A method and apparatus for detecting small scale targets at sea

Abstract

The application belongs to the field of real-time detection of marine targets, and relates to a marine small-scale target detection method and device, which comprises the following steps: performing preprocessing on an original image to obtain a first image and a second image, training an original yolov3-tiny network model to obtain a first-level yolov3-tiny network model and a second-level yolov3-tiny network model, inputting the first image into the first-level yolov3-tiny network model to obtain a first-level feature map, constructing an fdnet module, the fdnet module being used for predicting a region of interest image that needs to be cropped on the second image, inputting the region of interest image into the second-level yolov3-tiny network model to obtain a second-level feature map, and the first-level feature map and the second-level feature map forming a final feature map, the yolov3-tiny network model is expanded into two levels in the application, and an fdnet module is added after the fifth layer convolution pooling of the first-level yolov3-tiny network model, so that the accuracy of the yolov3-tiny network model in marine small-scale target detection is improved.

Description

Technical Field

[0001] This invention belongs to the field of real-time detection of maritime targets, and more specifically, relates to a method and apparatus for detecting small-scale maritime targets. Background Technology

[0002] Real-time small-scale target detection technology at sea is an important branch of the target detection field, with wide applications in both military and civilian sectors. In the military field, it is often used for precision guidance of missile electro-optical guidance systems, target detection by shipborne electro-optical detection systems, target detection by airborne electro-optical detection systems, and collaborative operations of sea-air unmanned cooperative systems. In the civilian field, it is often used for accident rescue, marine exploration, and fisheries management.

[0003] Currently, real-time small-scale target detection in the ocean is mainly based on four imaging methods: radar imaging, satellite remote sensing imaging, infrared imaging, and visible light imaging. Target detection based on visible light imaging has high resolution because the target image contains rich color, texture, and spectral information. It has good actual detection effect, high cost performance, and is easy to implement and popularize. By adding preprocessing operations such as image dehazing and histogram equalization to visible light images, problems such as sea surface fog, strong light interference, and sea-sky background fusion can be solved, making it a reliable detection and tracking solution for small-scale targets at sea.

[0004] At present, target detection and tracking technology based on visible light imaging is developing rapidly. The real-time target detection method for small scale at sea based on the yolov3-tiny network model has good real-time performance, but it also has problems such as a small number of available pixels, unclear visualization features, and low accuracy. Summary of the Invention

[0005] The technical problem that this invention needs to solve is that the number of available pixels for small-scale target detection at sea is small, the visualization features are not obvious, and the accuracy of small-scale target detection at sea is not high.

[0006] The present invention solves the above problems through the following technical solution:

[0007] Firstly, a method for detecting small-scale targets at sea is provided, including:

[0008] The original image is preprocessed to obtain the first image and the second image;

[0009] After training the original yolov3-tiny network model, we obtain the first-level yolov3-tiny network model and the second-level yolov3-tiny network model.

[0010] The first image is input into the first-level yolov3-tiny network model to obtain the first-level feature map;

[0011] Construct an fdnet module, which is used to predict the region of interest image to be cropped in the second image;

[0012] The region of interest image is input into the second-level yolov3-tiny network model to obtain the second-level feature map. The first-level feature map and the second-level feature map together form the final feature map.

[0013] Preferably, the preprocessing of the original image to obtain the first image and the second image specifically involves:

[0014] The collected raw images are scaled up to a first image with a first preset resolution and a second image with a second preset resolution.

[0015] Preferably, the process of training the original yolov3-tiny network model to obtain a first-level yolov3-tiny network model and a second-level yolov3-tiny network model includes:

[0016] The original dataset was evaluated based on the publicly available SeaShips ship target detection and recognition dataset to obtain the ship data training set;

[0017] Using the aforementioned ship data training set, the original yolov3-tiny network model is trained to obtain the first-level yolov3-tiny network model and the second-level yolov3-tiny network model after training.

[0018] Preferably, the step of inputting the first image into the first-level yolov3-tiny network model to obtain the first-level feature map includes:

[0019] The first-level feature map includes a first feature map with a first resolution and a second feature map with a second resolution, and the first feature map and the second feature map are part of the final feature map;

[0020] After the first image is input into the first-level yolov3-tiny network model, the output of the fifth convolutional pooling layer of the first-level yolov3-tiny network model is obtained as the third image, which is then used as input to the next module.

[0021] Preferably, the construction of the fdnet module, which is used to predict the region of interest image to be cropped in the second image, includes:

[0022] The third image is input into the fdnet module;

[0023] After the third image is input into the fdnet module, the region of interest image on the second image that needs to be input into the second-level yolov3-tiny network model is determined based on the output of the fdnet module.

[0024] Preferably, the step of inputting the region of interest image into the second-level YOLOv3-tiny network model to obtain a second-level feature map, wherein the first-level feature map and the second-level feature map constitute the final feature map, including:

[0025] The second-level feature map includes a third feature map at a first resolution and a fourth feature map at a second resolution, wherein the third feature map and the fourth feature map are another part of the final feature map;

[0026] Based on the information contained in the first-level feature map and the second-level feature map, the specific information of the target object in the original image is detected.

[0027] Secondly, a small-scale target detection device for sea is provided, comprising: a raw image preprocessing module, a YOLOv3-tiny network model training module, an FDNet module construction module, and a feature map acquisition module, wherein:

[0028] The original image preprocessing module is used to scale the original image into a first image with a first preset resolution and a second image with a second preset resolution;

[0029] The yolov3-tiny network model training module is used to train the original yolov3-tiny network model to obtain the trained first-level yolov3-tiny network model and the second-level yolov3-tiny network model.

[0030] The fdnet module construction module is used to construct the fdnet module to predict the region of interest image that needs to be cropped on the second image;

[0031] The feature map acquisition module is used to acquire feature maps for detection of the original image.

[0032] Preferably, the yolov3-tiny network model training module includes a ship data training set construction unit and a yolov3-tiny network model training unit;

[0033] The ship data training set construction unit is used to evaluate the original dataset based on the SeaShips ship target detection and recognition public dataset to obtain the ship data training set;

[0034] The yolov3-tiny network model training unit is used to train the original yolov3-tiny network model based on the ship data training set.

[0035] Preferably, the fdnet module construction module is used to construct the fdnet module to predict the region of interest image to be cropped on the second image, including:

[0036] The output of the fifth convolutional pooling layer of the first-level yolov3-tiny network model is used as the input of the fdnet module. Based on the output of the fdnet module, the region of interest image on the second image that needs to be input into the second-level yolov3-tiny network model is determined.

[0037] Preferably, the feature map acquisition module is used to acquire feature maps by inputting the first image into the first-level yolov3-tiny network model to obtain a first-level feature map, and inputting the region of interest image into the second-level yolov3-tiny network model to obtain a second-level feature map;

[0038] The specific information of the target object in the original image is detected by analyzing the first-level feature map and the second-level feature map.

[0039] The beneficial effects of this invention are:

[0040] Based on the lightweight YOLOv3-tiny network model, and by analyzing the characteristics of small-scale targets at sea, the YOLOv3-tiny network model is expanded into a two-level model. The two-level YOLOv3-tiny network models share parameters, and an FDNet module is added after the fifth convolutional pooling layer of the first-level YOLOv3-tiny network model to predict and crop regions containing small targets in the original image, thereby improving the accuracy of the YOLOv3-tiny network model in detecting small-scale targets at sea. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. The drawings are used to provide a further understanding of this disclosure and constitute a part of the specification. They are used together with the following detailed description to explain this disclosure, but do not constitute a limitation of this disclosure. In the drawings:

[0042] Figure 1This is a schematic diagram of a method for detecting small-scale targets at sea provided by an embodiment of the present invention;

[0043] Figure 2 This is a schematic diagram of the yolov3-tiny network model architecture in a small-scale target detection method at sea provided by an embodiment of the present invention;

[0044] Figure 3 This is a schematic diagram of the training process of the yolov3-tiny network model in a method for detecting small-scale targets at sea provided in an embodiment of the present invention.

[0045] Figure 4 This is a schematic diagram of the fdnet module architecture in a small-scale target detection method at sea provided by an embodiment of the present invention;

[0046] Figure 5 This is a schematic diagram of the parameters of the convolutional layer of the fdnet module in a method for detecting small-scale targets at sea provided in an embodiment of the present invention;

[0047] Figure 6 This is a schematic diagram of the overall process of a method for detecting small-scale targets at sea provided in an embodiment of the present invention;

[0048] Figure 7 This is a schematic diagram of the structure of a small-scale target detection device for sea provided in an embodiment of the present invention;

[0049] Figure 8 This is a schematic diagram of the structure of the yolov3-tiny network model training module in a small-scale maritime target detection device provided in an embodiment of the present invention;

[0050] Figure 9 This is a schematic diagram of the architecture of a small-scale target detection device for the sea provided in an embodiment of the present invention. Detailed Implementation

[0051] 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 embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0052] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0053] It should also be noted that, in this specification, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0054] Example 1

[0055] At present, target detection and tracking technology based on visible light imaging is developing rapidly. The real-time target detection method for small-scale targets at sea based on the yolov3-tiny network model has good real-time performance. However, the number of available pixels for small-scale target detection at sea at the image acquisition end is small, the visualization features are not obvious, and the accuracy of small-scale target detection at sea is not high.

[0056] Based on the above issues, such as Figure 1 As shown in the figure, this embodiment proposes a method for detecting small-scale targets at sea, including:

[0057] Step 101: Preprocess the original image to obtain the first image and the second image.

[0058] Real-time raw images of the sea surface are obtained through marine image acquisition equipment. The collected raw images are scaled up to a first image with a first preset resolution and a second image with a second preset resolution. Specifically, the raw images are scaled up using existing technology to obtain two images with resolutions of 416x416 (first preset resolution) and 1664x1664 (second preset resolution). The raw images are scaled up to two resolutions, 1664x1664 and 416x416, and then processed by neural networks because the design of the YOLOv3-tiny network model requires the detection of targets at different scales and computational efficiency.

[0059] Multi-scale detection: The yolov3-tiny network model uses multiple detection layers, each corresponding to a different output scale. Larger input resolutions (such as 1664x1664) can better detect smaller targets, while smaller input resolutions (such as 416x416) are suitable for detecting larger targets. This multi-scale detection design enables the model to detect targets of different sizes more comprehensively and accurately.

[0060] Computational efficiency: A larger input image resolution results in more pixels and parameters, which increases computation and memory consumption. To be more computationally efficient, a smaller input resolution (416x416) is provided, which can reduce computational costs to some extent and make the model more suitable for running on embedded devices or in environments with limited computing resources.

[0061] Step 102: For the original yolov3-tiny network model (e.g.) Figure 2 After training (as shown), the first-level yolov3-tiny network model and the second-level yolov3-tiny network model are obtained.

[0062] Specifically, it includes two steps, such as Figure 3 As shown, it includes:

[0063] Step 1021: The original dataset is evaluated based on the SeaShips ship target detection and recognition public dataset to obtain the ship data training set.

[0064] The hardware platform used for deployment was the Jetson Xavier NX edge computing platform. The training and testing datasets were extracted from the SeaShips dataset, which consists of 31,455 images covering six common ship types (ore carriers, bulk carriers, general cargo ships, container ships, fishing boats, and passenger ships). All images were obtained from approximately 10,800 real-world video clips acquired by surveillance cameras deployed in the coastline video surveillance system.

[0065] Download the original files of the SeaShips dataset by visiting its official website or relevant open data platforms. Evaluate the SeaShips dataset using appropriate evaluation metrics, such as mean squared accuracy. Then, preprocess the original dataset to meet training requirements. This may include image format conversion, data augmentation, and annotation format conversion.

[0066] Based on training requirements, the dataset is divided into a training set and a test set. The test set consists of 1,000 images, and the training set consists of 6,000 images. The training set is labeled using appropriate labeling tools to label the ship targets in each image, ensuring that the labels are accurate and consistent. Finally, the evaluated ship data training set is stored in an appropriate location, and a good data management system is established to facilitate subsequent training and use.

[0067] Step 1022: Using the aforementioned ship data training set, train the original YOLOv3-tiny network model to obtain the trained first-level YOLOv3-tiny network model and the second-level YOLOv3-tiny network model. This mainly includes the following:

[0068] (1) Data preparation: Ensure that the ship data training set has been prepared and that each image has been labeled accordingly, including the ship's bounding box and category information.

[0069] (2) Download the yolov3-tiny network model: Download the initial weight file of the yolov3-tiny network model from the open source code repository or related resources. These weight files can be loaded into the model as initial parameters before training begins.

[0070] (3) Model configuration: Create an appropriate model configuration file to define the network structure, hyperparameters and training settings of yolov3-tiny, including defining the input image size, number of classes, anchor box settings, etc.

[0071] (4) Data preprocessing: According to the requirements of the yolov3-tiny model, the ship data training set is preprocessed, including image size adjustment, normalization, data augmentation and other operations.

[0072] (5) Model training: Using the ship data training set and model configuration file, start training the yolov3-tiny network model. During the training process, the model will gradually optimize the weight parameters according to the input data and annotation information to achieve the detection and recognition of ship targets.

[0073] (6) Model evaluation and tuning: During the training process, the performance of the model is evaluated periodically using the validation set. Based on the evaluation results, the model is tuned. Hyperparameters such as learning rate, batch size, and number of training rounds can be adjusted to obtain better model performance.

[0074] It is worth noting that there is no difference between the first-level yolov3-tiny network model and the second-level yolov3-tiny network model. The weights are obtained by training the original yolov3-tiny network to distinguish them.

[0075] Step 103: Input the first image into the first-level yolov3-tiny network model to obtain the first-level feature map.

[0076] The first-level feature map includes a first feature map with a first resolution and a second feature map with a second resolution, and the first feature map and the second feature map are part of the final feature map.

[0077] The first resolution is 13x13x(n+5)x3, and the second resolution is 26x26x(n+5)x3, where n represents the number of object categories to be detected. Since the images in the training dataset and the test dataset are both from the SeaShips dataset, which is divided into 6 types, n=6 at this time. The first feature map and the second feature map are used as part of the final feature map.

[0078] To facilitate the implementation of subsequent methods, in a preferred embodiment, after the first image is input into the first-level yolov3-tiny network model, the output of the fifth convolutional pooling layer of the first-level yolov3-tiny network model is obtained as a third image, which is used as input to the next module.

[0079] For the YOLOv3-tiny network model, its fifth layer is a convolutional pooling layer. When the first image is input into the first-level YOLOv3-tiny network model, a new image, called the third image, can be obtained by extracting the output of the fifth layer. The third image can be used as input to the next-level module, and subsequent operations can improve the accuracy and precision of detection.

[0080] Step 104: Construct the fdnet module (its structure is as follows) Figure 4 As shown in the figure, the fdnet module is used to predict the region of interest image that needs to be cropped on the second image.

[0081] Specifically, the third image is input into the fdnet module. After the third image is input into the fdnet module, the region of interest image on the second image that needs to be input into the second-level yolov3-tiny network model is determined based on the output of the fdnet module.

[0082] The fdnet module internally includes two convolutional kernels and one pooling layer. Its convolutional layer parameters are as follows: Figure 5As shown, the second image is divided into 4x4 small regions. The cropping region is predicted by the FDNet module, using the coordinates of the region center and a confidence value indicating whether cropping is needed. The FDNet module predicts the cropping region using the same bounding box prediction method as the YOLOv3-tiny network model (the specific method is not elaborated in this embodiment). The second image is divided into 4x4 regions, and the center point of each cropping region falls within a certain region. This region is responsible for predicting the cropping module. The output layer of the FDNet module is 4x4x(2+1), where 4x4 represents the original image divided into 16 regions, 2 represents the coordinates of the predicted cropping region, and 1 represents whether cropping is needed. The cropping region predicted by the FDNet module is cropped into several 416x416 region images on the second image with a resolution of 1664x1664.

[0083] Step 105: Input the region of interest image into the second-level yolov3-tiny network model to obtain the second-level feature map. The first-level feature map and the second-level feature map together form the final feature map.

[0084] The second-level feature map includes a third feature map at a first resolution and a fourth feature map at a second resolution. The third and fourth feature maps are used as another part of the final feature map. The image with a cropped region resolution of 416x416 predicted by the fdnet module is fed back into the second-level yolov3-tiny network model as input. The first resolution is 13x13x(n+5)x3 and the second resolution is 26x26x(n+5)x3, where n represents the number of object categories to be detected. Since the training and test datasets are both from the publicly available SeaShips dataset, which is divided into 6 categories, n=6 at this time. The third and fourth feature maps are used as part of the final feature map.

[0085] Based on the information contained in the first-level feature map and the second-level feature map, the specific information of the target object in the original image is detected. Specifically, the first-level yolov3-tiny network model obtains a first feature map with a resolution of 13x13x33 and a second feature map with a resolution of 26x26x33. The second-level yolov3-tiny network model obtains a third feature map with a resolution of 13x13x33 and a fourth feature map with a resolution of 26x26x33. The first, second, third, and fourth feature maps together constitute the final feature map. The detection of the target object can be completed by extracting and analyzing the content of these four feature maps.

[0086] This embodiment provides a method for detecting small-scale targets at sea. The overall framework of the detection method is as follows: Figure 6 As shown, based on the lightweight YOLOv3-tiny network model, by analyzing the characteristics of small-scale targets at sea, the YOLOv3-tiny network model is expanded into a two-level model. The two-level YOLOv3-tiny network models share parameters, and an FDNet module is added after the fifth convolutional pooling layer of the first-level YOLOv3-tiny network model to predict and crop regions containing small targets in the original image, thereby improving the accuracy of the YOLOv3-tiny network model in detecting small-scale targets at sea.

[0087] Example 2:

[0088] Example 1 presented a method for detecting small-scale targets at sea. This example will present a device for detecting small-scale targets at sea, which is applicable to the detection method described in Example 1. Figure 7 As shown, it includes: a raw image preprocessing module, a YOLOv3-Tiny network model training module, an FDNET module construction module, and a feature map acquisition module, wherein:

[0089] The original image preprocessing module is used to scale the original image into a first image with a first preset resolution and a second image with a second preset resolution. The yolov3-tiny network model training module is used to train the original yolov3-tiny network model to obtain a trained first-level yolov3-tiny network model and a second-level yolov3-tiny network model. The fdnet module construction module is used to construct the fdnet module to predict the region of interest image that needs to be cropped on the second image. The feature map acquisition module is used to acquire feature maps to detect the original image.

[0090] The original image preprocessing module is used to preprocess the original image, scaling it to a first image with a first preset resolution and a second image with a second preset resolution. The original image is scaled using existing technology to obtain two images with resolutions of 416x416 (first preset resolution) and 1664x1664 (second preset resolution). The original image is scaled to two resolutions, 1664x1664 and 416x416, before neural network processing is performed because the design of the YOLOv3-tiny network model requires the detection of targets at different scales and computational efficiency.

[0091] The original image preprocessing module includes operations such as image resizing, normalization, and data augmentation to prepare the original image for input into the YOLOv3-Tiny network model for detection.

[0092] It is worth noting that, as described in Example 1, in this example, the first-level yolov3-tiny network model and the second-level yolov3-tiny network model are no different. The weights are obtained by training the original yolov3-tiny network only to distinguish them.

[0093] In a preferred embodiment, refer to Figure 8 The yolov3-tiny network model training module includes a ship data training set construction unit and a yolov3-tiny network model training unit. The ship data training set construction unit is used to obtain a ship data training set by evaluating the original dataset based on the SeaShips ship target detection and recognition public dataset. The yolov3-tiny network model training unit is used to train the original yolov3-tiny network model based on the ship data training set.

[0094] The specific method for obtaining the ship data training set by evaluating the original dataset based on the publicly available SeaShips ship target detection and recognition dataset is as follows: download the original dataset file by visiting the official website of the SeaShips dataset or related open data platforms, evaluate the SeaShips dataset using appropriate evaluation metrics, such as average accuracy, and then preprocess the original dataset to meet training requirements. This may include image format conversion, data augmentation, annotation format conversion, etc.

[0095] Based on training requirements, the dataset is divided into a training set and a test set. The test set consists of 1,000 images, and the training set consists of 6,000 images. The training set is labeled using appropriate labeling tools to label the ship targets in each image, ensuring that the labels are accurate and consistent. Finally, the evaluated ship data training set is stored in an appropriate location, and a good data management system is established to facilitate subsequent training and use.

[0096] The method for training the original yolov3-tiny network model has been specifically described in Example 1, and will not be repeated in this example.

[0097] In a preferred embodiment, the output of the fifth convolutional pooling layer of the first-level yolov3-tiny network model is used as the input of the fdnet module. Based on the output of the fdnet module, the region of interest image on the second image that needs to be input to the second-level yolov3-tiny network model is determined.

[0098] The fdnet module construction module is used to construct the fdnet module. The main purpose of constructing the fdnet module is to predict the region of interest image that needs to be cropped on the second image. By predicting the second image, the region that needs to be cropped on the second image can be obtained. Then, the images of these regions are input into the second-level yolov3-tiny network model to obtain the second-level feature map. The method of the fdnet module predicting the region of interest image is described in Example 1, and will not be elaborated on here.

[0099] In a preferred embodiment, the feature map acquisition module is used to acquire feature maps by inputting the first image into the first-level yolov3-tiny network model to obtain a first-level feature map, and inputting the region of interest image into the second-level yolov3-tiny network model to obtain a second-level feature map.

[0100] The specific information of the target object in the original image is detected by analyzing the first-level feature map and the second-level feature map.

[0101] The feature map acquisition module is used to acquire feature maps for detection of the original image. In the first-level YOLOv3-tiny network model and the second-level YOLOv3-tiny network model, the feature maps are obtained after the original image has completely passed through the first-level YOLOv3-tiny network model and the second-level YOLOv3-tiny network model. They are used to identify targets of different sizes and positions. By acquiring these feature maps, target detection and recognition can be performed, and the corresponding bounding boxes and category information can be obtained. The resolution of the specific acquired feature maps has been described in Example 1 and will not be elaborated further in this example.

[0102] Example 3

[0103] like Figure 9 The diagram shown is a schematic representation of the architecture of a small-scale maritime target detection device proposed in this embodiment of the invention. The device includes one or more processors 21 and a memory 22. Figure 9 Take a processor 21 as an example.

[0104] Processor 21 and memory 22 can be connected via a bus or other means. Figure 9 Taking the example of a connection between China and Israel via a bus.

[0105] The memory 22, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs and non-volatile computer-executable programs, such as the small-scale target detection method at sea described in Embodiment 1. The processor 21 executes the small-scale target detection method at sea by running the non-volatile software programs and instructions stored in the memory 22.

[0106] The memory 22 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 22 may optionally include memory remotely located relative to the processor 21, which can be connected to the processor 21 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0107] The program instructions / modules are stored in the memory 22 and, when executed by one or more processors 21, perform the small-scale maritime target detection method described in Embodiment 1 above.

[0108] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting small-scale targets at sea, characterized in that, include: The collected raw images are scaled up to a first image with a first preset resolution and a second image with a second preset resolution; After training the original yolov3-tiny network model, we obtain the first-level yolov3-tiny network model and the second-level yolov3-tiny network model. The first image is input into the first-level yolov3-tiny network model to obtain the first-level feature map; Construct an fdnet module, which is used to predict the region of interest image to be cropped in the second image; Wherein, after the first image is input into the first-level yolov3-tiny network model, the output of the fifth convolutional pooling layer of the first-level yolov3-tiny network model is obtained as the third image, and the third image is input into the fdnet module; After the third image is input into the fdnet module, the region of interest image on the second image that needs to be input into the second-level yolov3-tiny network model is determined based on the output of the fdnet module. The region of interest image is input into the second-level yolov3-tiny network model to obtain the second-level feature map. The first-level feature map and the second-level feature map together form the final feature map.

2. The method for detecting small-scale targets at sea according to claim 1, characterized in that, The process of training the original YOLOv3-tiny network model to obtain a first-level YOLOv3-tiny network model and a second-level YOLOv3-tiny network model includes: The original dataset was evaluated based on the publicly available SeaShips ship target detection and recognition dataset to obtain the ship data training set; Using the aforementioned ship data training set, the original yolov3-tiny network model is trained to obtain the first-level yolov3-tiny network model and the second-level yolov3-tiny network model after training.

3. The method for detecting small-scale targets at sea according to claim 1, characterized in that, The process involves inputting the region of interest image into the second-level YOLOv3-tiny network model to obtain a second-level feature map. The first-level feature map and the second-level feature map together form the final feature map, including: The second-level feature map includes a third feature map at a first resolution and a fourth feature map at a second resolution, wherein the third feature map and the fourth feature map are another part of the final feature map; Based on the information contained in the first-level feature map and the second-level feature map, the specific information of the target object in the original image is detected.

4. A small-scale target detection device for marine applications, characterized in that, The detection device is applicable to the detection method described in any one of claims 1 to 3, comprising: an original image preprocessing module, a YOLOv3-Tiny network model training module, an FDNet module construction module, and a feature map acquisition module, wherein: The original image preprocessing module is used to scale the original image into a first image with a first preset resolution and a second image with a second preset resolution; The yolov3-tiny network model training module is used to train the original yolov3-tiny network model to obtain the trained first-level yolov3-tiny network model and the second-level yolov3-tiny network model. The fdnet module construction module is used to construct the fdnet module to predict the region of interest image that needs to be cropped on the second image; The feature map acquisition module is used to acquire feature maps for detection of the original image.

5. The small-scale target detection device for marine applications according to claim 4, characterized in that, The yolov3-tiny network model training module includes a ship data training set construction unit and a yolov3-tiny network model training unit; The ship data training set construction unit is used to evaluate the original dataset based on the SeaShips ship target detection and recognition public dataset to obtain the ship data training set; The yolov3-tiny network model training unit is used to train the original yolov3-tiny network model based on the ship data training set.

6. The small-scale target detection device for marine applications according to claim 4, characterized in that, The fdnet module construction module is used to construct the fdnet module to predict the region of interest image to be cropped on the second image, including: The output of the fifth convolutional pooling layer of the first-level yolov3-tiny network model is used as the input of the fdnet module. Based on the output of the fdnet module, the region of interest image on the second image that needs to be input into the second-level yolov3-tiny network model is determined.

7. The small-scale target detection device for marine applications according to claim 4, characterized in that, The feature map acquisition module is used to acquire feature maps by inputting the first image into the first-level yolov3-tiny network model to obtain the first-level feature map, and by inputting the region of interest image into the second-level yolov3-tiny network model to obtain the second-level feature map. The specific information of the target object in the original image is detected by analyzing the first-level feature map and the second-level feature map.

Images