Indicates the first image of the target image Each channel, height position Related width-direction attention weights.

6. The method for detecting bird targets using the improved YOLOv5s model as described in claim 1, characterized in that, The expression for the WIOUv3 loss function is as follows: ； ； in, This represents the WIOUv3 loss function. Indicates intersection, union, and ratio. This represents the width of the area where the predicted bounding box overlaps with the ground truth bounding box. This indicates the height of the area where the predicted bounding box overlaps with the ground truth bounding box. This indicates the computational load for the relevant region. Indicates the width of the prediction box. Indicates the height of the predicted bounding box. This represents the width of the actual bounding box. This represents the height of the actual bounding box.

7. The method for detecting bird targets using the improved YOLOv5s model as described in claim 1, characterized in that, The target information includes target features.

8. The method for detecting bird targets using the improved YOLOv5s model as described in claim 1, characterized in that, The target features in the target image are labeled using the minimum bounding rectangle annotation method.

9. The method for detecting bird targets using the improved YOLOv5s model as described in claim 1, characterized in that, Multiple aircraft photos are collected, and multiple aircraft images are extracted from the multiple aircraft photos. The multiple aircraft images are scaled according to a predetermined size, rotated by a predetermined angle, and then randomly added to different target images. The aircraft images are used as negative samples.

Images