Computer vision object positioning system
By introducing lightweight backbone network and BEM modules into the drone target detection system, combining reverse attention exploration network and challenge area predictor, the detection accuracy of the drone target detection system in small targets and complex scenarios is solved. and the problem of low computing efficiency, more efficient and accurate object detection is achieved.
Patent Information
- Application Number
- CN202410408674.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-04-07
- Publication Date
- 2025-02-11
AI Technical Summary
When processing drone view images, existing drone target detection systems have problems of algorithm adaptability and low computing efficiency, especially when detecting small targets and complex scenarios, it is difficult to achieve high accuracy and real-time.
Adopting lightweight backbone network and branch enhancement module (BEM), combined with reverse attention exploration network and challenge area predictor, improve detection accuracy and calculation of challenging targets through feature extraction, image amplification and non-maximum suppression processing. efficiency.
While maintaining detection performance, it significantly reduces computing and storage requirements, adapts to the drone environment, and improves detection accuracy and robustness for small targets and complex scenarios.
Smart Images
Figure CN119417888A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of computer vision, and in particular to a computer vision object positioning system. Background Art
[0002] Object detection is a key task in computer vision, which aims to identify and locate specific objects in images or videos. This technology plays an important role in many applications, including autonomous driving, video surveillance, medical image analysis, industrial quality inspection, etc. The main difference between object detection and image classification is that object detection not only needs to identify the object category in the image, but also needs to accurately calibrate the location of the object, usually in the form of a bounding box.
[0003] The main challenges of existing drone target detection based on deep learning algorithms are the adaptability and computational efficiency of the algorithms. Since there are a large number of difficult targets such as small targets, occluded targets, and densely populated targets in drone view images, target detection remains challenging. When using existing models without special design, the detection performance of drone view images is significantly different from that of natural images. Drone small target detection is constrained by sensor resolution and lack of target features, which makes it difficult for the algorithm to accurately identify small-sized targets. The key to solving this problem is to develop target detection algorithms that can make full use of limited information, which may involve technologies such as feature enhancement and multi-scale processing.
[0004] In addition, most of the current drone target detection systems often have the problem of high computational complexity. High-complexity algorithms require more computing resources, but the computing power of drones is limited, which may lead to poor real-time performance and slow response speed, especially when dealing with a large number of small targets. Therefore, in order to meet the limited computing resources of drones, a lightweight target detection model must be designed. Summary of the invention
[0005] The present invention provides a computer vision object positioning system, which solves the problems raised by the above background technology.
[0006] The present invention provides the following technical solution: a computer vision object positioning system, comprising the following steps:
[0007] Step S1: Use the UAV aerial photography Visdrone2019 dataset for training and testing to ensure the adequacy and diversity of the dataset;
[0008] Step S2: The resolution of the dataset images is adjusted to adapt to the lightweight backbone network. The resolution of all dataset images is adjusted to 640×640 as the input of the lightweight backbone network of the global detection part. The images are downsampled through conv_1 and conv_2, and a lightweight feature extraction network composed of branch enhancement module (BEM) and conv is used for feature extraction and image downsampling.
[0009] Step S3: Apply the global features obtained in the above steps to global detection and use them as input to the reverse attention exploration network to generate the coordinates of the challenge area to assist subsequent local detection;
[0010] Step S4: amplifying and restoring the target image with high quality according to the generated challenge area coordinates;
[0011] Step S5: Feature extraction is performed on the enlarged challenging area image. After the global detector and the local detector output prediction results respectively, non-maximum suppression processing is performed to eliminate redundant prediction bounding boxes.
[0012] Preferably, step S2 specifically includes, in the lightweight feature extraction network, introducing a BEM module, using 16-way branch 1×1 convolution to integrate the input feature information, and then using 3×3 convolution to extract features, wherein the number of these low-dimensional feature channels is 1 / 32 of the input feature map, and then merging the low-dimensional feature map branches, and using 1×1 convolution feature splicing, and finally performing BN normalization to prevent gradient loss. The BEM module algorithm is as follows:
[0013]
[0014] Where x represents the input feature map, y represents the output feature map, conv1 represents 1×1 convolution, the channel is 1 / 2 of the input feature map channel, Γi(x) represents the input feature map downsampling to a low-dimensional feature map, and all Γi(x) have the same structure, including 1×1 convolution and 3×3 convolution, the number of output feature map channels is 1 / 32 of the number of input feature map channels, C represents the number of branches, and the network is set to 16, conv2 represents the 1×1 convolution of the combined feature, and BN represents the BatchNorm normalization operation.
[0015] Preferably, step S3 specifically includes, using a challenging region predictor (CRP) to estimate the probability of a difficult-to-detect object appearing in the image under the guidance of a reverse attention mechanism;
[0016] The Challenge Region Predictor (CRP) consists of an encoder, two feature extraction blocks, a fusion block, and two decoders;
[0017] The input global features are encoded, feature extracted, and decoded to generate a simple object distribution attention mask. The encoder first reduces the dimension of the global features to reduce the amount of calculation;
[0018] The algorithm is as follows:
[0019] E encoded =Sigmoid(BN(Conv(Conv(G global ,1 1),3 3))) (2)
[0020] The feature extraction block consists of two 3×3 convolutions, BN normalization, and Sigmoid activation layers. Each convolution layer is followed by a Sigmoid activation layer to introduce nonlinearity so that the network can better adapt to challenging scenarios. The algorithm is shown below.
[0021] F extracted =Sigmoid(BN(Conv(Conv(E encoded ,3 3),3 3))) (3)
[0022] The decoder consists of a 3×3 and a 1×1 convolution, BN normalization, and Sigmoid activation layer, followed by an inverted attention mask to generate a reverse mask containing the challenge area, which excludes the occupied simple object area. The algorithm is shown below.
[0023] M reverse =1-Sigmoid(BN(Conv(Conv(F extracted ,3 3),1 1))) (4)
[0024] Preferably, the specific steps of distinguishing and locating the difficult objects by the region predictor (CRP) include: pixel-wise multiplication of the inverted attention mask and the feature obtained by extracting features in the previous step and adding and splicing them to obtain the challenge region feature, and then extracting features and decoding operations to obtain the challenge region mask, which indicates the region where the difficult object is located through the decoder;
[0025] The feature extraction operation consists of two 3×3 convolutions, BN normalization, and Sigmoid activation layers. The decoding operation decodes the challenge area features after feature extraction through a 3×3 convolution layer and a 1×1 convolution layer, and then undergoes BN normalization and Sigmoid activation again.
[0026] Map high-level features to masks of the challenge area, and use the mask model to accurately indicate the area where the difficult object is located. The algorithm is as follows:
[0027]
[0028]
[0029] After obtaining the rough position of the difficult target, it is expanded into a vector, and nonlinear transformation and feature extraction are performed through multiple hidden layers to map the input data to the output result. During the forward propagation process, the input data is processed by multiple hidden layers to finally generate the prediction result.
[0030] Preferably, step S4 specifically includes, first, obtaining coordinate information of the challenge area, obtained from the above steps, the position of the challenging target indicated by the coordinates in the original high-resolution image;
[0031] Use bilinear interpolation to enlarge the image. For each pixel in the challenge area, according to the position of its coordinates in the original image, perform interpolation pixel by pixel in the entire challenge area, and complete the overall interpolation by linear weighted averaging the values of the four surrounding pixels to obtain the interpolated pixel value;
[0032] Finally, the interpolated image is integrated into the original high-resolution image to form an enlarged image.
[0033] Preferably, step S5 specifically includes, firstly, taking the magnified challenging area image in step S4 as input, constructing a feature extraction network, and gradually improving the abstract level of features through convolution operation;
[0034] In the feature extraction network, the BEM module is introduced through the fusion and concatenation of multi-way convolution branches and low-dimensional feature maps;
[0035] The activation function and normalization operation are applied to each convolution layer, and the feature map is downsampled through the pooling layer; the output feature map with a higher level representation is obtained.
[0036] Finally, after the global detector and the local detector each output their predictions, a subsequent processing step, Non-Maximum Suppression (NMS), is performed to eliminate redundant predicted bounding boxes.
[0037] Preferably, a weight allocation mechanism is also included to assign a higher weight to the result output by the local detector, and the algorithm is as follows:
[0038] S i =(1-a)×S i +a×S local (7)
[0039] Bounding boxes are removed using the following algorithm:
[0040] score final =score NMS ×Ⅱ valid (8)
[0041] The present invention has the following beneficial effects:
[0042] This computer vision object positioning system, by introducing a lightweight backbone network and BEM module, effectively reduces the computing and storage requirements while maintaining performance, and is adapted to the UAV environment. Secondly, through the reverse attention exploration network, especially the challenge region predictor (CRP), the encoding, feature extraction, decoding and other operations are carefully designed. Through the spatial information provided by the reverse mask, the ability to distinguish difficult objects from the background is better learned, and the detection accuracy of challenging targets is improved. The uniqueness lies in the use of a multi-layer perceptron (MLP) network to learn the coordinates of the challenge area, further extract and learn the key features of the challenge area, and provide key support for the model to more accurately locate difficult objects in local detection. In addition, the challenging area is enlarged and the interpolation algorithm is applied, and in the subsequent feature extraction step, the higher resolution image details are fully considered, which improves the meticulous and comprehensive feature extraction of challenging targets, helping the model to perceive and identify challenging targets more accurately. Finally, the non-maximum suppression (NMS) is fine-tuned, and the weight distribution mechanism and bounding box processing are introduced to effectively fuse the outputs of the global and local detectors, improving the accuracy and robustness of target detection. These combined advantages make this method perform better than traditional methods in target detection in complex scenes. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 It is the algorithm flow chart of the present invention;
[0044] Figure 2 This is a network structure diagram of the branch enhancement module of the present invention;
[0045] Figure 3 Reverse attention exploration network diagram for the present invention;
[0046] Figure 4 This is the network structure diagram of the challenge area predictor of the present invention. DETAILED DESCRIPTION
[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0048] See also Figure 1-4 , a computer vision object positioning system, comprising the following steps:
[0049] Step S1: Use the UAV aerial photography Visdrone2019 dataset for training and testing to ensure the adequacy and diversity of the dataset;
[0050] The Visdrone2019 dataset of drone aerial photography provides a rich variety of perspectives and scenes, which is a favorable resource for target detection tasks. When dividing the dataset, pay attention to maintaining a balance between the training set and the test set to ensure the generalization performance of the model in different scenarios;
[0051] Step S2: In order to adapt to the lightweight backbone network, the resolution of the dataset images is adjusted. The resolution of all dataset images is adjusted to 640×640 as the input of the lightweight backbone network of the global detection part. The images are downsampled through conv_1 and conv_2, and then a lightweight feature extraction network composed of a branch enhancement module (BEM) and conv is used for feature extraction and image downsampling;
[0052] Step S3: The global features obtained through the above steps are used for global detection and as input to the reverse attention exploration network to generate the coordinates of the challenge area to assist subsequent local detection;
[0053] Step S4: Enlarge the target image according to the challenge area coordinates output in the previous step, such as Figure 3 In the process of magnifying the coordinates of the challenging target, the graphics are restored with high quality to ensure accurate positioning and detailed presentation of the challenging target.
[0054] Step S5: Feature extraction of the enlarged challenging area image. Feature extraction is performed on the enlarged challenging area image. After the global detector and the local detector output the prediction results respectively, a subsequent processing step, namely non-maximum suppression (NMS), is performed to eliminate redundant prediction bounding boxes.
[0055] In particular, step S2 includes introducing a BEM module into the lightweight feature extraction network to further improve the feature representation capability. Figure 2 As shown in the figure, a 16-way branch 1×1 convolution is first used to integrate the input feature information, and then a 3×3 convolution is used for feature extraction, where the number of these low-dimensional feature channels is 1 / 32 of the input feature map. The low-dimensional feature map branches are then merged and concatenated using 1×1 convolution features, and finally BN normalization is performed to prevent gradient loss.
[0056] The network model effectively captures different scales and contextual information in the image. This design aims to ensure that the model can more accurately describe complex scenes while maintaining high efficiency. The BEM module algorithm is as follows:
[0057]
[0058] Where x represents the input feature map, y represents the output feature map, conv1 represents 1×1 convolution, the channel is 1 / 2 of the input feature map channel, Γi(x) represents the input feature map downsampling to a low-dimensional feature map, and all Γi(x) have the same structure, including 1×1 convolution and 3×3 convolution, the number of output feature map channels is 1 / 32 of the number of input feature map channels, C represents the number of branches, and the network is set to 16, conv2 represents the 1×1 convolution of the combined feature, and BN represents the BatchNorm normalization operation.
[0059] Among them, step S3 specifically includes, firstly, using the challenge region predictor (CRP) to estimate the probability of the difficult objects to detect appearing in the picture under the guidance of the reverse attention mechanism. The proposed challenge region predictor (CRP) consists of an encoder, two feature extraction blocks, a fusion block and two decoders. The reverse attention exploration network structure is as follows Figure 3 As shown;
[0060] like Figure 4 As shown in the figure, the input global features are encoded, extracted, and decoded to generate a simple object distribution attention mask. The encoder first reduces the dimension of the global features to reduce the amount of calculation. The encoder consists of a 1×1 and a 3×3 convolution, BN normalization, and Sigmoid activation layer. The algorithm is as follows:
[0061] E encoded =Sigmoid(BN(Conv(Conv(G global ,1 1),3 3))) (2)
[0062] Among them G global is a global feature, BN represents the normalization operation, and Sigmoid represents the Sigmoid activation function.
[0063] The feature extraction block consists of two 3×3 convolutions, BN normalization, and Sigmoid activation layers. These two convolution layers help capture the nonlinear relationship between features. The subsequent BN normalization operation helps stabilize the training process and prevent the gradient from disappearing or exploding. Each convolution layer is followed by a Sigmoid activation layer to introduce nonlinearity, so that the network can better adapt to challenging scenarios. The algorithm is shown below.
[0064] F extracted =Sigmoid(BN(Conv(Conv(E encoded ,3 3),3 3))) (3)
[0065] The decoder consists of a 3×3 and a 1×1 convolution, BN normalization, and Sigmoid activation layer, followed by an inverted attention mask to generate a reverse mask containing the challenge area, which excludes the occupied simple object area. The algorithm is shown below.
[0066] M reverse =1-Sigmoid(BN(Conv(Conv(F extracted ,3 3),1 1))) (4)
[0067] Then, the spatial information provided by the reverse mask is input into the fusion block. Under the guidance of the reverse mask, the feature extraction network learns to distinguish difficult objects from the background and locate challenging areas. The challenging region predictor (CRP) can better select complementary areas and detail areas to obtain the possible location of the challenging area.
[0068] Among them, the specific steps of the region predictor (CRP) to distinguish difficult objects and locate them include pixel-wise multiplication of the inverted attention mask with the features obtained by extracting features in the previous step and adding them to obtain the challenge area features, and then extracting features and decoding operations to obtain the challenge area mask. The mask indicates the area where the difficult object is located through the decoder.
[0069] The feature extraction operation consists of two 3×3 convolutions, BN normalization, and Sigmoid activation layers. The challenge area features after feature extraction in the decoding operation are decoded through a 3×3 convolution layer and a 1×1 convolution layer, and then undergo BN normalization and Sigmoid activation again.
[0070] This stage aims to map the learned high-level features into a mask of the challenge area. This mask model can accurately indicate the area where the difficult object is located. The algorithm is as follows:
[0071]
[0072]
[0073] After obtaining the rough position of the difficult target, it is expanded into a vector, and nonlinear transformation and feature extraction are performed through multiple hidden layers to map the input data to the output result. In the forward propagation process, the input data is processed by multiple hidden layers and finally generates the prediction result;
[0074] The training process of MLP is an iterative optimization process. By continuously updating parameters, the model can learn and adapt to the complex patterns and relationships of the input data.
[0075] MLP is a neural network with multiple fully connected layers for learning nonlinear mapping. In this application, MLP is used to further extract and learn the key features of the challenge region. Its learning process plays a key role in further refining and learning the challenge region information based on the Challenge Region Predictor (CRP).
[0076] The input of MLP is the challenge region features obtained by the reverse mask. The features are first processed through the feature extraction operation of two 3×3 convolutions, BN normalization, and Sigmoid activation layers to capture higher-level representations. Then, the mask of the challenge region is generated through a decoding operation of a 3×3 and a 1×1 convolution, BN normalization, and Sigmoid activation layer. This mask indicates the specific area where the difficult object is located through the decoder. The entire MLP process maps the challenge region features to the final coordinate output through step-by-step nonlinear transformations. In this process, the MLP network provides the model with the ability to more accurately locate difficult objects in local detection by learning the spatial distribution of challenging targets, and outputs their coordinates.
[0077] Wherein, step S4 specifically includes, first, obtaining coordinate information of the challenge area, obtained from the above steps, the position of the challenging target indicated by the coordinates in the original high-resolution image;
[0078] Use bilinear interpolation to enlarge the image. For each pixel in the challenge area, according to the position of its coordinates in the original image, perform interpolation pixel by pixel in the entire challenge area, and complete the overall interpolation by linearly weighted averaging the values of the four surrounding pixels to obtain the interpolated pixel value;
[0079] Finally, the interpolated image is integrated into the original high-resolution image to form an enlarged image, which can show the details of the challenging target more clearly and retain the characteristics of the original image.
[0080] Specifically, step S5 includes, firstly, taking the magnified challenging area image in step S4 as input, constructing a feature extraction network, and gradually improving the abstract level of features through convolution operation;
[0081] In the feature extraction network, the BEM module is introduced to further improve the feature expression ability of challenging areas through the fusion and splicing of multi-way convolution branches and low-dimensional feature maps.
[0082] The activation function and normalization operation are applied to each convolutional layer to enhance the nonlinear expression ability of the network and prevent the gradient vanishing problem. In addition, the feature map is downsampled through the pooling layer to reduce the amount of calculation while retaining key feature information.
[0083] Finally, an output feature map with a higher level representation is obtained.
[0084] Finally, after the global detector and the local detector each output their predictions, a subsequent processing step, Non-Maximum Suppression (NMS), is performed to eliminate redundant predicted bounding boxes.
[0085] It also includes a weight allocation mechanism that gives higher weight to the results output by the local detector. The algorithm is as follows:
[0086] S i =(1-a)×S i +a×S local (7)
[0087] Where S i is the raw confidence score of the i-th bounding box, S local is the confidence score of the corresponding local detector, and α is a weight adjustment parameter, which usually takes a value between 0 and 1. Through this mechanism, the influence of the local detector in the result can be more flexibly controlled, so that it can contribute more significantly to the final result when providing useful contextual information.
[0088] In addition, bounding boxes that are close to the edge of the challenging area may be truncated due to cropping, resulting in loss of target information. To solve this problem, bounding boxes are removed using the following algorithm:
[0089] score final =score NMS ×Ⅱ valid (8)
[0090] The score fina l is the final target detection score, score NMS is the score after NMS processing, Ⅱ valid is an indicator function that takes the value of 1 when the bounding box is valid and 0 otherwise. The implementation of this step ensures the accuracy and completeness of the final target detection result.
[0091] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.
[0092] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A computer vision object positioning system, characterized in that: The following steps are involved: Step S1: Use the UAV aerial photography Visdrone2019 dataset for training and testing to ensure the adequacy and diversity of the dataset; Step S2: The resolution of the dataset images is adjusted to adapt to the lightweight backbone network. The resolution of all dataset images is adjusted to 640×640 as the input of the lightweight backbone network of the global detection part. The images are downsampled through conv_1 and conv_2, and a lightweight feature extraction network composed of branch enhancement module (BEM) and conv is used for feature extraction and image downsampling. Step S3: Apply the global features obtained in the above steps to global detection and use them as input to the reverse attention exploration network to generate the coordinates of the challenge area to assist subsequent local detection; Step S4: amplifying and restoring the target image with high quality according to the generated challenge area coordinates; Step S5: Feature extraction is performed on the enlarged challenging area image. After the global detector and the local detector output prediction results respectively, non-maximum suppression processing is performed to eliminate redundant prediction bounding boxes.
2. A computer vision object positioning system according to claim 1, characterized in that: Step S2 specifically includes: introducing a BEM module in a lightweight feature extraction network, integrating the input feature information using a 16-way branch 1×1 convolution, and then extracting features using a 3×3 convolution, where the number of these low-dimensional feature channels is 1 / 32 of the input feature map, merging the low-dimensional feature map branches, and concatenating the features using a 1×1 convolution, and finally performing BN normalization to prevent gradient loss. The BEM module algorithm is as follows: Where x represents the input feature map, y represents the output feature map, conv1 represents 1×1 convolution, the channel is 1 / 2 of the input feature map channel, Γi(x) represents the input feature map downsampling to a low-dimensional feature map, and all Γi(x) have the same structure, including 1×1 convolution and 3×3 convolution, the number of output feature map channels is 1 / 32 of the number of input feature map channels, C represents the number of branches, and the network is set to 16, conv2 represents the 1×1 convolution of the combined feature, and BN represents the BatchNorm normalization operation.
3. A computer vision object positioning system according to claim 1, characterized in that: Step S3 specifically includes, using a challenging region predictor (CRP) to estimate the probability of difficult-to-detect objects appearing in the image under the guidance of the reverse attention mechanism; The Challenge Region Predictor (CRP) consists of an encoder, two feature extraction blocks, a fusion block, and two decoders; The input global features are encoded, feature extracted, and decoded to generate a simple object distribution attention mask. The encoder first reduces the dimension of the global features to reduce the amount of calculation; The algorithm is as follows: E encoded =Sigmoid(BN(Conv(Conv(G global ,11),33))) (2) The feature extraction block consists of two 3×3 convolutions, BN normalization, and Sigmoid activation layers. Each convolution layer is followed by a Sigmoid activation layer to introduce nonlinearity so that the network can better adapt to challenging scenarios. The algorithm is as follows: F extracted =Sigmoid(BN(Conv(Conv(E encoded ,33),33))) (3) The decoder consists of a 3×3 and a 1×1 convolution, BN normalization, and Sigma id activation layer, followed by an inverted attention mask to generate a reverse mask containing the challenge area, which excludes the occupied simple object area. The algorithm is as follows: M reverse =1-Sigmoid(BN(Conv(Conv(F extracted ,33),11))) (4) 4. A computer vision object positioning system according to claim 3, characterized in that: The specific steps of the region predictor (CRP) to distinguish and locate difficult objects include pixel-wise multiplication of the inverted attention mask and the features obtained by the previous step of feature extraction and adding them to obtain the challenge region feature, and then extracting features and decoding operations to obtain the challenge region mask, which indicates the region where the difficult object is located through the decoder; The feature extraction operation consists of two 3×3 convolutions, BN normalization, and Sigmoid activation layers. The decoding operation decodes the challenge area features after feature extraction through a 3×3 convolution layer and a 1×1 convolution layer, and then undergoes BN normalization and Sigmoid activation again. Map high-level features to masks of the challenge area, and use the mask model to accurately indicate the area where the difficult object is located. The algorithm is as follows: F extracted2 =Sigmoid(BN(Conv(Conv((F extracted M reverse F extracted ),33)),33)) (5) F challenge =Sigmoid(BN(Conv(Conv(F extracted2 ,33),11))) (6) After obtaining the rough position of the difficult target, it is expanded into a vector, and nonlinear transformation and feature extraction are performed through multiple hidden layers to map the input data to the output result. During the forward propagation process, the input data is processed by multiple hidden layers to finally generate the prediction result.
5. A computer vision object positioning system according to claim 1, characterized in that: Step S4 specifically includes, first, obtaining coordinate information of the challenge area, obtained from the above steps, the position of the challenging target indicated by the coordinates in the original high-resolution image; Use bilinear interpolation to enlarge the image. For each pixel in the challenge area, according to the position of its coordinates in the original image, perform interpolation pixel by pixel in the entire challenge area, and complete the overall interpolation by linear weighted averaging the values of the four surrounding pixels to obtain the interpolated pixel value. Finally, the interpolated image is integrated into the original high-resolution image to form an enlarged image.
6. A computer vision object positioning system according to claim 1, characterized in that: Step S5 specifically includes, firstly, taking the magnified challenging area image in step S4 as input, constructing a feature extraction network, and gradually improving the abstract level of features through convolution operations; In the feature extraction network, the BEM module is introduced through the fusion and concatenation of multi-way convolution branches and low-dimensional feature maps; The activation function and normalization operation are applied to each convolution layer, and the feature map is downsampled through the pooling layer; the output feature map with a higher level representation is obtained. Finally, after the global detector and the local detector each output their predictions, a subsequent processing step, Non-Maximum Suppression (NMS), is performed to eliminate redundant predicted bounding boxes.
7. A computer vision object positioning system according to claim 6, characterized in that: It also includes a weight distribution mechanism that gives higher weight to the results output by the local detector. The algorithm is as follows: S i =(1-a)×S i +a×S local (7) Bounding boxes are removed using the following algorithm: score final =score NMS ×Ⅱ valid (8)。