Ultrahigh-resolution image target detection method and system based on weak segmentation

Abstract

The invention discloses an ultrahigh-resolution image target detection method and system based on weak segmentation, and the method comprises the steps: obtaining an original ultrahigh-resolution image, segmenting the original ultrahigh-resolution image through a multi-overlapping-degree weakening segmentation mechanism, and obtaining a plurality of sub-image instance sets with different overlapping degrees; inputting the plurality of sub-graph instance sets into a convolutional neural network feature extractor in parallel, and fusing the multiple weak segmentation features of the plurality of sub-graph instance sets input in parallel to obtain a fused feature graph; and inputting the fused feature map into a target detection model for prediction to obtain a plurality of candidate frames detected by the same target in the sub-maps with different overlapping degrees, processing the plurality of candidate frames by adopting a weak segmentation boundary consistency optimization mechanism, and outputting a final target detection result. High calculation efficiency is kept while small target information is not sacrificed and a target truncation problem is effectively avoided.

Description

Technical Field

[0001] This invention relates to the field of computer vision recognition technology, specifically to a method and system for ultra-high resolution image target detection based on weak segmentation. Background Technology

[0002] In the field of target detection, with the development of imaging technology, ultra-high resolution images (e.g., exceeding 4K, 8K, and even reaching tens of millions of pixels) are becoming increasingly common in security monitoring, remote sensing analysis, and industrial quality inspection. However, the enormous size of these images poses a serious challenge to existing target detection methods.

[0003] Currently, the mainstream methods for processing ultra-high resolution images mainly include the following, along with their drawbacks:

[0004] 1. Global Downsampling: This is the most direct method, which involves scaling the entire ultra-high-resolution image to a standard, processable size (e.g., 1024x1024 pixels) before feeding it into a conventional object detector (e.g., YOLO, Faster R-CNN). Its core drawback is that the downsampling process severely loses image detail. For small targets in the image (such as faces in large security scenarios or small vehicles in remote sensing images), downsampling may leave only a few pixels or cause them to disappear completely, making them unrecognizable by the detector and resulting in a large number of missed detections.

[0005] 2. Tiling / Cropping and Detection: To avoid information loss, the industry commonly uses a tiling method. This method cuts the original large image into several overlapping or non-overlapping sub-images (patches), then detects each sub-image individually, and finally merges the results. This method suffers from the following difficult-to-solve technical problems:

[0006] Object truncation problem: At the boundaries of a slice, a complete object may be cut in half and scattered across different subgraphs. Detectors are prone to missing or false detections when processing these incomplete objects containing incomplete semantic information.

[0007] Limitations of a fixed overlap rate: While setting overlapping regions can alleviate target truncation issues, a fixed overlap rate cannot adapt to targets of different sizes and locations. A large overlap rate generates a lot of redundant computation, reducing detection efficiency; while a small overlap rate may still not guarantee that all targets are completely contained within a sub-image. Choosing the optimal overlap rate is itself a challenging problem.

[0008] Merging the results is difficult: When merging the detection results of each sub-image back into the original large image coordinate system, due to target truncation and overlapping detection, the same target may be detected multiple times, resulting in multiple overlapping bounding boxes. Traditional non-maximum suppression (NMS) algorithms are ineffective in handling this systematic redundancy introduced by the dicing method, and are prone to duplicate detection or false suppression. Summary of the Invention

[0009] The purpose of this invention is to provide a method and system for target detection in ultra-high resolution images based on weak segmentation, which solves the technical problem of how to maintain high computational efficiency while processing ultra-high resolution images without sacrificing small target information and effectively avoiding target truncation.

[0010] To achieve the above objectives, this application adopts the following approach:

[0011] On the one hand, this invention provides a method for target detection in ultra-high resolution images based on weak segmentation, specifically including the following steps:

[0012] S1. Obtain the original ultra-high resolution image, and use the multi-overlap weakening segmentation mechanism to segment the original ultra-high resolution image to obtain multiple sub-image instance sets with different overlap degrees.

[0013] S2. Input multiple subgraph instance sets into the convolutional neural network feature extractor in parallel, and fuse the multiple weak segmentation features of the multiple subgraph instance sets input in parallel to obtain a fused feature map.

[0014] S3. Input the fused feature map into the target detection model for prediction to obtain multiple candidate boxes detected by the same target in sub-maps with different overlap. Use a weak segmentation boundary consistency optimization mechanism to process the multiple candidate boxes and output the final target detection result.

[0015] In some specific implementations, the process of segmenting the original ultra-high resolution image by the multi-overlap weakening segmentation mechanism in step S1 is as follows:

[0016] S11. Define a set of overlap coefficients containing k different overlap coefficients, and set the corresponding virtual segmentation step size for each overlap.

[0017] S12. Using the same convolution kernel, perform convolution operations on the original ultra-high resolution image for each overlap coefficient in the overlap coefficient set according to the virtual segmentation step size corresponding to each overlap coefficient, to obtain k sets of sub-image instance sets with different overlap degrees.

[0018] In some specific implementation schemes, for the set of overlap coefficients R = {r1, r2, ..., r...} k}, where r i∈[0,1) represents the i-th overlap degree, and for each overlap degree r i The corresponding virtual segmentation step size is S i Si=P·(1-r i ), where P represents the window stride of the convolution kernel.

[0019] In some specific implementations, the size of the convolution kernel Ki is P×P, and all element values ​​of the convolution kernel are 1.

[0020] In some specific implementation schemes, the specific process of step S2 is as follows:

[0021] S21. Input multiple subgraph instance sets in parallel into a CNN feature extractor with shared weights to obtain multiple feature maps extracted from multiple subgraph instance sets;

[0022] S22. Multiple feature maps are aggregated using channel-level splicing or weighted fusion strategies to obtain a fused feature map.

[0023] In some specific implementation schemes, the specific process of step S3 is as follows:

[0024] S31. Input the fused feature map into the object detection model for prediction to obtain multiple candidate boxes, which come from different regions in the fused feature map.

[0025] S32. Multiple candidate boxes are deduplicated and fused using an overlap-aware and weighted average fusion strategy to obtain the final target detection result.

[0026] In some specific implementation schemes, the specific process of step S32 is as follows:

[0027] Each candidate box is represented as: Bj = (x j ,y j ,w j ,h j ,s j ,c j ), where (x j ,y j ,w j ,h j ) represents the coordinates and size of the candidate bounding box in the fused feature map, s j c represents the confidence score. j Indicates category;

[0028] Based on the coordinates of the candidate boxes, the feature maps corresponding to the candidate boxes are located, and the overlap information of the candidate boxes is obtained.

[0029] Based on the confidence scores of the detection boxes, multiple candidate boxes are sorted, the IoU between each candidate box and the current detection box is calculated, and candidate boxes with IoU higher than a preset threshold are filtered out.

[0030] The final target detection result is obtained by weighting each candidate box based on the overlap information and confidence score, and then performing a weighted average fusion.

[0031] In some specific implementations, the overlap information corresponding to the candidate boxes is obtained, including the position of the target in the candidate box in the feature map. Different weights are assigned to the remaining candidate boxes after filtering based on the target position in the candidate box. The distance from the target to the center of the feature map is calculated based on the target position, and the weights are assigned from low to high according to the distance.

[0032] In some specific implementations, when calculating IoU, a preset threshold for IoU is adjusted based on overlap information.

[0033] Secondly, this application provides an ultra-high resolution image target detection system based on weak segmentation, comprising:

[0034] One or more processors;

[0035] A storage unit is used to store one or more programs that, when executed by one or more processors, enable the one or more processors to implement the ultra-high resolution image target detection method based on weak segmentation as described in the first aspect.

[0036] The beneficial effects of this invention are as follows:

[0037] 1. This application proposes the application and fusion of the concept of "multiple overlap" in image segmentation, explicitly proposing and implementing multiple variable overlaps, and performing feature extraction and fusion through channel parallelism. This enables the model to integrate information from views with different degrees of overlap, far exceeding traditional segmentation methods with a single fixed overlap or no overlap.

[0038] 2. This application proposes a soft segmentation mechanism, using a large convolutional kernel initialized to 1 for all kernel values ​​as a special "region aggregation / perception" tool to generate a set of subgraphs with multiple overlaps, rather than traditional hard pruning or feature extraction. This is a novel input processing method specifically designed to address the problem of local information loss. In high-resolution scenarios, it ensures the integrity of local information without significantly increasing computational costs. By combining the channel-parallel computing capabilities of CNNs, it ensures information integrity at the input end through weak segmentation, while avoiding the computational explosion or information loss problems that may occur in traditional methods, thus achieving a balance between performance and efficiency.

[0039] 3. This application adopts weak segmentation boundary consistency optimization. To address the boundary problems that may occur after the fusion of multiple overlapping subgraphs, it proposes an overlap-aware deduplication and refinement strategy to ensure the smoothness and accuracy of the final detection results. This is a complete closed loop for the weak segmentation scheme. Attached Figure Description

[0040] Figure 1 A flowchart of a high-resolution image target detection method based on weak segmentation provided in an embodiment of the present invention;

[0041] Figure 2 A flowchart illustrating the multi-overlap weak segmentation mechanism provided in this embodiment of the invention;

[0042] Figure 3 This is a flowchart of the channel parallel feature fusion processing provided in an embodiment of the present invention;

[0043] Figure 4 A flowchart of weak segmentation boundary consistency optimization processing provided for embodiments of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0046] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0047] Furthermore, for clarity and brevity, descriptions of well-known structures, functions, and configurations may have been omitted. Those skilled in the art will recognize that various changes and modifications can be made to the examples described herein without departing from the spirit and scope of this disclosure.

[0048] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0049] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0050] Example 1

[0051] like Figure 1 As shown, this embodiment provides a method for target detection in ultra-high resolution images based on weak segmentation, specifically including the following steps:

[0052] S1. Obtain the original ultra-high resolution image, and use the multi-overlap weakening segmentation mechanism to segment the original ultra-high resolution image to obtain multiple sub-image instance sets with different overlap degrees.

[0053] Step S1, the multi-overlap weakening segmentation mechanism, performs the following segmentation process on the original ultra-high resolution image:

[0054] S11. Define a set of overlap coefficients containing k different overlap coefficients, and set the corresponding virtual segmentation step size for each overlap.

[0055] S12. Using the same convolution kernel, perform convolution operations on the original super-resolution image for each overlap coefficient in the overlap coefficient set, according to the virtual segmentation stride corresponding to each overlap coefficient, to obtain k sets of sub-image instance sets with different overlap degrees. For the overlap coefficient set R = {r1, r2, ..., r...} k}, where r i ∈[0,1) represents the i-th overlap degree, and for each overlap degree r i The corresponding virtual segmentation step size is S i Si=P·(1-r i P represents the window stride of the convolution kernel, the size of the convolution kernel Ki is P×P, and all elements of the convolution kernel have a value of 1.

[0056] S2. Input multiple subgraph instance sets into the convolutional neural network feature extractor in parallel, and fuse the multiple weak segmentation features of the multiple subgraph instance sets input in parallel to obtain a fused feature map.

[0057] The specific process of step S2 is as follows:

[0058] S21. Input multiple subgraph instance sets in parallel into a CNN feature extractor with shared weights to obtain multiple feature maps extracted from multiple subgraph instance sets;

[0059] S22. Multiple feature maps are aggregated using channel-level splicing or weighted fusion strategies to obtain a fused feature map.

[0060] S3. Input the fused feature map into the target detection model for prediction to obtain multiple candidate boxes detected by the same target in sub-maps with different overlap. Use a weak segmentation boundary consistency optimization mechanism to process the multiple candidate boxes and output the final target detection result.

[0061] The specific process of step S3 is as follows:

[0062] S31. Input the fused feature map into the object detection model for prediction to obtain multiple candidate boxes, which come from different regions in the fused feature map.

[0063] S32. Multiple candidate boxes are deduplicated and fused using an overlap-aware and weighted average fusion strategy to obtain the final target detection result.

[0064] The specific process of step S32 is as follows:

[0065] Each candidate box is represented as: Bj = (x j ,y j ,w j ,h j ,s j ,c j ), where (x j ,y j ,w j ,h j ) represents the coordinates and size of the candidate bounding box in the fused feature map, s j c represents the confidence score. j Indicates category;

[0066] Based on the coordinates of the candidate boxes, the feature maps corresponding to the candidate boxes are located, and the overlap information of the candidate boxes is obtained.

[0067] Based on the confidence scores of the detection boxes, multiple candidate boxes are sorted, and the IoU between each candidate box and the current detection box is calculated. Candidate boxes with IoU higher than a preset threshold are filtered out. When calculating IoU, the preset threshold of IoU can also be adjusted according to the overlap information.

[0068] The final target detection result is obtained by weighting each candidate box based on the overlap information and confidence score, and then performing a weighted average fusion.

[0069] The overlap information corresponding to the candidate boxes is obtained, including the position of the target in the candidate box in the feature map. Different weights are assigned to the remaining candidate boxes after filtering based on the target position in the candidate box. The distance from the target to the center of the feature map is calculated based on the target position, and the weights are assigned from low to high according to the distance.

[0070] This application is applicable to various scenarios, including: security monitoring, detecting distant, small-sized personnel, vehicles, or abnormal behavior in ultra-high-definition surveillance video streams; remote sensing image analysis, performing feature identification, target counting, and change detection (such as aircraft, ships, and buildings) on large-area, high-resolution images acquired by satellites and drones; industrial defect detection, detecting minute cracks, scratches, foreign objects, and other defects in high-resolution industrial product surface images; medical imaging diagnosis, detecting minute lesions and cellular abnormalities in high-resolution pathological sections or images; and intelligent transportation, performing refined detection of vehicles, pedestrians, and traffic signs in urban road and highway surveillance images. To better illustrate this, the following example, "aircraft detection in high-altitude remote sensing images," will be used to explain the specific implementation process of this invention.

[0071] Scenario setting: Suppose we have an ultra-high resolution remote sensing image of 20,000x20,000 pixels, and the task is to detect all the aircraft in it. These aircraft vary in size and are randomly distributed.

[0072] Implementation steps:

[0073] Step 1: Weak segmentation with multiple overlap does not directly scale or hard-segment the 20000x20000 image. Instead, it employs the weak segmentation mechanism with multiple overlap of this invention. A basic sub-image (receptive field) size is defined, for example, P = 800 pixels, and two sets of overlap coefficients are set, for example, R = {r1 = 0.2, r2 = 0.5}.

[0074] For an overlap of r1 = 0.2, a convolution kernel of size 800×800 with all values ​​set to 1 is used to perform a convolution operation on the original large image with a stride of S1 = 800·(1-0.2) = 640 pixels. This generates the first set of "soft sub-images" Isub_r1. Simultaneously, for an overlap of r2 = 0.5, the same convolution kernel is used with a stride of S2 = 800·(1-0.5) = 400 pixels to generate the second set of "soft sub-images" Isub_r2.

[0075] Key point: The convolution operation here is not for extracting complex image features, but rather as a "region-aware" tool. It generates local views from different perspectives in a "soft" way (rather than hard cutting). The view with an overlap of 0.5 contains more contextual redundancy than the view with an overlap of 0.2.

[0076] Step 2: Parallel feature extraction and fusion of channels. The two sets of "soft subgraphs" Isub_r1 and Isub_r2 generated above are fed in parallel into a CNN feature extraction network with shared weights (e.g., ResNet as the backbone network).

[0077] After passing through a CNN, Isub_r1 is processed to obtain feature map Xsub_r1 (e.g., with dimensions [N1, 256, H, W]).

[0078] After passing through a CNN, Isub_r2 is used to obtain the feature map Xsub_r2 (e.g., with dimensions [N2, 256, H, W]).

[0079] Then, a channel-parallel multi-weak segmentation feature fusion strategy is adopted. The two sets of feature maps are concatenated along the channel dimension:

[0080] Xfused=Concat(Xsub_r1,Xsub_r2)

[0081] A fused feature map with dimensions [N_fused, 512, H, W] is obtained. This feature map contains information under two levels of overlap, making it richer and more robust than the features of a single view.

[0082] Step 3: Detection and Boundary Consistency Optimization. The fused feature map is fed into an object detection head (such as the YOLOv8 detection head or Faster R-CNN) for prediction. Due to the overlap, an aircraft located at the boundary of two original slices is now likely to be completely detected in multiple "soft sub-maps".

[0083] For example, if an aircraft A is detected three times, each time from a different overlapping subgraph, three candidate boxes B1, B2, and B3 are generated. In this case, a weak segmentation boundary consistency optimization mechanism is activated.

[0084] It not only considers the confidence scores and IoU of the three boxes, but also analyzes whether aircraft A is located in the central or peripheral region of the "soft submap" that generated these three candidate boxes.

[0085] Suppose B1 comes from a subgraph where the aircraft is at its center, while B2 and B3 come from subgraphs where the aircraft is at its edge. In this case, overlap-aware NMS would give B1 a higher weight.

[0086] Finally, through weighted average fusion, B1, B2, and B3 are merged into a unique final bounding box B_final with a more precise location.

[0087] Implementation Results: Through the above steps, this invention successfully detected all aircraft in remote sensing images, effectively avoiding the problem of aircraft targets being "cut off" due to hard slicing. At the same time, through the fusion of multiple views and intelligent result merging, the accuracy and robustness of detection are significantly improved.

[0088] It is understandable that the solution in this application can be broken down into three important steps: image segmentation based on a multi-overlap weak segmentation mechanism, multi-weak segmentation feature fusion based on channel parallelism, and a weak segmentation boundary consistency optimization strategy. The following explains these three steps from a more easily understandable perspective:

[0089] 1. Weak segmentation mechanism with multiple overlaps

[0090] Brief explanation and examples:

[0091] Imagine trying to read a huge painting using a fixed frame (like a square cut out of an A4 sheet of paper) to find the bird depicted. The traditional "hard-slice" method is like moving the frame flat without overlap. If a bird lands on the boundary between two frames, you can only see half of the bird each time, and you might not be able to recognize it.

[0092] The "multiple overlap weak segmentation mechanism" of this invention completely changes this approach.

[0093] The unique aspect of "weak segmentation" is that it no longer uses "scissors" to rigidly cut the image. Instead, it designs a special "viewing mirror"—a large convolutional kernel with all values ​​set to 1. This "viewing mirror" does not change the content of the image; instead, it acts like a region aggregator, packaging all pixel information within the mirror into a "local view." Specifically, when this convolutional kernel (of size P×P) sweeps across the image, it performs a convolution operation with a P×P region in the image. However, because the kernel values ​​are all 1, it actually aggregates all pixel information within that region, rather than extracting specific features as in ordinary convolution. The output of this convolution operation retains the complete information of the original region, simply reorganizing it into a new representation. This method is "soft," completely preserving all information within the region rather than rigidly cutting it.

[0094] The advantage of "multiple overlap" is that it moves this "observation mirror" in more than one way. It moves it with multiple different overlap amplitudes. For example, the first move overlaps only 20% of the field of view each time, meaning the convolutional kernel moves with a stride of 80% of the window size each time; the second move overlaps 50% of the field of view, with the kernel stride at 50% of the window size. By adjusting the stride parameter of the convolution operation, the degree of overlap for each scan can be precisely controlled. This is like observing the scroll back and forth in different, more careful ways, ensuring that the bird that landed on the boundary will eventually be fully and clearly presented in the center of the "observation mirror." This combination of multi-stride convolutions allows any target, regardless of its location in the image, to be completely captured in at least one "soft sub-image." In this way, multiple sets of high-quality, lossless local views are prepared for subsequent recognition work, fundamentally solving the problem of targets being cut off at boundaries.

[0095] Mathematical explanation and logical description:

[0096] Traditional image segmentation methods typically employ fixed window sizes and strides, or non-overlapping tile generation, which can easily lead to the target being truncated at the segmentation boundaries, resulting in the loss of crucial local information, especially for small targets in high-resolution images. This application introduces a multi-overlap weak segmentation mechanism, aiming to segment high-resolution images in a "soft" manner while ensuring the integrity of local information.

[0097] like Figure 2 As shown, assume the original high-resolution input image is I∈R H×W×C Where H is the height, W is the width, and C is the number of channels. Define a set of overlap coefficients R = {r1, r2, ..., r k}, where ri∈[0,1) represents the i-th overlap. For each overlap ri, it is achieved by adjusting the virtual "segmentation step size" Si. When the window size is P×P, the traditional step size is P, and the overlap is 0. In order to achieve the overlap of ri, the effective step size is Si=P·(1-ri).

[0098] To avoid the rigid segmentation of traditional sliding windows, a "soft" segmentation method based on a large convolutional kernel with all kernel values ​​set to 1 is adopted to generate sub-images with different overlap levels. Specifically, for each overlap level ri, a special convolution operation Ci(·) is constructed, with a kernel Ki of size P×P and all element values ​​set to 1 (i.e., Ki(x,y)=1). This convolutional kernel acts similarly to an unweighted average pooling or a "region-aware" operation. It is not for feature extraction, but rather for aggregating pixel information within a specific size region to form a local view within a "receptive field." By adjusting the stride Si of the convolution, a set of "soft sub-images" {Isub_r1, Isub_r2, …, Isub_rk} with different overlap levels ri can be generated. Each "soft sub-image" Isub_ri is not a strict crop of the original image, but rather an aggregated representation of local regions in the original image, with its boundary information effectively preserved and redundant through the overlap level.

[0099] 2. Multi-weak segmentation feature fusion based on channel parallelism

[0100] Brief explanation and examples:

[0101] Continuing the example above, we now have several sets of "photographs" recording local information of the painting (one set with a 20% overlap and another with a 50% overlap) obtained through the "multiple overlap" observation method. How can we utilize these photographs? The traditional approach might be to look at one set after another, which is inefficient and results in isolated information. This application provides a more efficient channel-parallel feature fusion scheme.

[0102] Special features and advantages: These different "photo sets" are simultaneously fed into a powerful recognition brain (CNN network). This brain can process information in parallel. The key to fusion lies in the "channel" dimension. You can think of a channel as the "color layer" of an image (such as red, green, and blue). The information from the set of photos with a 20% overlap rate is "overlaid" with the information from the set of photos with a 50% overlap rate along the "channel" dimension.

[0103] This is akin to creating a new, richer "color photograph." It not only contains red, green, and blue, but also two entirely new "color layers": a "20% overlap perspective" and a "50% overlap perspective." In this way, the brain, during analysis, can simultaneously see all the information from different viewing methods, resulting in a more comprehensive and profound understanding of the content. This parallel processing and fusion approach neither adds much extra time cost nor significantly enriches the dimensions of information.

[0104] Mathematical explanation and logical description:

[0105] Through the previous step's multi-overlap weakening segmentation mechanism, multiple sets of sub-images with different overlap degrees were obtained. To fully utilize this information and address the computational burden caused by high-resolution images, a channel-parallel multi-weak segmentation feature fusion strategy is further proposed.

[0106] like Figure 3 As shown, for each overlap ri, the resulting set of subgraphs Isub_ri is fed in parallel into a shared or independent convolutional neural network (CNN) feature extractor F(·). CNNs are inherently good at channel-parallel computation, which means that feature extraction of different subgraphs can be performed efficiently in parallel on computing units such as GPUs.

[0107] Let X = sub_riF(Isub_ri) be the feature map extracted from the set of sub-maps with the corresponding overlap ri. Since these feature maps all come from different overlapping regions of the same original image, they have a natural correspondence in the channel dimension. Channel-level concatenation or weighted fusion strategies are used to aggregate these feature maps.

[0108] Channel-level splicing: X fused =Concat(Xsub_r1,Xsub_r2,…,Xsub_rk)

[0109] Concat(·) indicates concatenation along the channel dimension. This method preserves feature information across all overlap levels and integrates them into a wider feature vector.

[0110] Weighted fusion:

[0111]

[0112] Here, αi can be learnable weights (e.g., dynamically generated via attention mechanisms such as SENet or CBAM modules). This approach allows the model to adaptively assign importance to features with varying degrees of overlap.

[0113] By employing this parallel fusion approach, the model can simultaneously perceive and integrate local information at different levels of overlap, effectively avoiding the loss of local information. Simultaneously, the computational load is distributed across parallel channels, thus not significantly increasing the overall computational cost. The fused features (Xfused) contain richer and more robust local representations, providing high-quality input for subsequent object detection.

[0114] 3. Optimization of consistency at weak partition boundaries

[0115] Brief explanation and examples:

[0116] Through the first two steps, the intelligent detection system is now very powerful; it may have detected the same bird in different "local views" and provided its location. For example, it found it in a view with 20% overlap and also in a view with 50% overlap. Now it has multiple "reports" about the bird's location. How do we determine its most accurate and unique location? Traditional NMS methods are like simple referees, choosing whichever "report" is loudest (highest confidence). But this is not intelligent enough. This application proposes a smarter weak segmentation boundary consistency optimization decision mechanism.

[0117] The unique feature and advantage of this mechanism: The weak segmentation boundary consistency optimization decision-making mechanism not only listens to who has the loudest voice, but also traces the source of each "report". It considers: Was this "report" generated in a view with a wide field of vision and the bird in the center, or in a view with the bird at the edge of the field of vision and somewhat blurry?

[0118] For example, the "report" that captures the bird completely in the center of the field of vision will be given higher weight and trust. Then, all the "reports" about this bird will be subject to a "weighted democratic vote" (weighted average fusion) instead of a crude "one-vote-wins-all" approach. Through this more refined and context-aware optimization method, targets that have been detected multiple times can be smoothly and accurately merged into a single result, greatly improving the accuracy and reliability of the final localization.

[0119] Mathematical explanation and logical description:

[0120] Although weak segmentation preserves local information through overlap, inconsistencies in detection results across subgraph boundaries may still exist during the final target detection result aggregation stage. To address this issue, this application introduces a weak segmentation boundary consistency optimization mechanism. For target boxes detected in subgraphs with different overlap levels, deduplication and refinement are required during the fusion stage. Traditional non-maximum suppression (NMS) may not adequately handle redundant or conflicting predictions from views with different overlap levels. Therefore, an overlap-aware NMS or weighted average fusion strategy is proposed.

[0121] For each detected candidate box, the boundary Bj = (x j ,y j ,w j ,h j ,s j ,c j ), where (x j ,y j ,w j ,h j) represents the bounding box coordinates and dimensions, sj represents the confidence score, and cj represents the class. When multiple submaps detect the same object, multiple predicted candidate boxes will be generated.

[0122] Overlap-Aware NMS: Traditional NMS typically only considers IoU (Intersection over Union). This application introduces an additional dimension: the overlap information of the subgraphs from which the predicted candidate boxes originate. For example, candidate boxes originating near the center of the overlapping region can be given higher priority or weight. Alternatively, when calculating IoU, the IoU threshold can be adjusted based on overlap information, thereby more intelligently filtering redundant boxes.

[0123] Weighted average fusion: For multiple candidate boxes identified as the same target, their position, size, and confidence scores can be averaged based on their confidence scores and the overlap of their respective subgraphs. For example:

[0124] Here, wj is the weight, which can be determined by the confidence score sj and the overlap ri of the subgraph containing the bounding box. For example, w j ∝s j ·f(r i ), where f(ri) is a weighted function based on overlap.

[0125] This optimization ensures that targets detected under different weakly segmented views can be smoothly and accurately aggregated, reducing duplicate detections and false alarms, and improving the robustness and accuracy of the final detection results.

[0126] The fundamental difference between the weak segmentation boundary consistency optimization mechanism proposed in this application and the traditional WBF is that the fusion strategy of the weak segmentation boundary consistency optimization mechanism specifically utilizes the contextual information (overlap itself) of candidate boxes from weak segmentation sub-graphs with different overlap to solve the segmentation boundary problem, while WBF is a more general bounding box fusion method that mainly performs weighted merging based on confidence.

[0127] The proposed solution differs from existing solutions in the following ways:

[0128] 1. Unlike traditional segmentation techniques, the "weak segmentation" proposed in this application is not a simple sliding window or tile-based approach. Its multiple overlaps and "soft" segmentation method based on large convolution kernel values ​​of all 1s are fundamentally different from existing technologies in both concept and implementation. This paper will emphasize the limitations of traditional methods (such as information fragmentation and redundant computation) and how this patent effectively overcomes these limitations.

[0129] 2. Unlike multi-scale feature fusion networks (such as FPN / PANet), existing techniques like FPN primarily focus on fusing feature maps from different levels within a CNN to handle multi-scale targets. This application proposes a "weak segmentation" strategy for the input image layer, which ensures the integrity of local information through multiple overlaps before feature extraction. This is a more upfront and fundamental optimization. The method emphasized here is a complement to, rather than a replacement for, existing feature fusion techniques, addressing the problem of earlier information loss.

[0130] 3. Distinguishing itself from existing convolutional kernel applications: Traditional convolutional kernels are mainly used for feature extraction, and their weights are learned. This application uses large convolutional kernels with all values ​​of 1 to generate overlapping subgraphs, which is a functional innovation rather than traditional feature extraction. This specific purpose will be highlighted to differentiate it from conventional convolutional operations.

[0131] 4. Emphasis on comprehensive solutions: This application is not just about a single technical point, but forms a systematic combination of "weak segmentation mechanism + channel parallel fusion + boundary consistency optimization" to jointly solve the core problems in high-resolution image detection.

[0132] Example 2

[0133] This embodiment provides an ultra-high resolution image target detection system based on weak segmentation, including:

[0134] One or more processors;

[0135] A storage unit is used to store one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the ultra-high resolution image target detection method based on weak segmentation of Embodiment 1.

[0136] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Based on the technical essence of the present invention, any simple modifications, equivalent substitutions, and improvements made to the above embodiments within the spirit and principles of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A target detection method for ultra-high resolution images based on weak segmentation, characterized in that, Specifically, the following steps are included: S1. Obtain the original ultra-high resolution image, and use the multi-overlap weakening segmentation mechanism to segment the original ultra-high resolution image to obtain multiple sub-image instance sets with different overlap degrees. S2. Input multiple subgraph instance sets into the convolutional neural network feature extractor in parallel, and fuse the multiple weak segmentation features of the multiple subgraph instance sets input in parallel to obtain a fused feature map. S3. Input the fused feature map into the target detection model for prediction to obtain multiple candidate boxes detected by the same target in sub-maps with different overlap. Use a weak segmentation boundary consistency optimization mechanism to process the multiple candidate boxes and output the final target detection result.

2. The ultra-high resolution image target detection method based on weak segmentation according to claim 1, characterized in that, Step S1, the multi-overlap weakening segmentation mechanism, performs the following segmentation process on the original ultra-high resolution image: S11. Define a set of overlap coefficients containing k different overlap coefficients, and set the corresponding virtual segmentation step size for each overlap. S12. Using the same convolution kernel, perform convolution operations on the original ultra-high resolution image for each overlap coefficient in the overlap coefficient set according to the virtual segmentation step size corresponding to each overlap coefficient, to obtain k sets of sub-image instance sets with different overlap degrees.

3. The ultra-high resolution image target detection method based on weak segmentation according to claim 2, characterized in that, For the set of overlap coefficients R = {r1, r2, ..., r...} k }, where r i ∈[0,1) represents the i-th overlap degree, and for each overlap degree r i The corresponding virtual segmentation step size is S i Si=P·(1-r i ), where P represents the window stride of the convolution kernel.

4. The ultra-high resolution image target detection method based on weak segmentation according to claim 3, characterized in that, The kernel Ki has a size of P×P, and all elements of the kernel have a value of 1.

5. The ultra-high resolution image target detection method based on weak segmentation according to claim 1, characterized in that, The specific process of step S2 is as follows: S21. Input multiple subgraph instance sets in parallel into a CNN feature extractor with shared weights to obtain multiple feature maps extracted from multiple subgraph instance sets; S22. Multiple feature maps are aggregated using channel-level splicing or weighted fusion strategies to obtain a fused feature map.

6. The ultra-high resolution image target detection method based on weak segmentation according to claim 5, characterized in that, The specific process of step S3 is as follows: S31. Input the fused feature map into the object detection model for prediction to obtain multiple candidate boxes, which come from different regions in the fused feature map. S32. Multiple candidate boxes are deduplicated and fused using an overlap-aware and weighted average fusion strategy to obtain the final target detection result.

7. The ultra-high resolution image target detection method based on weak segmentation according to claim 6, characterized in that, The specific process of step S32 is as follows: Each candidate box is represented as: Bj = (x j ,y j ,w j ,h j ,s j ,c j ), where (x j ,y j ,w j ,h j ) represents the coordinates and size of the candidate bounding box in the fused feature map, s j c represents the confidence score. j Indicates category; Based on the coordinates of the candidate boxes, the feature maps corresponding to the candidate boxes are located, and the overlap information of the candidate boxes is obtained. Based on the confidence scores of the detection boxes, multiple candidate boxes are sorted, the IoU between each candidate box and the current detection box is calculated, and candidate boxes with IoU higher than a preset threshold are filtered out. The final target detection result is obtained by weighting each candidate box based on the overlap information and confidence score, and then performing a weighted average fusion.

8. The ultra-high resolution image target detection method based on weak segmentation according to claim 7, characterized in that, The overlap information corresponding to the candidate boxes is obtained, including the position of the target in the candidate box in the feature map. Different weights are assigned to the remaining candidate boxes after filtering according to the position of the target in the candidate box. The distance from the target to the center of the feature map is calculated according to the target position, and the weights are assigned from low to high according to the distance.

9. The ultra-high resolution image target detection method based on weak segmentation according to claim 7, characterized in that, When calculating IoU, the preset threshold of IoU is adjusted based on the overlap information.

10. A target detection system for ultra-high resolution images based on weak segmentation, characterized in that, include: One or more processors; A storage unit is used to store one or more programs that, when executed by one or more processors, enable the one or more processors to implement a weak segmentation-based ultra-high resolution image target detection method as described in any one of claims 1-9.

Images