Sea target determination method, device and equipment

Abstract

The invention provides a method, a device and equipment for determining a sea target. The method comprises the following steps: acquiring target image information of the sea target acquired by information acquisition equipment; performing feature extraction and attention calculation processing on the target image information to obtain a target detection result; obtaining a target local image according to the target detection result; performing image enhancement processing on the target local image to obtain a water surface target image; the water surface target image comprises a detected marine target. According to the scheme, target detection can be carried out in a low-visibility and low-illumination environment, and the efficiency of maritime search and rescue work is improved.

Description

Technical Field

[0001] The present invention relates to the field of computer information technology processing technology, and in particular to a method, device and equipment for determining a target at sea. Background Art

[0002] Maritime search and rescue work is of vital importance, and the current success rate of maritime search and rescue work is already very high. However, due to the tight time constraints of maritime rescue work, it is often necessary to operate in harsh marine environments. It is of great significance to study methods to promptly and accurately detect the target to be rescued under conditions of low visibility and low illumination.

[0003] Thanks to the continuous development of intelligent recognition algorithms, excellent network structures for target detection in visible light images and infrared images continue to emerge. However, the existing network structures are not competent for target detection tasks in low visibility and low illumination environments, and cannot achieve the goal of improving the efficiency of maritime search and rescue work. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a method, device and equipment for determining targets at sea, which can perform target detection in low visibility and low illumination environments, thereby improving the efficiency of search and rescue work at sea.

[0005] In order to solve the above technical problems, the technical solution of the present invention is as follows:

[0006] A method for determining a marine target, comprising:

[0007] Acquire target image information of marine targets collected by information collection equipment;

[0008] Performing feature extraction and attention calculation processing on the target image information to obtain a target detection result;

[0009] Obtaining a target local image according to the target detection result;

[0010] The target local image is subjected to image enhancement processing to obtain a surface target image; the surface target image includes the detected sea target.

[0011] Optionally, feature extraction and attention calculation processing are performed on the target image information to obtain a target detection result, including:

[0012] Preprocessing the target image information to obtain target data;

[0013] Inputting the target data into a convolutional neural network module of a target detection network structure for feature extraction to obtain a first feature matrix;

[0014] Inputting the first feature matrix into the attention mechanism module of the target detection network structure for attention feature processing to obtain a second feature matrix after attention is allocated;

[0015] Inputting the second feature matrix into the feature pyramid network module of the target detection network structure for feature fusion and upsampling processing to obtain a third feature matrix;

[0016] The third feature matrix is ​​input into the detection head module of the target detection network structure for feature fusion and conversion to obtain a target detection result.

[0017] Optionally, the first feature matrix is ​​input into the attention mechanism module of the target detection network structure for attention feature processing to obtain a second feature matrix after attention is allocated, including:

[0018] The first feature matrix is ​​equally divided to obtain G first sub-feature matrices, where G is a positive integer;

[0019] Perform channel attention feature extraction and spatial attention feature extraction on each of the first sub-feature matrices to obtain G first sub-feature matrices after attention processing;

[0020] The G first sub-feature matrices after attention processing are subjected to matrix superposition and channel shuffling processing to obtain a second feature matrix.

[0021] Optionally, channel attention feature extraction and spatial attention feature extraction are performed on each of the first sub-feature matrices to obtain G attention-processed first sub-feature matrices, including:

[0022] Divide each of the first sub-feature matrices equally, so that each of the first sub-feature matrices is divided into two partial first sub-feature matrices;

[0023] Perform channel attention feature extraction on the first part of the first sub-feature matrix, and multiply the result of the channel attention feature extraction by the first part of the first sub-feature matrix to obtain the first part of the sub-feature matrix after channel attention processing;

[0024] Perform spatial attention feature extraction on the first sub-feature matrix of the second part, and multiply the result of the spatial attention feature extraction by the first sub-feature matrix of the second part to obtain the first sub-feature matrix of the part after spatial attention processing;

[0025] The part of the first sub-feature matrix after the channel attention processing and the part of the first sub-feature matrix after the spatial attention processing are superimposed to obtain the first sub-feature matrix after the attention processing; the G first sub-feature matrices are equally divided for attention feature extraction to obtain a total of G first sub-feature matrices after the attention processing.

[0026] Optionally, the third feature matrix is ​​input into the detection head module of the target detection network structure for feature fusion and conversion to obtain a target detection result, including:

[0027] Convolving the third feature matrix with a preset convolution kernel to obtain a third sub-feature matrix;

[0028] Performing feature fusion and conversion processing on the third sub-feature matrix to obtain a target type detection result;

[0029] The third sub-feature matrix is ​​subjected to feature fusion and conversion processing to obtain a target area detection result and an intersection-over-union detection result;

[0030] The target detection result is obtained according to the target type detection result, the target area detection result and the intersection-over-union detection result.

[0031] Optionally, the third sub-feature matrix is ​​subjected to feature fusion and conversion processing to obtain a target type detection result, including:

[0032] Performing a convolution operation on the third sub-feature matrix and a preset convolution kernel to obtain a third sub-feature matrix after the first convolution operation;

[0033] According to the third sub-feature matrix after the first convolution process, the target category detection result is obtained.

[0034] Optionally, the third sub-feature matrix is ​​subjected to feature fusion and conversion processing to obtain a target area detection result and an intersection-over-union detection result, including:

[0035] Performing a convolution operation on the third sub-feature matrix and a preset convolution kernel to obtain a third sub-feature matrix after the second convolution operation;

[0036] According to the third sub-feature matrix after the second convolution processing, the target area detection result and the intersection-over-union detection result are obtained.

[0037] The present invention also provides a marine target determination device, comprising:

[0038] An acquisition module, used to acquire target image information of marine targets acquired by information acquisition equipment;

[0039] A processing module is used to perform feature extraction and attention calculation processing on the target image information to obtain a target detection result; obtain a target local image based on the target detection result; perform image enhancement processing on the target local image to obtain a surface target image; the surface target image includes the detected sea target.

[0040] The present invention also provides a computing device, comprising: a processor and a memory storing a computer program, wherein when the computer program is run by the processor, the method as described above is executed.

[0041] The present invention also provides a computer-readable storage medium storing instructions, which, when executed on a computer, enable the computer to execute the above method.

[0042] The above solution of the present invention includes at least the following beneficial effects:

[0043] The above scheme of the present invention obtains target image information of the sea target collected by the information collection device; performs feature extraction and attention calculation processing on the target image information to obtain a target detection result; obtains a target partial image according to the target detection result; performs image enhancement processing on the target partial image to obtain a surface target image; the surface target image contains the detected sea target. Target detection can be performed in low visibility and low illumination environments, thereby improving the efficiency of sea search and rescue work. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is a first flow chart of a method for determining a marine target according to an embodiment of the present invention;

[0045] Figure 2 is a schematic diagram of the overall process of the method for determining a marine target according to an embodiment of the present invention;

[0046] Figure 3 is an architecture diagram of a target detection network structure according to an embodiment of the present invention;

[0047] Figure 4 is an attention mechanism module diagram of a target detection network structure according to an embodiment of the present invention;

[0048] Figure 5 is a detection head module diagram of a target detection network structure according to an embodiment of the present invention;

[0049] Figure 6 4 is a structural diagram of a marine target determination device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present invention and to enable the scope of the present invention to be fully communicated to those skilled in the art.

[0051] like Figure 1As shown, an embodiment of the present invention provides a method for determining a target at sea, comprising:

[0052] Step 11, obtaining target image information of the marine target collected by the information collection device;

[0053] Here, the information acquisition device refers to a device capable of collecting target images on the sea surface, preferably a high-performance camera that can collect high-resolution video images on the water surface. The target image information is an image of a target on the water surface collected by the information acquisition device, and the target can be real target information to be detected or false target information.

[0054] Step 12, performing feature extraction and attention calculation processing on the target image information to obtain a target detection result;

[0055] Here, by using a pre-trained target detection network structure, the target image information collected by the information collection device is pre-processed and then input into the target detection network structure to obtain a target detection result, which includes detecting the target or not detecting the target.

[0056] Step 13, obtaining a target local image according to the target detection result;

[0057] Here, when the target detection network structure detects that the input target image information contains a target, it indicates that the input target image information contains a target that needs to be rescued, and then a valuable target local image is extracted from the image in which the target is detected.

[0058] Step 14, performing image enhancement processing on the target local image to obtain a surface target image; the surface target image includes the detected sea target.

[0059] Here, according to the weather environment of the currently collected target image information, a suitable image enhancement module is selected to enhance the extracted target partial image to obtain a surface target image. After the surface target image is processed by the image enhancement module, the target contained in the image can be clearly displayed. At this time, the rescuer can determine whether to implement rescue according to the display of the target in the surface target image.

[0060] The following is attached to the instruction manual Figure 2 Further explanation of the application process:

[0061] First, the information acquisition device collects image data, preferably video information of the surface target to be rescued collected by a high-performance camera. The collected target image information is preprocessed and input into a preset target detection network structure. The target detection network structure is a high-performance target detection model obtained by training after adding an attention mechanism and decoupling a single detection head on the basis of the conventional fifth version of the target detection algorithm of the YOLO series. The target detection network structure performs a series of convolution operations on the input image to extract features, calculates the attention allocation weight, and outputs the detection result after feature fusion. If the target detection network structure detects a suspected search and rescue target, the image of the detected target is locally extracted, and according to the rescue environment, an appropriate image enhancement module is selected to enhance the local image of the suspected search and rescue mission target; a dark channel defogging algorithm module is used in foggy days, and a Retinex algorithm module is used in low-light environments. The present invention can also flexibly add new image enhancement modules based on the specific application environment to adapt to different maritime search and rescue environments; the processing results are displayed in a human-computer interaction display device. The search and rescue personnel determine whether the suspected search and rescue mission target displayed on the display device is the mission target. If not, they continue to perform the search mission. If it is the mission target, they perform the rescue mission and this process ends.

[0062] During the training process of the target detection network structure, the video data of the surface target to be rescued in the simulated scene or real scene collected by the high-performance camera is used as the basic data for training. At the same time, the data set collected by the high-performance camera is also expanded to adapt to the target detection tasks in more different environments. In the process of collecting the data collected by the camera, it is necessary to ensure that the collected image data covers a variety of surface targets to be rescued, a variety of meteorological scenes and a variety of application scenarios of maritime rescue environments, and to ensure the performance of the present invention from the perspective of the data set construction method; the material is annotated using an annotation tool to generate a data set that meets the input format of the convolutional neural network. In the process of data expansion, in addition to the classical method, the real scene of maritime rescue in a low-light environment is simulated by reducing the image brightness, and the real scene of maritime rescue in a rainy and foggy environment is simulated by adding Gaussian noise, etc., and the surface target data set is targeted for data enhancement.

[0063] like Figure 3 As shown, in an optional embodiment of the present invention, step 12 may include:

[0064] Step 121, preprocessing the target image information to obtain target data;

[0065] Step 122, inputting the target data into a convolutional neural network module of a target detection network structure for feature extraction to obtain a first feature matrix;

[0066] Here, the architecture of the target detection network structure is as follows Figure 3 As shown, it includes: a convolutional neural network module, an attention mechanism module, a feature pyramid network module and a detection head module. The convolutional neural network module includes multiple ConvBNSiLU (convolution), C3 and an SPPF module. The ConvBNSiLU consists of three modules: convolution, batch normalization and activation. In the convolutional neural network module, each time the feature matrix passes through a ConvBNSiLU, the size of the feature map is reduced by half and the number of channels is doubled. The C3 and SPPF modules adopt the classic C3 and SPPF network structure in the YOLO series, wherein the C3 module contains three convolutional layers for extracting and fusing features from different dimensions. The SPPF consists of a convolutional layer and a series of pooling layers in series, which are used to fuse features of different scales and integrate multi-scale features into a feature vector of fixed length.

[0067] In such Figure 3 In the convolutional neural network module shown, layers 0 to 4 are used to predict small targets, layers 5 to 6 are used to predict medium targets, and layers 7 to 9 are used to predict large targets. In the process of processing the feature matrix from layer 0 to layer 9, the size of the feature map gradually decreases and the number of channels gradually increases. Among them, the first feature matrix output by the 4th layer C3, the 6th layer C3 and the 9th layer SPPF is respectively input into the following attention mechanism module for attention feature processing.

[0068] Step 123, inputting the first feature matrix into the attention mechanism module of the target detection network structure to perform attention feature processing, and obtaining a second feature matrix after attention is allocated;

[0069] Specifically, step 123 may include:

[0070] Step 1231, equally divide the first feature matrix to obtain G first sub-feature matrices, where G is a positive integer;

[0071] Here, if Figure 4 As shown, the first feature matrix is ​​a feature matrix of size W*H*N, where W is the feature map width, H is the feature map height, and N is the number of channels. The first feature matrix of N channels is divided into G parts, each of which is a first sub-feature matrix of size W*H*K. In the process of matrix division, K parts can be randomly selected or extracted in sequence, as long as the N channels are divided into G parts, each of which contains K non-repeating channels. Preferably, divide them in sequence. For example, 9 channels can be divided into three parts: [1,2,3], [4,5,6], and [7,8,9].

[0072] Step 1232, performing channel attention feature extraction and spatial attention feature extraction on each of the first sub-feature matrices to obtain G first sub-feature matrices after attention processing; including:

[0073] Step 12321, equally divide each of the first sub-feature matrices into two partial first sub-feature matrices;

[0074] Here, if Figure 4 As shown, the first sub-feature matrix of size W*H*K is divided into two equal parts, namely, two feature matrices of size W*H*K / 2.

[0075] Step 12322, performing channel attention feature extraction on the first part of the first sub-feature matrix, multiplying the result of the channel attention feature extraction by the first part of the first sub-feature matrix, to obtain the first part of the sub-feature matrix after channel attention processing;

[0076] Here, channel attention feature extraction is used to learn channel dependency. The specific operation method of channel attention feature extraction is to first average pool the feature matrix and obtain the feature vector, then perform the function operation F(x)=Wx+b and use the σ(·) threshold to process the result. The matrix after channel attention feature extraction is multiplied by the W*H*K / 2 feature matrix to obtain the first sub-feature matrix of part of the size of W*H*K / 2.

[0077] Step 12323, extracting spatial attention features from the first sub-feature matrix of the second part, and multiplying the result of the spatial attention feature extraction by the first sub-feature matrix of the second part to obtain the first sub-feature matrix of the part after spatial attention processing;

[0078] Here, spatial attention feature extraction is used to learn pixel-level pairwise relationships. The specific operation method of spatial attention feature extraction is to first perform Group Normalization on the feature matrix, then perform the function operation F(x)=Wx+b and use the σ(·) threshold to process the result. The matrix after spatial attention feature extraction is multiplied by the W*H*K / 2 feature matrix to obtain the first sub-feature matrix of the size of W*H*K / 2.

[0079] Step 12324, superimpose the part of the first sub-feature matrix after the channel attention processing and the part of the first sub-feature matrix after the spatial attention processing to obtain the first sub-feature matrix after the attention processing; G first sub-feature matrices are equally divided for attention feature extraction to obtain a total of G first sub-feature matrices after the attention processing.

[0080] Here, after learning the sub-features in parallel in the spatial and channel dimensions respectively, the first sub-feature matrices of the two parts are superimposed to obtain a feature matrix of size W*H*K.

[0081] Step 1233, performing matrix superposition and channel shuffling processing on the G first sub-feature matrices after attention processing to obtain a second feature matrix.

[0082] Here, G W*H*K-sized first sub-feature matrices after attention processing are superimposed to obtain a W*H*N feature matrix. The feature layers are randomly arranged through channel shuffle, so that the sub-features are aggregated together to obtain the second feature matrix, and the performance of the neural network is improved.

[0083] Step 124, inputting the second feature matrix into the feature pyramid network module of the target detection network structure for feature fusion and upsampling processing to obtain a third feature matrix;

[0084] Here, in Figure 3 In the target detection network structure shown, the feature pyramid network module still uses the network structure of V5 in the YOLO series, including ConvBNSiLU (convolution), C3, Upsample (upsampling) and Concat (splicing). Among them, Upsample is used to change the size of the feature map to facilitate splicing with the previous feature map. Through the feature pyramid network module, the fusion of shallow graphic features and deep semantic features can be achieved, making target detection more accurate. The three second feature matrices output by the three attention mechanism modules are respectively input into the convolution layer, splicing layer and splicing layer of the feature pyramid network module, and the feature pyramid network module outputs the third feature matrix after processing.

[0085] Step 125, input the third feature matrix into the detection head module of the target detection network structure for feature fusion and conversion to obtain a target detection result.

[0086] Specifically, Figure 5 As shown, step 125 may include:

[0087] Step 1251, convolve the third feature matrix with a preset convolution kernel to obtain a third sub-feature matrix; here, the preset convolution kernel is preferably a convolution kernel of size 1*1; the third sub-feature matrix is ​​a third sub-feature matrix of size W*H*N.

[0088] Step 1252, performing feature fusion and conversion processing on the third sub-feature matrix to obtain a target type detection result; including:

[0089] Step 12521, convolve the third sub-feature matrix with a preset convolution kernel to obtain a third sub-feature matrix after the first convolution processing; here, the preset convolution kernel is preferably a 3*3 convolution kernel, and the third sub-feature matrix of size W*H*N is convolved twice with the 3*3 convolution kernel to obtain a feature matrix of size W*H*N.

[0090] Step 12522, obtaining the target category detection result according to the third sub-feature matrix after the first convolution processing.

[0091] Step 1253, performing feature fusion and conversion processing on the third sub-feature matrix to obtain a target area detection result and an intersection-over-union detection result; including:

[0092] Step 12531, convolve the third sub-feature matrix with a preset convolution kernel to obtain a third sub-feature matrix after the second convolution process; the preset convolution kernel is preferably a 3*3 convolution kernel, and the third sub-feature matrix of size W*H*N is convolved twice with the 3*3 convolution kernel to obtain a feature matrix of size W*H*N.

[0093] Step 12532, according to the third sub-feature matrix after the second convolution processing, obtain the target area detection result and the intersection-over-union detection result. Figure 5 As shown, the region is predicted based on some channels of N channels, and the IOU (intersection over union) is predicted based on some channels of N channels. Preferably, N is 5.

[0094] Step 1254, obtaining a target detection result according to the target type detection result, the target area detection result and the intersection-over-union detection result.

[0095] The target detection network structure of the present invention sets an attention mechanism module between the convolutional neural network module and the feature pyramid network module. By introducing the attention mechanism, the neural network can focus on the relevant elements of the input more accurately, thereby effectively improving the performance of the neural network. The detection head module is set to two prediction branches, one prediction branch only performs target type judgment, and the other prediction branch performs target area and intersection-over-union detection, thereby improving the convergence speed of the neural network and enhancing the performance of the neural network.

[0096] In an optional embodiment of the present invention, step 14 may include:

[0097] Step 141, determining a target image enhancement module corresponding to the current weather environment from among a plurality of image enhancement modules according to the weather environment when the information acquisition device acquires the target image information;

[0098] Step 142: performing image enhancement processing on the target local image according to the target image enhancement module to obtain a water surface target image.

[0099] In this embodiment, the image enhancement module can be a dark channel defogging algorithm module, a Retinex algorithm module, or other algorithm modules. The image enhancement module in this application can be expanded according to different external environment requirements, and this is not limited in this application. The dark channel defogging algorithm module processes each color channel of the input image to generate a dark channel map; uses the dark channel map to estimate the atmospheric light intensity in the scene; calculates the transmittance map using the atmospheric scattering model; combines the transmittance map and the atmospheric light intensity, and uses the inverse atmospheric scattering model to restore the original fog-free image, thereby removing the influence of fog in the image. The Retinex algorithm module applies a series of Gaussian filters of different sizes to the input image to obtain illumination estimates at different scales; for each scale, the original image is subtracted from the illumination estimate of the corresponding scale to obtain a reflectance image; the reflectance images at all scales are combined to form the final enhanced image. The Retinex algorithm module is suitable for images collected in low-light environments.

[0100] The hardware equipment relied on by the above-mentioned offshore target determination method of the present invention includes:

[0101] 1. Information collection equipment, mainly used for real-time collection of high-resolution surface video streams, providing high-quality real-time data for surface target detection;

[0102] 2. Data processing unit, mainly used to receive and process surface video streams, use the target detection network structure to detect suspected search and rescue mission targets, and call appropriate image enhancement modules according to the working environment to locally enhance the detection results;

[0103] 3. Human-computer interactive display equipment is mainly used to display high-quality image information of suspected search and rescue mission targets to search and rescue personnel, providing information support for search and rescue personnel to take rescue actions.

[0104] The above method of the present invention uses intelligent algorithms to assist people in conducting maritime search and rescue in the loop, which can improve the efficiency of maritime search and rescue work in harsh environments such as low visibility and low illumination. This method has the function of improving the work efficiency of search and rescue personnel and reducing the workload of search and rescue personnel. By using high-performance cameras, search and rescue personnel can observe targets in a larger range; using the attention mechanism module, the detection head module and the modular local image enhancement algorithm, high-quality images of suspected search and rescue mission targets are provided to search and rescue personnel, thereby improving the accuracy and efficiency of search and rescue work.

[0105] like Figure 6 As shown, an embodiment of the present invention further provides a marine target determination device 60, comprising:

[0106] An acquisition module 61 is used to acquire target image information of a marine target acquired by an information acquisition device;

[0107] The processing module 62 is used to perform feature extraction and attention calculation processing on the target image information to obtain a target detection result; obtain a target local image based on the target detection result; perform image enhancement processing on the target local image to obtain a surface target image; the surface target image includes the detected sea target.

[0108] Optionally, feature extraction and attention calculation processing are performed on the target image information to obtain a target detection result, including:

[0109] Preprocessing the target image information to obtain target data;

[0110] Inputting the target data into a convolutional neural network module of a target detection network structure for feature extraction to obtain a first feature matrix;

[0111] Inputting the first feature matrix into the attention mechanism module of the target detection network structure for attention feature processing to obtain a second feature matrix after attention is allocated;

[0112] Inputting the second feature matrix into the feature pyramid network module of the target detection network structure for feature fusion and upsampling processing to obtain a third feature matrix;

[0113] The third feature matrix is ​​input into the detection head module of the target detection network structure for feature fusion and conversion to obtain a target detection result.

[0114] Optionally, the first feature matrix is ​​input into the attention mechanism module of the target detection network structure for attention feature processing to obtain a second feature matrix after attention is allocated, including:

[0115] The first feature matrix is ​​equally divided to obtain G first sub-feature matrices, where G is a positive integer;

[0116] Perform channel attention feature extraction and spatial attention feature extraction on each of the first sub-feature matrices to obtain G first sub-feature matrices after attention processing;

[0117] The G first sub-feature matrices after attention processing are subjected to matrix superposition and channel shuffling processing to obtain a second feature matrix.

[0118] Optionally, channel attention feature extraction and spatial attention feature extraction are performed on each of the first sub-feature matrices to obtain G attention-processed first sub-feature matrices, including:

[0119] Divide each of the first sub-feature matrices equally, so that each of the first sub-feature matrices is divided into two partial first sub-feature matrices;

[0120] Perform channel attention feature extraction on the first part of the first sub-feature matrix, and multiply the result of the channel attention feature extraction by the first part of the first sub-feature matrix to obtain the first part of the sub-feature matrix after channel attention processing;

[0121] Perform spatial attention feature extraction on the first sub-feature matrix of the second part, and multiply the result of the spatial attention feature extraction by the first sub-feature matrix of the second part to obtain the first sub-feature matrix of the part after spatial attention processing;

[0122] The part of the first sub-feature matrix after the channel attention processing and the part of the first sub-feature matrix after the spatial attention processing are superimposed to obtain the first sub-feature matrix after the attention processing; the G first sub-feature matrices are equally divided for attention feature extraction to obtain a total of G first sub-feature matrices after the attention processing.

[0123] Optionally, the third feature matrix is ​​input into the detection head module of the target detection network structure for feature fusion and conversion to obtain a target detection result, including:

[0124] Convolving the third feature matrix with a preset convolution kernel to obtain a third sub-feature matrix;

[0125] Performing feature fusion and conversion processing on the third sub-feature matrix to obtain a target type detection result;

[0126] The third sub-feature matrix is ​​subjected to feature fusion and conversion processing to obtain a target area detection result and an intersection-over-union detection result;

[0127] The target detection result is obtained according to the target type detection result, the target area detection result and the intersection-over-union detection result.

[0128] Optionally, the third sub-feature matrix is ​​subjected to feature fusion and conversion processing to obtain a target type detection result, including:

[0129] Performing a convolution operation on the third sub-feature matrix and a preset convolution kernel to obtain a third sub-feature matrix after the first convolution operation;

[0130] According to the third sub-feature matrix after the first convolution process, the target category detection result is obtained.

[0131] Optionally, the third sub-feature matrix is ​​subjected to feature fusion and conversion processing to obtain a target area detection result and an intersection-over-union detection result, including:

[0132] Performing a convolution operation on the third sub-feature matrix and a preset convolution kernel to obtain a third sub-feature matrix after the second convolution operation;

[0133] According to the third sub-feature matrix after the second convolution processing, the target area detection result and the intersection-over-union detection result are obtained.

[0134] It should be noted that the device is a device corresponding to the above method, and all implementation methods in the above method embodiments are applicable to the embodiments of the device and can achieve the same technical effects.

[0135] The embodiment of the present invention further provides a computing device, comprising: a processor, a memory storing a computer program, wherein when the computer program is executed by the processor, the method described above is executed. All implementations in the above method embodiment are applicable to this embodiment and can achieve the same technical effect.

[0136] The embodiment of the present invention further provides a computer-readable storage medium storing instructions, which, when executed on a computer, enable the computer to execute the above method. All implementations in the above method embodiment are applicable to this embodiment and can achieve the same technical effect.

[0137] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0138] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0139] In the embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0140] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0141] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0142] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, ROM, RAM, magnetic disks, or optical disks.

[0143] In addition, it should be noted that in the apparatus and method of the present invention, it is obvious that each component or each step can be decomposed and / or recombined. These decompositions and / or recombinations should be regarded as equivalent schemes of the present invention. Moreover, the steps of performing the above-mentioned series of processing can naturally be performed in chronological order according to the order of description, but it is not necessary to perform them in chronological order, and some steps can be performed in parallel or independently of each other. For those of ordinary skill in the art, it is understood that all or any steps or components of the method and apparatus of the present invention can be implemented in any computing device (including processors, storage media, etc.) or a network of computing devices in hardware, firmware, software or a combination thereof, which can be achieved by those of ordinary skill in the art using their basic programming skills after reading the description of the present invention.

[0144] Therefore, the purpose of the present invention can also be achieved by running a program or a group of programs on any computing device. The computing device can be a well-known general device. Therefore, the purpose of the present invention can also be achieved by simply providing a program product containing a program code that implements the method or device. That is to say, such a program product also constitutes the present invention, and the storage medium storing such a program product also constitutes the present invention. Obviously, the storage medium can be any well-known storage medium or any storage medium developed in the future. It should also be pointed out that in the device and method of the present invention, it is obvious that each component or each step can be decomposed and / or recombined. These decompositions and / or recombinations should be regarded as equivalent schemes of the present invention. In addition, the steps of performing the above-mentioned series of processing can naturally be performed in chronological order according to the order of description, but it is not necessary to perform them in chronological order. Some steps can be performed in parallel or independently of each other.

[0145] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for determining a target at sea, characterized in that: include: Acquire target image information of marine targets collected by information collection equipment; Performing feature extraction and attention calculation processing on the target image information to obtain a target detection result; Obtaining a target local image according to the target detection result; The target local image is subjected to image enhancement processing to obtain a surface target image; the surface target image includes the detected sea target.

2. The method for determining a marine target according to claim 1, characterized in that: Perform feature extraction and attention calculation processing on the target image information to obtain a target detection result, including: Preprocessing the target image information to obtain target data; Inputting the target data into a convolutional neural network module of a target detection network structure for feature extraction to obtain a first feature matrix; Inputting the first feature matrix into the attention mechanism module of the target detection network structure for attention feature processing to obtain a second feature matrix after attention is allocated; Inputting the second feature matrix into the feature pyramid network module of the target detection network structure for feature fusion and upsampling processing to obtain a third feature matrix; The third feature matrix is ​​input into the detection head module of the target detection network structure for feature fusion and conversion to obtain a target detection result.

3. The method for determining a marine target according to claim 2, characterized in that: Inputting the first feature matrix into the attention mechanism module of the target detection network structure for attention feature processing to obtain a second feature matrix after allocating attention, including: The first feature matrix is ​​equally divided to obtain G first sub-feature matrices, where G is a positive integer; Perform channel attention feature extraction and spatial attention feature extraction on each of the first sub-feature matrices to obtain G first sub-feature matrices after attention processing; The G first sub-feature matrices after attention processing are subjected to matrix superposition and channel shuffling processing to obtain a second feature matrix.

4. The method for determining a marine target according to claim 3, characterized in that: Channel attention feature extraction and spatial attention feature extraction are performed on each of the first sub-feature matrices to obtain G first sub-feature matrices after attention processing, including: Divide each of the first sub-feature matrices equally, so that each of the first sub-feature matrices is divided into two partial first sub-feature matrices; Perform channel attention feature extraction on the first part of the first sub-feature matrix, and multiply the result of the channel attention feature extraction by the first part of the first sub-feature matrix to obtain the first part of the sub-feature matrix after channel attention processing; Perform spatial attention feature extraction on the first sub-feature matrix of the second part, and multiply the result of the spatial attention feature extraction by the first sub-feature matrix of the second part to obtain the first sub-feature matrix of the part after spatial attention processing; The part of the first sub-feature matrix after the channel attention processing and the part of the first sub-feature matrix after the spatial attention processing are superimposed to obtain the first sub-feature matrix after the attention processing; the G first sub-feature matrices are equally divided for attention feature extraction to obtain a total of G first sub-feature matrices after the attention processing.

5. The method for determining a marine target according to claim 2, characterized in that: Inputting the third feature matrix into the detection head module of the target detection network structure for feature fusion and conversion to obtain a target detection result, including: Convolving the third feature matrix with a preset convolution kernel to obtain a third sub-feature matrix; Performing feature fusion and conversion processing on the third sub-feature matrix to obtain a target type detection result; The third sub-feature matrix is ​​subjected to feature fusion and conversion processing to obtain a target area detection result and an intersection-over-union detection result; The target detection result is obtained according to the target type detection result, the target area detection result and the intersection-over-union detection result.

6. The method for determining a marine target according to claim 5, characterized in that: The third sub-feature matrix is ​​subjected to feature fusion and conversion processing to obtain a target type detection result, including: Performing a convolution operation on the third sub-feature matrix and a preset convolution kernel to obtain a third sub-feature matrix after the first convolution operation; According to the third sub-feature matrix after the first convolution process, the target category detection result is obtained.

7. The method for determining a marine target according to claim 5, characterized in that: The third sub-feature matrix is ​​subjected to feature fusion and conversion processing to obtain a target area detection result and an intersection-over-union detection result, including: Performing a convolution operation on the third sub-feature matrix and a preset convolution kernel to obtain a third sub-feature matrix after the second convolution operation; According to the third sub-feature matrix after the second convolution processing, the target area detection result and the intersection-over-union detection result are obtained.

8. A marine target determination device, characterized in that: include: An acquisition module, used to acquire target image information of marine targets acquired by information acquisition equipment; A processing module is used to perform feature extraction and attention calculation processing on the target image information to obtain a target detection result; and obtain a target partial image according to the target detection result; The target local image is subjected to image enhancement processing to obtain a surface target image; the surface target image includes the detected sea target.

9. A computing device, characterized in that include: A processor and a memory storing a computer program, wherein when the computer program is executed by the processor, the method according to any one of claims 1 to 7 is performed.

10. A computer-readable storage medium, characterized in that: Instructions are stored, and when the instructions are executed on a computer, the computer is caused to execute the method according to any one of claims 1 to 7.

Images