Ship target detection method and system based on remote sensing image

Abstract

The invention discloses a ship target detection method and system based on a remote sensing image, and relates to the technical field of target detection, and the method comprises the specific steps: constructing an initial target detection model which comprises a self-adaptive expansion scale module, a self-adaptive kernel adjustment module and a frequency optimization module; obtaining a training data set, inputting the training data set into the initial target detection model for training, and performing parameter optimization on the initial target detection model based on the training target data and the annotation target to obtain a target detection model; acquiring a to-be-detected image; and inputting a to-be-detected image into the trained target detection model to obtain a detection target. Through combination of the adaptive expansion scale module and the dynamic kernel adjustment module, the capability of capturing details of a ship target in a remote sensing image is effectively enhanced, and the feature response of each pixel can be optimized according to the local context of the pixel due to introduction of the adaptive expansion rate. Therefore, the adaptability to complex backgrounds and targets with different sizes is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of target detection, and more particularly to a ship target detection method and system based on remote sensing images. Background Art

[0002] With the rapid development of remote sensing technology, high-resolution remote sensing images acquired by satellites or drones play an increasingly important role in environmental monitoring, urban planning, military reconnaissance, and disaster assessment. In these applications, the detection and segmentation of ship targets is one of the key tasks, especially in ocean monitoring and shipping safety. Ship targets usually have complex backgrounds and diverse appearances. Affected by natural lighting, sea surface fluctuations, and other environmental factors, traditional image processing methods face many challenges.

[0003] In recent years, the rise of deep learning technology has provided new ideas for image target detection. In particular, convolutional neural networks have shown superior performance in image segmentation and target detection tasks. However, the characteristics of remote sensing images make the standard convolutional neural network architecture not always effective in ship detection. Specifically, ship targets in remote sensing images may show significant size and shape changes due to different shooting angles and environmental factors. At the same time, the complexity of the background may cause the target to be occluded or confused, which further increases the difficulty of detection.

[0004] In order to solve the above problems, many researchers have begun to explore adaptive convolution and multi-scale feature extraction techniques. These methods aim to adapt to targets of different scales by adjusting the size and shape of the convolution kernel. However, while dynamically adjusting the convolution kernel, these methods often fail to effectively process information of different frequency features, resulting in information loss and affecting the final detection results.

[0005] In addition, although some methods have been successfully applied to specific scenarios, their computational complexity and demand for hardware resources often limit their feasibility in real-time applications. Especially in resource-constrained environments, such as mobile devices and edge computing platforms, how to balance detection accuracy and computational efficiency remains an urgent problem to be solved. Therefore, how to maintain computational efficiency while being able to adaptively adjust the receptive field and convolution kernel frequency response is an urgent problem that technicians in this field need to solve. Summary of the invention

[0006] In view of this, the present invention provides a ship target detection method and system based on remote sensing images, which overcome the above-mentioned defects.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] A ship target detection method based on remote sensing images, the specific steps are:

[0009] Target detection model construction method:

[0010] Constructing an initial target detection model, wherein the initial target detection model includes an adaptive expansion scale module, an adaptive kernel adjustment module and a frequency optimization module;

[0011] A training data set is obtained, and the training data set is input into the initial target detection model for training, the adaptive expansion scale module predicts the initial expansion rate of the received training data, and adaptively adjusts the initial expansion rate based on the modulation mechanism to obtain the optimal expansion rate of each pixel; and the optimal dynamic frequency is obtained based on the adaptive kernel adjustment module and the frequency optimization module; training target data is extracted based on the optimal expansion rate of each pixel and the optimal dynamic frequency, and the initial target detection model is optimized based on the training target data and the labeled target to obtain the target detection model;

[0012] Target detection methods:

[0013] Acquire the image to be detected;

[0014] The image to be detected is input into the trained target detection model to obtain the detection target.

[0015] Optionally, the adaptive expansion scale module includes a convolution layer, an activation function layer and a modulation mechanism;

[0016] The convolution layer is used to predict the initial expansion rate of each pixel in the training data;

[0017] The activation function layer is used to perform non-negative processing on the initial expansion rate to obtain a non-negative expansion rate;

[0018] The modulation mechanism is used to dynamically adjust the non-negative dilation rate based on high-frequency power to maximize the receptive field of each pixel while minimizing the lost frequency information.

[0019] Optionally, the modulation mechanism needs to convert the training data into the frequency domain before modulation to obtain a complex output array.

[0020] Optionally, the modulation steps of the modulation mechanism are:

[0021] Obtaining a receptive field of any region, wherein the size of the receptive field is proportional to the dilation rate;

[0022] Calculate the high frequency power, and quantify the high frequency loss of the image caused by the increase of the expansion rate according to the high frequency power;

[0023] The expansion rate is adaptively adjusted using a preset optimization function based on the receptive field and the high-frequency loss of the image.

[0024] Optionally, the expression of the preset optimization function is:

[0025]

[0026] Where RF(p) is the receptive field at point p; HP(p) is the high-frequency power at point p.

[0027] Optionally, the working principle of the adaptive core adjustment module is:

[0028] The convolution kernel parameters are decomposed into low-frequency components and high-frequency components, and a lightweight global pooling layer and a convolution layer are used to predict the dynamic weights of each channel. The ratio of the dynamic weights of each channel is dynamically adjusted according to the context so that the ratio of the low-frequency component to the high-frequency component reaches the optimal frequency ratio.

[0029] Optionally, the working principle of the frequency optimization module is:

[0030] Different masks are applied in Fourier domain to decompose the features into multiple predefined frequency bands, and the frequency components are spatially reweighted in different frequency bands dynamically to obtain the optimal dynamic frequency.

[0031] Optionally, during the training process, the diversity of the training data is increased by using data enhancement technology.

[0032] Optionally, during the training process, parameters of the initial target detection model are optimized through a loss function, where the loss function includes a target detection loss function and a regularization loss function.

[0033] A ship target detection system based on remote sensing images, comprising:

[0034] A model building module, used to build an initial target detection model, wherein the initial target detection model includes an adaptive expansion scale module, an adaptive kernel adjustment module and a frequency optimization module;

[0035] A model training module is used to obtain a training data set and input the training data set into the initial target detection model for training; the adaptive expansion scale module predicts the initial expansion rate of the received training data, and adaptively adjusts the initial expansion rate based on a modulation mechanism to obtain an optimal expansion rate of each pixel; and obtains an optimal dynamic frequency based on the adaptive kernel adjustment module and the frequency optimization module; extracts training target data based on the optimal expansion rate of each pixel and the optimal dynamic frequency, and optimizes the parameters of the initial target detection model based on the training target data and the labeled target to obtain the target detection model;

[0036] A data acquisition module, used for acquiring an image to be detected;

[0037] The target detection module is used to input the image to be detected into the trained target detection model to obtain the detection target.

[0038] It can be seen from the above technical solutions that the present invention discloses a ship target detection method and system based on remote sensing images, which has the following beneficial effects compared with the prior art:

[0039] 1) Through the combination of the adaptive expansion scale module and the dynamic kernel adjustment module, the ability to capture details of ship targets in remote sensing images is effectively enhanced. The introduction of the adaptive expansion rate enables the feature response of each pixel to be optimized according to its local context, thereby improving the adaptability to complex backgrounds and targets of different sizes. In addition, the dynamic kernel adjustment ensures high-precision detection under various interference conditions through fine control of low-frequency and high-frequency information, which not only improves the accuracy of detection, but also effectively reduces the incidence of missed detection and false detection, especially in highly complex environments.

[0040] 2) The present invention adopts a lightweight module design, which significantly reduces the computational complexity of the model while maintaining high performance. The implementation of the adaptive kernel adjustment strategy only adds a small amount of computational overhead, ensuring high efficiency in the training and reasoning stages; it can run quickly in resource-constrained computing environments, such as mobile devices or edge computing platforms, to meet practical application needs. In addition, the model's efficient parameter fine-tuning capability and flexibility enable it to adapt to different remote sensing image processing tasks, reducing the computing resources required for model training and deployment, and further improving its usability in practical scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0042] Figure 1 A schematic diagram of the overall model architecture provided by the present invention;

[0043] Figure 2 A schematic diagram of an adaptive expansion scale module provided by the present invention;

[0044] Figure 3 This is a schematic diagram of the dynamic core adjustment module provided by the present invention. DETAILED DESCRIPTION

[0045] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] The present invention discloses a ship target detection method based on remote sensing images, such as Figure 1 As shown, the specific steps are:

[0047] Target detection model construction method:

[0048] Step 1: construct an initial target detection model, which includes an adaptive expansion scale module, an adaptive kernel adjustment module and a frequency optimization module;

[0049] Step 2: Obtain a training data set, and input the training data set into the initial target detection model for training. The adaptive expansion scale module predicts the initial expansion rate of the received training data, and adaptively adjusts the initial expansion rate based on the modulation mechanism to obtain the optimal expansion rate of each pixel; and obtains the optimal dynamic frequency based on the adaptive kernel adjustment module and the frequency optimization module; extracts training target data based on the optimal expansion rate and the optimal dynamic frequency of each pixel, and optimizes the parameters of the initial target detection model based on the training target data and the labeled target to obtain the target detection model;

[0050] Target detection methods:

[0051] Step 3: Obtain the image to be detected;

[0052] Step 4: Input the image to be detected into the trained target detection model to obtain the detection target.

[0053] In one embodiment, the steps for constructing the initial target detection model are as follows: first, a backbone network based on EfficientNet is constructed, and adaptive dilated convolution technology is integrated, and an adaptive kernel adjustment strategy and a frequency selection optimization strategy are introduced.

[0054] In one embodiment, if Figure 2 As shown, the adaptive expansion scale module includes a convolution layer, an activation function layer and a modulation mechanism;

[0055] Convolutional layers, used to predict the initial dilation rate for each pixel in the training data;

[0056] The activation function layer is used to process the initial expansion rate to obtain a non-negative expansion rate;

[0057] A modulation mechanism is used to dynamically adjust the non-negative dilation rate based on the high-frequency power to maximize the receptive field of each pixel while minimizing the lost frequency information.

[0058] In one embodiment, the modulation mechanism needs to convert the training data into the frequency domain before modulation to obtain a complex output array.

[0059] In one embodiment, the modulation steps of the modulation mechanism are:

[0060] Get the receptive field of any region, the size of the receptive field is proportional to the expansion rate;

[0061] Calculate the high frequency power, and quantify the high frequency loss of the image caused by the increase of the expansion rate according to the high frequency power;

[0062] The expansion rate is adaptively adjusted using a preset optimization function based on the receptive field and high-frequency loss of the image.

[0063] Furthermore, the adaptive expansion scale module balances the large receptive field and effective bandwidth of the expanded convolution by assigning different expansion rates to each pixel; then, by introducing the activation function layer and the modulation mechanism, the non-negativity of the expansion rate is ensured, and the receptive field of each pixel is maximized while minimizing the lost frequency information. In this embodiment, the lost frequency information is measured by calculating the high-frequency power, and the threshold is optimized to maximize the receptive field and minimize the loss of high-frequency power. Finally, this is achieved by directly optimizing the variable of the adaptive expansion rate. It ensures that the network can effectively capture high-frequency detail information while expanding the receptive field, laying a solid foundation for subsequent target detection tasks.

[0064] Furthermore, in the adaptive expansion scale module, the image to be detected is first transformed using discrete Fourier transform (DFT) Convert to frequency domain and get X DFT , whose expression is:

[0065]

[0066] In the formula, X DFT is a real matrix of size H×W, representing the complex output array from DFT; H and W represent the height and width of the image to be detected, respectively; h and ω represent the coordinates of the image X to be detected; u represents the frequency component in the horizontal direction in the frequency domain, usually corresponding to the horizontal frequency information of the image or signal, and v represents the frequency component in the vertical direction in the frequency domain, usually corresponding to the vertical frequency information of the image or signal; the normalized frequencies in the height and width directions are given by |u| and |v|.

[0067] Then, considering that the input feature map has spatial variability, the optimal expansion rate for each pixel may be different. Therefore, an adaptive expansion rate is introduced to achieve a better balance. This strategy assigns a different expansion rate to each pixel, and its expression is:

[0068]

[0069] Where Y(p) is the pixel value at position p in the output feature map; D(p) is an adaptive dilation rate (variable), which may be different at different pixels p. It can be predicted by convolutional neural networks and reflects the expansion or enlargement of a local area at a certain position; i and j are integers from 1 to K, indicating the position in this summation operation. Because there are two nested sums in the formula, they are used to traverse the elements in a certain space; K is the kernel size, W i is the weight parameter of the kernel; Δp i and Δp j are offsets used to adjust the position of local regions in an image, and they can be viewed as adjustments to the spatial position in the input image.

[0070] The ReLU layer is introduced to ensure the non-negativity of the expansion rate, and a modulation mechanism is adopted to maximize the receptive field and minimize the frequency information lost for each pixel; the relationship between the adaptive expansion rate and the receptive field and the frequency information loss is specifically: it first defines the local feature centered on p, and the feature with a window size of s is X(p,s). The receptive field (RF) of the local feature is proportional to the expansion rate D(p), and the size of the receptive field increases with the increase of the expansion rate. Then, when the expansion rate increases, the high-frequency information of the image becomes difficult to capture accurately, especially the high-frequency components within the frequency range. This is due to the limitation of the expansion operation, which causes some frequency components to be "lost". In order to quantify this lost high-frequency information, the high-frequency power (HP) is introduced, which measures the lost information by calculating the power within a specific frequency range. The high-frequency power is calculated by integrating the high-frequency area in the frequency domain and calculating its energy, thereby measuring the amount of high-frequency information lost. Finally, the goal of optimizing the expansion rate function θ is to minimize the loss of this frequency information, thereby improving the model's perception of local features while maintaining effective frequency information capture capabilities. This process optimizes the performance of the model by adjusting the expansion rate D(p).

[0071] Therefore, the goal of optimizing θ can be written as Through this optimization, the optimal dilation rate for each pixel can be adaptively selected while maximizing the effectiveness of its receptive field, expressed as:

[0072]

[0073] However, due to the frequency set Due to the discreteness of , and the fact that the computation of HP is not differentiable, direct optimization may be impractical. Therefore, we choose to directly optimize D(p), that is, increase the dilation rate at p positions with low HP(p) to encourage a large receptive field, and suppress the dilation rate at positions with high HP(p) to reduce the loss of frequency information.

[0074] In one embodiment, if Figure 3 As shown in the figure, the working principle of the adaptive core adjustment module is:

[0075] The convolution kernel parameters are decomposed into low-frequency components and high-frequency components. The lightweight global pooling layer and convolution layer are used to predict the dynamic weights of each channel. The proportion of the dynamic weights of each channel is dynamically adjusted according to the context so that the ratio of low-frequency components to high-frequency components reaches the optimal frequency ratio.

[0076] Furthermore, the effective bandwidth of the network is further enhanced by an adaptive kernel adjustment strategy, which decomposes the convolution kernel parameters into low-frequency and high-frequency components and adjusts the frequency response by dynamic weighting, and then dynamically adjusts to balance the high-frequency and low-frequency components in the feature representation, thereby enhancing the receptive field. Different masks are applied in the Fourier domain to decompose the features into different frequency bands, dynamically re-weighting the frequency components of different frequency bands spatially.

[0077] Furthermore, in order to further improve the effective bandwidth, the convolution kernel parameters are decomposed into low-frequency and high-frequency components, and then dynamic weighting is introduced to adjust the frequency response. This process adds only a few additional parameters and computational overhead. For a static convolution kernel, its weight W can be decomposed into:

[0078]

[0079] Here, represents the average W according to the kernel weight. This average weight acts as a low-pass K×K mean filter, followed by a parameter represented by The 1×1 convolution defined. After decomposition, the high-frequency and low-frequency components are dynamically adjusted, and the expression is:

[0080]

[0081] In the formula, λ l ,λ h Dynamic weights for each channel are predicted by a simple lightweight global pooling + convolution layer, and λ is dynamically adjusted according to the input context. l ,λ hThe ratio of , enables the network to focus on specific frequency bands and adapt to the complexity of visual patterns in features. This dynamic frequency adaptation method enhances the network's ability to capture low-frequency context and high-frequency local details. This in turn increases the effective bandwidth, thereby improving the performance of segmentation tasks that require extracting different features at different frequencies.

[0082] In one embodiment, the working principle of the frequency optimization module is:

[0083] Different masks are applied in Fourier domain to decompose the features into multiple predefined frequency bands, and the frequency components are spatially reweighted in different frequency bands dynamically to obtain the optimal dynamic frequency.

[0084] Furthermore, the features are decomposed into different frequency bands by applying different masks in the Fourier domain:

[0085]

[0086] where F -1 represents inverse fast Fourier transform, M b is a binary mask used to extract the corresponding frequencies, and its expression is:

[0087]

[0088] Here, φ b and φ b+1 From a set of B+1 predefined frequency thresholds Then, the frequency components are spatially reweighted dynamically in different frequency bands, and the expression is:

[0089]

[0090] In the formula, is the learned frequency balance feature after processing, A selection diagram showing the b-th frequency band.

[0091] In one embodiment, during the training process, the diversity of training data is increased through data enhancement technology.

[0092] In one embodiment, during the training process, parameters of the initial target detection model are optimized using a loss function, where the loss function includes a target detection loss function and a regularization loss function.

[0093] Furthermore, the EfficientNet network integrating the modules and strategies in the first two steps is trained and optimized end-to-end. During the training process, the labeled remote sensing image dataset is used, and the network parameters are adjusted to minimize the detection error through the back propagation algorithm and the appropriate loss function. In addition, data enhancement, regularization and other techniques are used to improve the generalization ability and robustness of the model. After the training is completed, the model is fine-tuned to adapt to the specific remote sensing image ship target detection task to ensure high efficiency and accuracy in practical applications.

[0094] Specifically: During the training process, data augmentation techniques are used to increase the diversity of training samples, such as random cropping, rotation, scaling, and color transformation, to improve the generalization and robustness of the model. These techniques can be implemented through the following operations:

[0095] O′=Augment(O);

[0096] In the formula, O is the original input image, and O′ is the enhanced image, which is Figure 1 After the initial convolution layer, normalization and activation function, the input feature X is obtained, with a size of C×H×W. Then after the above operations, we get Figure 1 The output feature Y in the moving convolution block has a size of C×H×W. In the moving convolution block, a depthwise separable convolution operation is used. Unlike standard convolution, the depthwise separable convolution first performs a convolution operation on each input channel separately without involving information mixing between channels. Each channel is processed using an independent convolution kernel, which can greatly reduce the amount of calculation. Next, a point convolution (usually a 1x1 convolution) is performed to mix the features of different channels, thereby further enhancing the expressiveness of the model. After the moving convolution block is processed, the network uses a global average pooling operation to perform an average calculation on each channel in the spatial dimension. This operation effectively compresses the spatial information, retaining only the global information of each channel, and generates a feature vector of size 1×1×C. Next, the feature vector enters the fully connected layer and is further processed to obtain the final feature representation. Finally, the output of the fully connected layer is converted into a probability distribution through the Softmax layer, thereby representing the predicted probability p of each category.

[0097] The EfficientNet network integrating the above modules and strategies is trained and optimized end-to-end. During the training process, the labeled remote sensing image dataset is used to adjust the network parameters to minimize the detection error through the back propagation algorithm and the appropriate loss function. The loss function can be expressed as:

[0098] L=L det +αL reg ;

[0099] Where, Ldet is the target detection loss, L reg is the regularization loss, and α is the weight hyperparameter that balances the two.

[0100]

[0101] Where N is the number of samples, y i is the true label, p i is the predicted probability, L bbox is the bounding box regression loss, λ bbox is a weight hyperparameter that balances classification loss and regression loss.

[0102] In addition, in order to prevent overfitting, the L2 regularization strategy is used to limit the complexity of the network, which can be expressed as:

[0103]

[0104] Where W is the weight parameter of the network and λ is the regularization strength hyperparameter.

[0105] After the training is completed, the model is fine-tuned to adapt to the specific remote sensing image ship target detection task to ensure high efficiency and accuracy in practical applications. During the fine-tuning process, a smaller learning rate is used for further training to refine the model parameters. The formula can be expressed as:

[0106]

[0107] In the formula, θ new and θ old Denote the fine-tuned parameters and the original parameters respectively, and η is the learning rate during fine-tuning. Through the above steps, the remote sensing image ship target detection method of the present invention not only improves the detection accuracy, but also maintains high efficiency during model training and deployment, and can be applied to remote sensing image processing tasks in a variety of computing environments.

[0108] On the other hand, this embodiment discloses a ship target detection system based on remote sensing images, including:

[0109] A model building module is used to build an initial target detection model, which includes an adaptive expansion scale module, an adaptive kernel adjustment module and a frequency optimization module;

[0110] The model training module is used to obtain a training data set and input the training data set into the initial target detection model for training. The adaptive expansion scale module predicts the initial expansion rate of the received training data and adaptively adjusts the initial expansion rate based on the modulation mechanism to obtain the optimal expansion rate of each pixel; and obtains the optimal dynamic frequency based on the adaptive kernel adjustment module and the frequency optimization module; extracts the training target data based on the optimal expansion rate and the optimal dynamic frequency of each pixel, optimizes the parameters of the initial target detection model based on the training target data and the labeled target, and obtains the target detection model;

[0111] A data acquisition module, used for acquiring an image to be detected;

[0112] The target detection module is used to input the image to be detected into the trained target detection model to obtain the detection target.

[0113] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0114] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A ship target detection method based on remote sensing images, characterized in that: The specific steps are: Target detection model construction method: Constructing an initial target detection model, wherein the initial target detection model includes an adaptive expansion scale module, an adaptive kernel adjustment module and a frequency optimization module; A training data set is obtained, and the training data set is input into the initial target detection model for training, the adaptive expansion scale module predicts the initial expansion rate of the received training data, and adaptively adjusts the initial expansion rate based on the modulation mechanism to obtain the optimal expansion rate of each pixel; and the optimal dynamic frequency is obtained based on the adaptive kernel adjustment module and the frequency optimization module; training target data is extracted based on the optimal expansion rate of each pixel and the optimal dynamic frequency, and the initial target detection model is optimized based on the training target data and the labeled target to obtain the target detection model; Target detection methods: Acquire the image to be detected; The image to be detected is input into the trained target detection model to obtain the detection target.

2. The method for detecting ship targets based on remote sensing images according to claim 1, characterized in that: The adaptive expansion scale module includes a convolution layer, an activation function layer and a modulation mechanism; The convolution layer is used to predict the initial expansion rate of each pixel in the training data; The activation function layer is used to perform non-negative processing on the initial expansion rate to obtain a non-negative expansion rate; The modulation mechanism is used to dynamically adjust the non-negative dilation rate based on high-frequency power to maximize the receptive field of each pixel while minimizing the lost frequency information.

3. The ship target detection method based on remote sensing images according to claim 2 is characterized in that: The modulation mechanism also needs to convert the training data into the frequency domain before modulation to obtain a complex output array.

4. The method for detecting ship targets based on remote sensing images according to claim 3 is characterized in that: The modulation steps of the modulation mechanism are: Obtaining a receptive field of any region, wherein the size of the receptive field is proportional to the dilation rate; Calculate the high frequency power, and quantify the high frequency loss of the image caused by the increase of the expansion rate according to the high frequency power; The expansion rate is adaptively adjusted using a preset optimization function based on the receptive field and the high-frequency loss of the image.

5. The method for detecting ship targets based on remote sensing images according to claim 2, characterized in that: The expression of the preset optimization function is: Where RF(p) is the receptive field of point p; HP(p) is the high frequency power at point p.

6. The method for detecting ship targets based on remote sensing images according to claim 1, characterized in that: The working principle of the adaptive core adjustment module is: The convolution kernel parameters are decomposed into low-frequency components and high-frequency components, and a lightweight global pooling layer and a convolution layer are used to predict the dynamic weights of each channel. The ratio of the dynamic weights of each channel is dynamically adjusted according to the context so that the ratio of the low-frequency component to the high-frequency component reaches the optimal frequency ratio.

7. The method for detecting ship targets based on remote sensing images according to claim 1, characterized in that: The working principle of the frequency optimization module is: Different masks are applied in Fourier domain to decompose the features into multiple predefined frequency bands, and the frequency components are spatially reweighted in different frequency bands dynamically to obtain the optimal dynamic frequency.

8. The method for detecting ship targets based on remote sensing images according to claim 1, characterized in that: During the training process, the diversity of the training data is increased through data enhancement techniques.

9. The ship target detection method based on remote sensing images according to claim 1 is characterized in that: During the training process, the parameters of the initial target detection model are optimized through a loss function, and the loss function includes a target detection loss function and a regularization loss function.

10. A ship target detection system based on remote sensing images, characterized in that: include: A model building module, used to build an initial target detection model, wherein the initial target detection model includes an adaptive expansion scale module, an adaptive kernel adjustment module and a frequency optimization module; A model training module is used to obtain a training data set and input the training data set into the initial target detection model for training. The adaptive expansion scale module predicts the initial expansion rate of the received training data and adaptively adjusts the initial expansion rate based on a modulation mechanism to obtain an optimal expansion rate for each pixel; and obtains an optimal dynamic frequency based on the adaptive kernel adjustment module and the frequency optimization module; extracts training target data based on the optimal expansion rate of each pixel and the optimal dynamic frequency, and optimizes the parameters of the initial target detection model based on the training target data and the labeled target to obtain the target detection model; A data acquisition module, used for acquiring an image to be detected; The target detection module is used to input the image to be detected into the trained target detection model to obtain the detection target.

Images