Dynamic channel bayonet ship identification method and system for complex environment

Abstract

The invention provides a channel bayonet ship dynamic identification method and system for a complex environment, and relates to the technical field of image processing. The method comprises the steps of obtaining a dynamic monitoring image set of a target ship, constructing a plurality of local feature images of each to-be-analyzed image, performing feature fusion to generate a structural feature image, and extracting a plurality of target component areas of the to-be-analyzed images; performing regional evolution analysis on the plurality of to-be-analyzed images to generate a plurality of component regional evolution trajectories; dividing the dynamic monitoring image set into a plurality of local monitoring image sets, performing local intensity analysis on each local monitoring image set, and constructing to obtain a local feature set; and performing region coverage detection on the plurality of to-be-analyzed images in the local feature set, generating a global feature set corresponding to the local feature set, processing the global feature set through a pre-trained ship identification model, and generating a target identification result of the target ship. According to the invention, the accuracy and stability of ship dynamic identification in a complex environment are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and in particular to a channel portal ship dynamic identification method and system for complex environments. BACKGROUND

[0002] In key traffic nodes such as port channels and waterways, image recognition technology for passing ships is widely used in security management, traffic scheduling, intelligent supervision and other fields. In particular, in the channel portal scene, by setting up a front-end camera device, multiple images of the ship passing through the portal are continuously obtained, and then target detection, identity recognition, type discrimination and other processing are performed, which is an important part of the existing intelligent ship supervision system.

[0003] In actual application, a common processing method is to directly input the collected multiple images into a pre-trained recognition model for batch analysis. However, due to differences in image clarity, angle, and occlusion, some images may have local blur, incomplete ship structure, occlusion, glare interference, and other problems, which can cause unstable model recognition results, and even misidentification or missed identification. In order to improve the accuracy of identification, some existing technologies attempt to filter out blurred and insufficient brightness frames before the images enter the model, or perform denoising and enhancement processing on some frames, thereby improving the overall quality of the input image set. However, this process usually analyzes the image as a whole, which can easily ignore the structural integrity of the target ship in the image, i.e., whether the image truly and comprehensively presents the key components of the ship, and can easily ignore the structural complementarity between multiple images, making it difficult to utilize the temporal evolution relationship in the image sequence for inter-frame information fusion, which can easily cause key frame omission or excessive frame redundancy, and can affect the accuracy of ship identification in complex environments. SUMMARY

[0004] To solve the above technical problems, the present application provides a channel portal ship dynamic identification method and system for complex environments, which considers the structural saliency of local regions in the image, combines the inter-frame structural evolution trend and image expression quality information, and provides a more targeted and information complete image input basis for the recognition model, thereby improving the accuracy and stability of ship dynamic identification in complex environments.

[0005] The first aspect of the present application provides a channel portal ship dynamic identification method for complex environments, comprising: obtaining a dynamic monitoring image set of a target ship, the dynamic monitoring image set comprising a plurality of to-be-analyzed images, performing feature extraction on each to-be-analyzed image, and constructing a plurality of local feature images for each to-be-analyzed image; perform feature fusion on the plurality of local feature images of the to-be-analyzed image to generate a structural feature image, determine a plurality of candidate center nodes of each structural feature image, and extract a plurality of target component regions of the to-be-analyzed image based on the plurality of candidate center nodes; perform region evolution analysis on the plurality of to-be-analyzed images, determine a plurality of local evolution combinations between any two adjacent to-be-analyzed images, and fuse the plurality of local evolution combinations to generate a plurality of component region evolution trajectories; divide the dynamic monitoring image set into a plurality of local monitoring image sets, perform local intensity analysis on each local monitoring image set according to the plurality of component region evolution trajectories, output a local intensity image of each local monitoring image set, and construct a local feature set; perform region coverage detection on the plurality of to-be-analyzed images in the local feature set, generate a global feature set corresponding to the local feature set, process the global feature set through a pre-trained ship recognition model, and generate a target recognition result of the target ship.

[0006] Preferably, the plurality of candidate center nodes of each structural feature image are determined, and the plurality of target component regions of the to-be-analyzed image are extracted based on the plurality of candidate center nodes, including: For the structural feature image, the plurality of local feature images of the to-be-analyzed image include edge feature images, texture feature images, and direction gradient feature images, the edge feature, the texture feature, and the direction gradient feature of each pixel point in the to-be-analyzed image are determined, and weighted fusion is performed to obtain a structural feature value of each pixel point, thereby generating a structural feature image corresponding to the to-be-analyzed image; perform sliding window analysis on the structural feature image, determine a plurality of local feature extreme points of the structural feature image in different sliding window regions based on the structural feature value, and mark the plurality of local feature extreme points as a plurality of candidate center nodes of the structural feature image; perform region growing analysis on the plurality of candidate center nodes respectively, construct a candidate growing region of each candidate center node, determine a plurality of region repeat nodes in the structural feature image, calculate a region allocation index corresponding to each region repeat node in the associated plurality of candidate growing regions, perform region reconstruction on the plurality of candidate growing regions based on the region allocation index of the region repeat node, and generate a plurality of target component regions of the to-be-analyzed image.

[0007] Preferably, the plurality of local evolution combinations between any two adjacent to-be-analyzed images are determined, and the plurality of local evolution combinations are fused to generate a plurality of component region evolution trajectories, including: generate a plurality of candidate evolution combinations according to the plurality of target component regions corresponding to the two adjacent to-be-analyzed images, calculate a region overlap parameter and a structure matching parameter of each candidate evolution combination, and calculate a local matching index of the candidate evolution combination based on the region overlap parameter and the structure matching parameter; Determine a plurality of local evolution combinations from a plurality of candidate evolution combinations according to a local matching index, perform evolution splicing on the plurality of local evolution combinations, and obtain a plurality of component region evolution trajectories about the plurality of to-be-analyzed images.

[0008] Preferably, local intensity analysis is performed on each local monitoring image set according to the plurality of component region evolution trajectories, a local intensity image of each local monitoring image set is output, and a local feature set is constructed. According to the structural feature image, a region response intensity parameter corresponding to each target component region in the local monitoring image set is calculated, a plurality of trajectory aggregations are generated by performing trajectory aggregation on a plurality of target component regions in the local monitoring image set according to the plurality of component region evolution trajectories, a local trajectory response intensity parameter of each target component region in each trajectory aggregation is calculated, a local intensity parameter is generated by fusing a plurality of local trajectory response intensity parameters of the to-be-analyzed image, a local intensity image of each local monitoring image set is determined according to the local intensity parameter, and a local feature set is constructed according to a plurality of local intensity images.

[0009] Preferably, region coverage detection is performed on a plurality of to-be-analyzed images in the local feature set, a global feature set corresponding to the local feature set is generated, and the global feature set comprises: A global trajectory response intensity parameter of each target component region in the component region evolution trajectory is calculated, a trajectory distribution parameter corresponding to a plurality of target component regions of each to-be-analyzed image in the local feature set is determined, the trajectory distribution parameter of the to-be-analyzed image is weighted and fused according to the global trajectory response intensity parameter, a global intensity parameter corresponding to each to-be-analyzed image in the local feature set is generated, and image screening is performed on the local feature set based on the global intensity parameter to generate a global feature set.

[0010] Preferably, a ratio between a number of pixel points overlapped by two target component regions in the candidate evolution combination and a total number of pixel points contained is calculated as a region overlap parameter of the candidate evolution combination.

[0011] The second aspect of the present application provides a channel portal ship dynamic identification system for complex environments, which is used to realize the above-mentioned ship dynamic identification method for complex environments, and comprises: A local feature extraction module is configured to acquire a dynamic monitoring image set of a target ship, the dynamic monitoring image set comprises a plurality of to-be-analyzed images, perform feature extraction on each to-be-analyzed image, and construct a plurality of local feature images of each to-be-analyzed image. A component region generation module is configured to perform feature fusion on the plurality of local feature images of the to-be-analyzed image to generate a structural feature image, determine a plurality of candidate center nodes of each structural feature image, and extract a plurality of target component regions of the to-be-analyzed image based on the plurality of candidate center nodes. The regional evolution analysis module is used to perform regional evolution analysis on multiple images to be analyzed, determine multiple local evolution combinations between any two adjacent images to be analyzed, and fuse multiple local evolution combinations to generate multiple component regional evolution trajectories; A local intensity analysis module is used to divide the dynamic monitoring image set into multiple local monitoring image sets, perform local intensity analysis on each local monitoring image set according to the evolution trajectories of multiple component regions, output a local intensity image of each local monitoring image set, and construct a local feature set; The ship recognition optimization module is used to perform regional coverage detection on multiple images to be analyzed in the local feature set, generate a global feature set corresponding to the local feature set, process the global feature set through the pre-trained ship recognition model, and generate the target recognition result of the target ship.

[0012] The present invention has the following beneficial effects: The present invention constructs an image screening mechanism based on structural feature response. For the dynamically collected ship image sequence, multiple significant areas that can characterize the ship's structural components are extracted through multi-dimensional feature fusion and region growing algorithm, avoiding the problem of low representativeness caused by the prior definition of specific ship structure. Combined with the structural region matching between multiple frame images, the regional evolution trajectory across time series is constructed, and the stable distribution characteristics of the components are effectively captured. By accumulating the response intensity and comparing between frames in a local window, the representative images are optimized. Finally, through the coverage screening mechanism, the information integrity and spatial distribution breadth of the screened images in the structural dimension are guaranteed. A high-quality image input recognition model with clear structure and reasonable distribution can be selected from a large number of dynamic monitoring images, effectively reducing the interference of redundant images on the recognition model, and improving the accuracy and stability of ship dynamic recognition in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 A flow chart of a method for dynamic identification of ships at waterway checkpoints in complex environments provided by an embodiment of the present invention.

[0014] Figure 2 A schematic structural diagram of a dynamic ship identification system for a waterway checkpoint in a complex environment provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0015] In order to enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0016] Figure 1A flow chart of a channel bottleneck ship dynamic recognition method for complex environment provided by an embodiment of the present application is shown, please refer to Figure 1 A channel bottleneck ship dynamic recognition method for complex environment, comprising the following steps: Step S1, acquiring a dynamic monitoring image set of a target ship, the dynamic monitoring image set comprising a plurality of to-be-analyzed images, performing feature extraction on each to-be-analyzed image, and constructing a plurality of local feature images of each to-be-analyzed image.

[0017] Specifically, a set of continuous monitoring images, i.e. a plurality of continuous to-be-analyzed images, collected during the passing of the target ship through the channel bottleneck are acquired, and constitute a dynamic monitoring image set. Further, a plurality of channel feature images of each to-be-analyzed image are extracted. In this embodiment, edge features, texture features and direction gradient features are taken as examples. For the extraction of edge features, the to-be-analyzed image is processed by Sobel to extract the horizontal gradient map and the vertical gradient map of the to-be-analyzed image, and the gradient intensity value is calculated as the edge intensity feature according to the horizontal gradient feature and the vertical gradient feature of the pixel point, so as to construct the edge feature image. For the extraction of direction gradient features, the gradient direction features of different pixel points are calculated based on the horizontal gradient map and the vertical gradient map to constitute the direction gradient feature image. For the extraction of texture features, the to-be-analyzed image is processed by the LBP operator to generate the LBP texture feature value of each pixel point to construct the texture feature image. The response characteristics of different regions in the image are described in more detail, the region separability of the image is improved, and the potential component region in the image is determined subsequently.

[0018] Step S2, performing feature fusion on the plurality of local feature images of the to-be-analyzed image to generate a structure feature image, determining a plurality of candidate center nodes of each structure feature image, and extracting a plurality of target component regions of the to-be-analyzed image based on the plurality of candidate center nodes.

[0019] Specifically, the plurality of local feature images are fused, for example, the edge feature image, the texture feature image and the direction gradient feature image are weighted by equal weight, and finally the structure feature image of the to-be-analyzed image is obtained, so as to enhance the response contrast of the potential component region in the image. Local peak detection is performed on the structure feature image to identify a plurality of candidate structure center nodes, and structure region growing is performed with each center node as the core to obtain a plurality of block-shaped regions in the image which have significant continuity and edge complexity in space.

[0020] As an optional implementation process, the extraction of the plurality of target component regions of the to-be-analyzed image specifically comprises: The structural feature image is analyzed by using a sliding window. A fixed size, for example, 5x5, sliding window is used to scan the structural feature image, and in each sliding window region, it is determined whether there is a significant local feature extreme point based on the structural feature value of the pixel point. If the structural feature value of a pixel point is a local maximum in its corresponding window range, it is marked as a candidate center node.

[0021] The region growing analysis is performed on each candidate center node to construct a candidate growing region of each candidate center node. In this process, for each candidate center node, the region growing analysis based on feature consistency is performed. Specifically, the center node is taken as a seed point or an initial growing region, and a plurality of adjacent pixel points associated with the center node are determined. If the difference between the structural feature value of the adjacent pixel point and the average structural feature value of the plurality of pixel points of the initial growing region is less than a preset growing threshold, the initial growing region is updated according to the corresponding adjacent pixel point. The process of analyzing the adjacent pixel points by iteration, that is, initially determining a plurality of adjacent pixel points and updating the region, continues to determine the adjacent pixel points and update the region based on the new growing region, and gradually expands to construct a candidate growing region with each candidate center node as the core. When there is no new adjacent pixel point on the region boundary that satisfies the growing condition, the growing process of the region is terminated.

[0022] A plurality of region repeat nodes in the structural feature image are determined, and a region allocation index of each region repeat node in the associated plurality of candidate growing regions is calculated. Based on the region allocation index of the region repeat node, the plurality of candidate growing regions are regionally reconstructed to generate a plurality of target component regions of the image to be analyzed.

[0023] In this embodiment, considering the independence of the growing process, there may be some pixel points belonging to multiple candidate growing regions in the plurality of candidate growing regions finally obtained. Such pixel points are marked as region repeat nodes. For each region repeat node, a plurality of candidate growing regions associated with the region repeat node are determined, and the region allocation index of the region repeat node in different candidate growing regions is calculated. For the calculation of the region allocation index, the following method can be used. If the node is allocated to a candidate growing region in the current growing process, the initial state of the candidate growing region in the current growing process is taken as a reference, the difference between the structural feature value of the region repeat node and the average structural feature value of the plurality of pixel points of the growing region in the reference state is calculated, and the candidate growing region with the smallest difference is selected as the allocation object of the region repeat node. In this way, the plurality of region repeat nodes are redistributed, and the plurality of target component regions without repeated pixel points are finally obtained by reconstructing the plurality of candidate growing regions.

[0024] It is worth noting that in this step, by analyzing the local feature response of the structural feature image, a plurality of candidate center nodes are extracted and the target component region is constructed based on structural consistency, which can effectively avoid the problem of insufficient generalization caused by relying on ship semantic component labels. Due to the significant differences in structural arrangement, size ratio, cargo state and other aspects of different types of ships, if semantic parts such as bow, deck, engine room, etc. are directly used as analysis units, it is often difficult to stably obtain a region division with universal expression ability. The target component region constructed by the image local structure response in the application can better reflect the significant structure region presented by the image itself, has clear boundaries, strong internal consistency, and is more likely to form an evolutionary correspondence in multiple images, thereby improving the robustness of subsequent region evolution tracking and the extraction efficiency of representative images.

[0025] Step S3, performing region evolution analysis on the plurality of to-be-analyzed images to determine a plurality of local evolution combinations between any two adjacent to-be-analyzed images, and fusing the plurality of local evolution combinations to generate a plurality of component region evolution trajectories.

[0026] Specifically, on the basis of constructing a plurality of target component regions of each to-be-analyzed image, region evolution matching analysis is performed on any two adjacent to-be-analyzed images in the dynamic monitoring image set, considering the overall structural response feature strength and position distribution information, determining the local evolution combination of the region evolution relationship existing in the adjacent to-be-analyzed images, and sequentially correlating the local evolution combinations between multiple frames to construct a plurality of component region evolution trajectories, each trajectory representing the expression process of a structural region in consecutive frame images.

[0027] As an optional implementation process, for the generation of a plurality of component region evolution trajectories, specifically comprising: According to the plurality of target component regions corresponding to the adjacent two to-be-analyzed images, a plurality of candidate evolution combinations are generated, the region overlap parameter and the structure matching parameter of each candidate evolution combination are calculated, and the local matching index of the candidate evolution combination is calculated based on the region overlap parameter and the structure matching parameter.

[0028] In the embodiment, for any pair of adjacent image frames to be analyzed, a plurality of target component regions pre-generated in the current frame and the next frame are extracted respectively, all possible cross-frame region combination pairs, i.e., combinations between each current frame region and all regions in the next frame, are constructed to form a plurality of candidate evolution combinations for describing the possible temporal continuity and structural inheritance between two component regions. For each pair of candidate region combinations, a region overlap parameter reflecting the geometric consistency of the two regions in the image space is calculated, which can be obtained by calculating the ratio between the number of overlapping pixel points and the total number of pixel points involved. A structure matching parameter reflecting the response consistency of the two regions in the structural feature dimension is calculated, which can be based on the structural feature values of the pixel points in the structural feature image, for example, the mean values of the structural feature values of a plurality of pixel points in the two regions are calculated respectively and then represented by the difference between the mean values, and the region overlap parameter is modified by the structure matching parameter, and the ratio between the region overlap parameter and the structure matching parameter is calculated to obtain the local matching index of the candidate evolution combination.

[0029] A plurality of local evolution combinations are determined from the plurality of candidate evolution combinations according to the local matching index, and the plurality of local evolution combinations are evolutionally spliced to obtain a plurality of component region evolution trajectories for the plurality of image frames to be analyzed.

[0030] In the embodiment, the selection of the local evolution combination is specifically based on the matching degree and the evolution uniqueness, i.e., the current frame region corresponds to only one region in the next frame. Specifically, a plurality of candidate evolution combinations corresponding to the current frame region can be determined, the candidate evolution combination with the largest local matching index is selected and recorded as the local evolution combination. At the same time, considering the representativeness of the combination, the local matching index corresponding to the local evolution combination cannot be lower than a preset difference threshold, so as to avoid mistaking a combination with a very low matching degree as a normal region evolution phenomenon. Finally, the plurality of local evolution combinations constructed between the continuous frames are analyzed in series, and according to the matching relationship between the regions, the plurality of region combinations are spliced in the time sequence direction to form continuous component region evolution trajectories across multiple frames, each evolution trajectory represents the continuous expression process of a certain significant structure region, such as a part of a ship body, in the time dimension. The continuity tracking of the structural region in multiple images is realized, and the error accumulation problem caused by direct rough comparison of the whole image is avoided.

[0031] Step S4, dividing the dynamic monitoring image set into a plurality of local monitoring image sets, performing local intensity analysis on each local monitoring image set according to the plurality of component region evolution trajectories, outputting a local intensity image of each local monitoring image set and constructing a local feature set.

[0032] Specifically, the dynamic monitoring image set is divided into multiple local monitoring image sets in a time sliding manner, each local monitoring image set including multiple continuous images to be analyzed, for example, 5 continuous frames divided into a local monitoring image set, thereby forming multiple image subsets with local time sequence correlation. For each local monitoring image set, response strength analysis is performed based on the component region evolution track, the most representative image of the structural expression is extracted, and a local feature set is constructed to effectively exclude images with missing local information caused by occlusion, shaking, etc. during continuous shooting.

[0033] As an optional implementation process, for the construction of the local feature set, specifically includes: According to the structural feature image, the region response strength parameter corresponding to the target component region of each image to be analyzed in the local monitoring image set is calculated, and multiple track aggregates are generated by aggregating multiple target component regions in the local monitoring image set according to multiple component region evolution tracks.

[0034] In this embodiment, the region response strength parameter can be calculated based on the structural feature value of multiple pixels in the target component region, for example, represented by the mean value of the structural feature value. According to the multiple component region evolution tracks constructed as described above, all target component regions are aggregated within the local monitoring image set. The region in each track is aggregated as a track aggregate in each image frame corresponding to the track. Each track aggregate reflects the performance process of the same component evolving over time in multiple image frames.

[0035] The local track response strength parameter of each target component region in each track aggregate is calculated, specifically generated by normalizing the region response strength parameters of multiple target component regions of the track aggregate. If there is no image frame covered, the local track response strength parameter is set to 0. Then, multiple local track response strength parameters of the images to be analyzed are fused to generate a local intensity parameter, specifically the sum of the local track response strength parameters of multiple target component regions. Finally, according to the local intensity parameter, the image to be analyzed with the maximum local intensity parameter is selected as the local intensity image of each local monitoring image set, and a local feature set is constructed according to multiple local intensity images.

[0036] It is worth noting that due to the interference phenomena such as occlusion, water disturbance, etc. are more short-term local phenomena, which are manifested as the interruption of the trajectory in part of the image frames, and reflected as the absence of response intensity in some frames. In this case, the frames with high overall response intensity mean that they contain more complete structural trajectory coverage, thus effectively excluding local interference images. Some traditional frame selection methods such as priority of clarity are easily affected by local interference, and misselect some images with high surface quality but missing structural information. The trajectory response intensity fusion considers the continuous expression ability of regional structure in multiple frames, which can avoid the misleading caused by looking at the instantaneous quality of a frame, and more accurately capture the image frames with complete expression of target structure.

[0037] Step S5, region coverage detection is performed on the plurality of to-be-analyzed images in the local feature set to generate a global feature set corresponding to the local feature set, and the global feature set is processed by the pre-trained ship identification model to generate a target identification result of the target ship.

[0038] Specifically, to further improve the overall coverage ability of the image set on the ship structure information, the images in the local feature set are subjected to region coverage detection, the response intensity and spatial distribution of each target component region are comprehensively evaluated, and finally the image set with the best coverage is selected as the global feature set, which is input into the pre-trained ship identification model for analysis and processing to generate the final target identification result.

[0039] As an optional implementation process, for the construction of the global feature set, specifically includes: The global trajectory response intensity parameter of each target component region in the component region evolution trajectory is calculated, and the trajectory distribution parameter corresponding to each target component region in the local feature set is determined. The global trajectory response intensity parameter is the normalized processing of the corresponding regional response intensity parameter of the target component region in different frames in the entire trajectory, which represents the local response level of the structural intensity in different frames in the trajectory in the evolution process. The trajectory distribution parameter is specifically the reciprocal of the number of evolution frames involved. If the component region evolution trajectory involves ten consecutive frames, the trajectory distribution parameter of the target component region in different evolution frames is one tenth, which represents the scarcity of the intensity in a single evolution frame in the overall trajectory.

[0040] Then, the trajectory distribution parameters of the to-be-analyzed images are weighted and fused according to the global trajectory response intensity parameter, the global trajectory response intensity parameter is used as a weight parameter, the global trajectory response intensity parameter of the target component region under each evolution frame is a fusion object, and the global intensity parameters corresponding to the plurality of to-be-analyzed images in the local feature set are calculated. It should be noted that the fewer the number of image covers of the component region evolution trajectory, that is, the fewer the evolution frames involved, the higher the scarcity of the trajectory, and the greater the global trajectory response intensity parameter of the target component region, which indicates that the feature capture of the current frame to the trajectory region of the region is more structural, for example, the structure is clear and the definition is high. Those regions that appear continuously in multiple frames will not be selected too much, and some regions, such as the ship stern, involve a very short evolution trajectory and appear in only a few frames. In this case, the greater the global intensity parameter.

[0041] After the global intensity parameters corresponding to the plurality of to-be-analyzed images in the local feature set are finally generated, image screening is performed on the local feature set based on the global intensity parameters, so that the most trajectories are covered by as few images as possible, and some images with greater global intensity parameters are screened out to construct a global feature set, thereby effectively improving the structural integrity and region coverage of the global feature set, avoiding redundant frames, and improving the subsequent model processing efficiency.

[0042] Further, the ship identification model can be a model pre-trained based on a neural network, which can be used for intelligent identification of an input global feature image set and output an identification result of a target ship. The specific identification target can be identification information of the ship such as a ship name, a number, and the like, and a type classification of the ship such as a cargo ship, a fishing boat, a passenger ship, and the like, which can be obtained by training a data set constructed by manually labeled samples. The specific structure, training algorithm, and network architecture of the model and the like are common technical means known to those skilled in the art, and the specific design can be flexibly selected according to application requirements. The ship identification model is a well-known technical means to those skilled in the art, and will not be described and limited herein.

[0043] The key innovation of the present application lies in that the feature screening and optimization of multiple images are performed through the structure response intensity and the region evolution relationship, thereby improving the input quality and identification effect of the final identification model. Compared with directly inputting all images into the identification model for processing or performing simple denoising and enhancement preprocessing before inputting into the model, the present application introduces a structure region extraction and trajectory evolution analysis mechanism to screen and process the dynamic monitoring image set, which can extract image frames that have representativeness and continuity in the structure level under the condition that the original image quality is uneven and the local region is shielded or blurred.

[0044] On this basis, through local response intensity analysis and atlas coverage optimization operation, the image frame finally used for identification processing not only has higher structural integrity, but also is more comprehensive in covering different target component areas, which helps the subsequent identification model to obtain more stable and accurate feature information. For the ship dynamic identification scene of the channel portal in a complex environment, due to the interference conditions such as target structure blur, local occlusion, abnormal illumination and the like in the image continuous acquisition process, if all images are used for identification processing without discrimination, it is easy to cause the redundancy images to affect the overall identification performance, and even the model input information may be too messy to cause misjudgment or identification drift and the like. The image screening mechanism of the present application not only can effectively reduce the interference of the redundant images on the identification result, but also can improve the accuracy and robustness of the overall identification under the premise of ensuring the model input efficiency, so that the ship identification system is more suitable for the ship dynamic monitoring and identification scene in a complex environment.

[0045] Figure 2 The structure diagram of a ship dynamic identification system for a channel portal in a complex environment provided by an embodiment of the present application is shown. Please refer to Figure 2 A ship dynamic identification system for a channel portal in a complex environment, comprising: A local feature extraction module, configured to acquire a dynamic monitoring image set of a target ship, the dynamic monitoring image set comprising a plurality of to-be-analyzed images, perform feature extraction on each to-be-analyzed image, and construct a plurality of local feature images of each to-be-analyzed image; A component region generation module, configured to perform feature fusion on the plurality of local feature images of the to-be-analyzed image to generate a structural feature image, determine a plurality of candidate center nodes of each structural feature image, and extract a plurality of target component regions of the to-be-analyzed image based on the plurality of candidate center nodes; A region evolution analysis module, configured to perform region evolution analysis on the plurality of to-be-analyzed images, determine a plurality of local evolution combinations between any two adjacent to-be-analyzed images, and fuse the plurality of local evolution combinations to generate a plurality of component region evolution trajectories; A local intensity analysis module, configured to divide the dynamic monitoring image set into a plurality of local monitoring image sets, perform local intensity analysis on each local monitoring image set according to the plurality of component region evolution trajectories, output a local intensity image of each local monitoring image set, and construct a local feature set; A ship identification optimization module, configured to perform region coverage detection on the plurality of to-be-analyzed images in the local feature set, generate a global feature set corresponding to the local feature set, process the global feature set through a pre-trained ship identification model, and generate a target identification result of the target ship.

[0046] The foregoing is considered as illustrative of the principles of the application. Numerous modifications and changes will readily occur to those skilled in the art, which modifications we desire to protect the scope of the application disclosed in the foregoing description and represented by the drawings.

Claims

1. A method for dynamic identification of ships at waterway checkpoints in complex environments, characterized by: include: Acquire a dynamic monitoring image set of the target ship, the dynamic monitoring image set including a plurality of images to be analyzed, perform feature extraction on each image to be analyzed, and construct a plurality of local feature images of each image to be analyzed; Performing feature fusion on multiple local feature images of the image to be analyzed to generate a structural feature image, determining multiple candidate central nodes of each structural feature image, and extracting multiple target component areas of the image to be analyzed based on the multiple candidate central nodes; Perform regional evolution analysis on multiple images to be analyzed, determine multiple local evolution combinations between any two adjacent images to be analyzed, and fuse multiple local evolution combinations to generate multiple component regional evolution trajectories; The dynamic monitoring image set is divided into multiple local monitoring image sets, local intensity analysis is performed on each local monitoring image set according to the evolution trajectories of multiple component regions, a local intensity image of each local monitoring image set is output, and a local feature set is constructed; Perform regional coverage detection on multiple images to be analyzed in the local feature set to generate a global feature set corresponding to the local feature set. The global feature set is processed by the pre-trained ship recognition model to generate the target ship recognition result.

2. A method for dynamic identification of ships at waterway bayonet points in complex environments according to claim 1, characterized in that: Determine multiple candidate central nodes for each structural feature image, and extract multiple target component regions of the image to be analyzed based on the multiple candidate central nodes, including: For the structural feature image, multiple local feature images of the image to be analyzed include edge feature images, texture feature images and directional gradient feature images. The edge features, texture features and directional gradient features of each pixel in the image to be analyzed are determined, and weighted fusion is performed to obtain the structural feature value of each pixel to generate the structural feature image corresponding to the image to be analyzed; Performing sliding window analysis on the structural feature image, determining multiple local feature extreme value points of the structural feature image in different sliding window areas based on the structural feature values, and marking them as multiple candidate central nodes of the structural feature image; A region growing analysis is performed on multiple candidate central nodes respectively, and a candidate growth region of each candidate central node is constructed. Multiple regional repeated nodes in the structural feature image are determined, and the region allocation index corresponding to each regional repeated node in the associated multiple candidate growth regions is calculated. Based on the region allocation index of the regional repeated nodes, the multiple candidate growth regions are regionally reconstructed to generate multiple target component regions of the image to be analyzed.

3. The method for dynamic identification of ships at waterway bayonet points in complex environments according to claim 2 is characterized in that: Determine multiple local evolution combinations between any two adjacent images to be analyzed, and fuse multiple local evolution combinations to generate multiple component region evolution trajectories, including: Generate multiple candidate evolution combinations according to multiple target component regions corresponding to two adjacent images to be analyzed, calculate the regional overlap parameter and structural matching parameter of each candidate evolution combination, and calculate the local matching index of the candidate evolution combination based on the regional overlap parameter and the structural matching parameter; A plurality of local evolution combinations are determined from a plurality of candidate evolution combinations according to the local matching index, and the plurality of local evolution combinations are evolutionarily spliced ​​to obtain the evolution trajectories of a plurality of component regions of a plurality of images to be analyzed.

4. The method for dynamic identification of ships at waterway bayonet points in complex environments according to claim 3 is characterized in that: Perform local intensity analysis on each local monitoring image set based on the regional evolution trajectories of multiple components, output the local intensity image of each local monitoring image set, and construct a local feature set, including: The regional response intensity parameters corresponding to the target component region of each image to be analyzed in the local monitoring image set are calculated based on the structural feature image. The multiple target component regions in the local monitoring image set are trajectory aggregated according to the evolution trajectories of the multiple component regions to generate multiple trajectory aggregates. The local trajectory response intensity parameters of each target component region in each trajectory aggregate are calculated. The multiple local trajectory response intensity parameters of the image to be analyzed are fused to generate local intensity parameters. The local intensity image of each local monitoring image set is determined based on the local intensity parameters. The local feature set is constructed based on the multiple local intensity images.

5. The method for dynamic identification of ships at waterway checkpoints in complex environments according to claim 4 is characterized in that: Perform regional coverage detection on multiple images to be analyzed in the local feature set to generate a global feature set corresponding to the local feature set, including: The global trajectory response intensity parameter of each target component region in the component region evolution trajectory is calculated, and the trajectory distribution parameters corresponding to multiple target component regions of each image to be analyzed in the local feature set are determined. The trajectory distribution parameters of the image to be analyzed are weightedly fused according to the global trajectory response intensity parameter to generate the global intensity parameters corresponding to the multiple images to be analyzed in the local feature set. Based on the global intensity parameters, the local feature set is image-screened to generate a global feature set.

6. The method for dynamic identification of ships at waterway bayonet points in complex environments according to claim 5 is characterized in that: The ratio between the number of overlapping pixels of the two target components in the candidate evolution combination and the total number of pixels included is calculated as the regional overlap parameter of the candidate evolution combination.

7. A dynamic identification system for ships at waterway bayonets in complex environments, characterized by: The system is used to implement the method for dynamic identification of ships at waterway checkpoints in complex environments as described in any one of claims 1 to 6, comprising: A local feature extraction module is used to obtain a dynamic monitoring image set of the target ship, the dynamic monitoring image set including multiple images to be analyzed, perform feature extraction on each image to be analyzed, and construct multiple local feature images for each image to be analyzed; A component region generation module is used to perform feature fusion on multiple local feature images of the image to be analyzed to generate a structural feature image, determine multiple candidate central nodes of each structural feature image, and extract multiple target component regions of the image to be analyzed based on the multiple candidate central nodes; The regional evolution analysis module is used to perform regional evolution analysis on multiple images to be analyzed, determine multiple local evolution combinations between any two adjacent images to be analyzed, and fuse multiple local evolution combinations to generate multiple component regional evolution trajectories; A local intensity analysis module is used to divide the dynamic monitoring image set into multiple local monitoring image sets, perform local intensity analysis on each local monitoring image set according to the evolution trajectories of multiple component regions, output a local intensity image of each local monitoring image set, and construct a local feature set; The ship recognition optimization module is used to perform regional coverage detection on multiple images to be analyzed in the local feature set, generate a global feature set corresponding to the local feature set, process the global feature set through the pre-trained ship recognition model, and generate the target recognition result of the target ship.

Images