Target detection method and device combining computer vision and big data mining, and computer equipment

Abstract

The application relates to the technical field of target detection, in particular to a target detection method and device combining computer vision and big data mining and a computer device. Based on a preliminary detection result, relevant enhancement and inhibition templates are searched in a pre-constructed context enhancement template library and a context inhibition template library, a spatial attention graph is generated through similarity matching calculation, fusion enhancement feature maps and fusion inhibition feature maps are obtained through fusion operation, and a general scene enhancement graph generated in combination with scene features is used to finely bidirectionally modulate a target feature extraction graph extracted based on a to-be-detected image, features conforming to the context of the target can be adaptively strengthened, and interference signals violating logical relations can be weakened, so that the detection capability for occluded targets, small targets and rare targets can be effectively improved in a complex scene, false detection and missed detection are reduced, and the accuracy and efficiency of target detection are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target detection, in particular to a target detection method and device combining computer vision and big data mining and a computer device. BACKGROUND

[0002] In the current field of computer vision, although the target detection technology has made significant progress, especially the deep learning-based methods such as Faster R-CNN, YOLO and SSD, which can directly predict the target bounding box and class label from the image through end-to-end training, these methods mainly rely on the analysis of the visual features within a single image, and have inherent limitations. Firstly, they lack the use of the semantic co-occurrence rules and spatial context relationships between targets in large-scale data, for example, the frequent co-occurrence of "keyboard" and "mouse" or the mutual exclusion relationship of "ship" and "desert", which are difficult to be parameterized and learned by a single model; secondly, the robustness of traditional methods to complex scenes such as occlusion, blur or small targets is insufficient, because they do not use the visual patterns learned from historical data about specific targets in typical contexts, such as the common appearance of monitors in office table scenes, to enhance feature expression and suppress unreasonable detection; thirdly, the existing technology cannot dynamically guide the model to focus on the context area of the primary target detected, such as "computer desk", such as the "monitor" that may exist in the area above the desktop, or suppress false detection that violates spatial common sense, such as "car" appearing in "sky", which limits the detection accuracy in complex real scenes.

[0003] Therefore, it is necessary to provide a target detection system and method combining computer vision and big data mining to solve the above technical problems. SUMMARY

[0004] Based on this, the purpose of the present application is to provide a target detection method and device combining computer vision and big data mining, a computer device and a storage medium, based on the preliminary detection result, the relevant enhancement and suppression templates are retrieved from the pre-constructed context enhancement template library and context suppression template library, and the spatial attention map is generated through similarity matching calculation, the fusion enhancement feature map and the fusion suppression feature map are obtained through fusion operation, and the general scene enhancement map generated by combining the scene features, the target feature extraction map extracted based on the image to be detected is finely modulated in both directions, which can adaptively strengthen the feature expression consistent with the target context and weaken the interference signals that violate the logical relationship, thereby effectively improving the detection ability of occluded targets, small targets and rare targets in complex scenes, reducing false detection and missing detection, and improving the accuracy and efficiency of target detection.

[0005] In a first aspect, the embodiments of the present application provide a target detection method combining computer vision and big data mining, comprising the following steps:

[0006] obtaining a to-be-detected image; performing feature extraction on the to-be-detected image to obtain a target feature extraction map and a general scene enhancement map of the to-be-detected image;

[0007] performing target detection frame extraction based on the target feature extraction map to obtain a plurality of initial target detection frames of the to-be-detected image and a category label corresponding to each initial target detection frame;

[0008] According to each initial target detection frame and the category label corresponding to each initial target detection frame, a corresponding template feature vector is queried from a pre-constructed context enhancement template library and a context suppression template library, respectively, to obtain an enhanced template feature vector and a suppression template feature vector corresponding to each initial target detection frame;

[0009] Performing similarity matching calculation on the target feature extraction map and the enhanced template feature vector and the suppression template feature vector corresponding to each initial target detection frame, respectively, to obtain an enhanced context attention map and a suppression context attention map corresponding to each initial target detection frame;

[0010] Fusing the enhanced context attention map corresponding to each initial target detection frame to obtain a fused enhanced feature map; and fusing the suppression context attention map corresponding to each initial target detection frame to obtain a fused suppression feature map;

[0011] According to the fused enhanced feature map, the fused suppression feature map, and the general scene enhancement map, performing bidirectional modulation on the target feature extraction map to obtain a target modulation feature map; and performing target recognition processing on the target modulation feature map to obtain a target detection result of the to-be-detected image.

[0012] In a second aspect, the embodiments of the present application provide a target detection device combining computer vision and big data mining, comprising:

[0013] A feature extraction module is configured to obtain a to-be-detected image; perform feature extraction on the to-be-detected image to obtain a target feature extraction map and a general scene enhancement map of the to-be-detected image;

[0014] A detection frame extraction module is configured to perform target detection frame extraction based on the target feature extraction map to obtain a plurality of initial target detection frames of the to-be-detected image and a category label corresponding to each initial target detection frame;

[0015] The template feature query module is configured to perform corresponding template feature vector query from a pre-constructed context enhancement template library and a context suppression template library according to each initial target detection frame and a category label corresponding to each initial target detection frame, to obtain an enhancement template feature vector and a suppression template feature vector corresponding to each initial target detection frame.

[0016] The context attention extraction module is configured to perform similarity matching calculation on the target feature extraction map and the enhancement template feature vector and the suppression template feature vector corresponding to each initial target detection frame, to obtain an enhancement context attention map and a suppression context attention map corresponding to each initial target detection frame.

[0017] The feature fusion module is configured to fuse the enhancement context attention map corresponding to each initial target detection frame to obtain a fused enhancement feature map, and fuse the suppression context attention map corresponding to each initial target detection frame to obtain a fused suppression feature map.

[0018] The target detection module is configured to perform bidirectional modulation on the target feature extraction map according to the fused enhancement feature map, the fused suppression feature map and a general scene enhancement map, to obtain a target modulation feature map, and perform target recognition processing on the target modulation feature map to obtain a target detection result of the image to be detected.

[0019] In a third aspect, an embodiment of the present application provides a computer device, including a processor, a memory, and a computer program stored in the memory and executable on the processor; when the computer program is executed by the processor, the steps of the target detection method combining computer vision and big data mining according to the first aspect are implemented.

[0020] In a fourth aspect, an embodiment of the present application provides a storage medium, which stores a computer program; when the computer program is executed by a processor, the steps of the target detection method combining computer vision and big data mining according to the first aspect are implemented.

[0021] In the embodiment of the present application, a target detection method, device, computer equipment and storage medium combining computer vision and big data mining are provided. Based on the preliminary detection result, the relevant enhancement and inhibition templates are retrieved from the pre-constructed context enhancement template library and context inhibition template library, and the spatial attention map is generated through similarity matching calculation. The fusion enhancement feature map and the fusion inhibition feature map are obtained through fusion operation, and the general scene enhancement map is generated in combination with the scene features. The target feature extraction map extracted based on the to-be-detected image is finely bidirectionally modulated, which can adaptively strengthen the feature expression consistent with the target context and weaken the interference signals violating the logical relationship, thereby effectively improving the detection capability of the occluded target, small target and rare target in a complex scene, reducing the false detection and missing detection, and improving the accuracy and efficiency of target detection.

[0022] In order to better understand and implement, the present application is described in detail below with reference to the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 The flowchart of the target detection method combining computer vision and big data mining provided by an embodiment of the present application is shown in the figure;

[0024] Figure 2 The flowchart of S1 in the target detection method combining computer vision and big data mining provided by an embodiment of the present application is shown in the figure;

[0025] Figure 3 The flowchart of S7 in the target detection method combining computer vision and big data mining provided by another embodiment of the present application is shown in the figure;

[0026] Figure 4 The flowchart of S73 in the target detection method combining computer vision and big data mining provided by an embodiment of the present application is shown in the figure;

[0027] Figure 5 The flowchart of S73 in the target detection method combining computer vision and big data mining provided by another embodiment of the present application is shown in the figure;

[0028] Figure 6 The flowchart of S4 in the target detection method combining computer vision and big data mining provided by an embodiment of the present application is shown in the figure;

[0029] Figure 7 The flowchart of S5 in the target detection method combining computer vision and big data mining provided by an embodiment of the present application is shown in the figure;

[0030] Figure 8 The flowchart of S6 in the target detection method combining computer vision and big data mining provided by an embodiment of the present application is shown in the figure;

[0031] Figure 9 A structural schematic diagram of a target detection device combining computer vision and big data mining is provided for an embodiment of the present application.

[0032] Figure 10 A structural schematic diagram of a computer device is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0033] The exemplary embodiments will be described in detail herein with reference to the attached drawings. The description of the exemplary embodiments is intended to apply to various alternative embodiments as well. The following description is not limited to the exemplary embodiments but applies equally well to other embodiments as long as they are within the scope of the appended claims.

[0034] The terminology used in the present application is for the purpose of describing particular embodiments only and is not intended to be limiting of the present application. As used in the present application and the following claims, the singular forms “a,” “an” and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms “comprises” and / or “comprising,” when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0035] It is to be understood that the singular forms “a,” “an,” and “the” include plural referents unless the context clearly dictates otherwise. It will be further understood that the terms “comprises” and / or “comprising,” when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0036] Please refer to Figure 1 , Figure 1 A flowchart of a target detection method combining computer vision and big data mining is provided for an embodiment of the present application, and the method comprises the following steps:

[0037] S1: obtaining a to-be-detected image; performing feature extraction on the to-be-detected image to obtain a target feature extraction image and a general scene enhancement image of the to-be-detected image.

[0038] The execution subject of the target detection method combining computer vision and big data mining in the present application is a detection device (hereinafter referred to as a detection device) of the target detection method combining computer vision and big data mining. In an optional embodiment, the detection device can be a computer device, can be a server, or a server cluster formed by multiple computer devices.

[0039] In an optional embodiment, the detection device can obtain a user-inputted image to be detected, or can extract an image to be detected from a preset database.

[0040] The detection device performs feature extraction on the image to be detected, and obtains a target feature extraction map and a general scene enhancement map of the image to be detected.

[0041] Please refer to Figure 2 , Figure 2 The flowchart of S1 in the target detection method combining computer vision and big data mining provided by an embodiment of the present application includes steps S11-S12, and the details are as follows.

[0042] S11: inputting the image to be detected into a preset convolutional neural network for multi-level feature extraction, obtaining an intermediate-level feature extraction map and a target-level feature extraction map, and taking the intermediate-level feature extraction map as a target feature extraction map.

[0043] In the present embodiment, the detection device inputs the image to be detected into a preset convolutional neural network for multi-level feature extraction, and obtains an intermediate-level feature extraction map and a target-level feature extraction map, wherein the target-level feature extraction map is a feature extraction map of a deeper level than the intermediate-level feature extraction map.

[0044] The detection device takes the intermediate-level feature extraction map as a target feature extraction map, wherein the target feature extraction map is a high-resolution feature map, and the spatial resolution of the high-resolution feature map is not lower than one-eighth of the original image resolution of the image to be detected.

[0045] Specifically, the convolutional neural network adopts ResNet50 and VGG16, the backbone network of the convolutional neural network is responsible for extracting multi-level feature representations, from the intermediate level of the backbone network, for example, from the output of the third residual stage, a target feature extraction map is extracted, the spatial resolution of the target feature extraction map is ensured to be not lower than one-eighth of the original input image resolution by controlling the down-sampling rate, so as to reserve sufficient spatial detail information for accurate positioning, and at the same time, from a deeper level of the same backbone network, for example, from the output after the fourth residual stage, a target-level feature extraction map is extracted.

[0046] S12: performing a global average pooling operation on the feature extraction graph of the target level to obtain a global average pooling feature graph as a global scene feature vector; inputting the global scene feature vector into a preset deconvolution network to perform upsampling and nonlinear mapping, and obtaining a general scene enhancement graph.

[0047] In this embodiment, the detection device performs a global average pooling operation on the feature extraction graph of the target level to obtain a global average pooling feature graph as a global scene feature vector, wherein the global scene feature vector carries macro scene semantic information of the image.

[0048] The detection device inputs the global scene feature vector into a preset deconvolution network to perform upsampling and nonlinear mapping, and obtains a general scene enhancement graph, wherein the general scene enhancement graph is used to adaptively enhance the feature response of the relevant semantic region according to the overall scene type of the image in the subsequent modulation process.

[0049] Specifically, the deconvolution network is sequentially connected by at least one fully connected layer and two deconvolution layers with a step of 2, the detection device inputs the global scene feature vector into the deconvolution network, the fully connected layer first raises the dimension of the global scene feature vector to a size suitable for deconvolution operation, and then the deconvolution layer gradually performs upsampling and uses a ReLU activation function after each deconvolution layer to introduce nonlinear mapping, and finally outputs a single-channel weight graph with the same spatial size as the feature extraction graph of the target level as the general scene enhancement graph.

[0050] S2: performing target detection frame extraction based on the target feature extraction graph to obtain a plurality of initial target detection frames of the image to be detected and a class label corresponding to each initial target detection frame.

[0051] In this embodiment, the detection device performs target detection frame extraction based on the target feature extraction graph to obtain a plurality of initial target detection frames of the image to be detected and a class label corresponding to each initial target detection frame.

[0052] Specifically, the detection device uses a region proposal network to generate a plurality of candidate regions for the target feature extraction graph, and outputs an initial target detection frame and a class label probability distribution corresponding to each candidate region through classification and regression operations to obtain a plurality of initial target detection frames of the image to be detected and a class label corresponding to each initial target detection frame, wherein the class label is determined by a softmax classifier.

[0053] S3: According to each of the initial target detection frame and the category label corresponding to each of the initial target detection frame, a corresponding template feature vector query is performed from a pre-constructed context enhancement template library and a context suppression template library respectively, to obtain an enhancement template feature vector and a suppression template feature vector corresponding to each initial target detection frame.

[0054] In the embodiment, the detection device performs a corresponding template feature vector query from a pre-constructed context enhancement template library and a context suppression template library respectively according to each of the initial target detection frame and the category label corresponding to each of the initial target detection frame, to obtain an enhancement template feature vector and a suppression template feature vector corresponding to each initial target detection frame.

[0055] In an optional embodiment, it further includes a step S7 of constructing a context enhancement template library and a context suppression template library; the step S7 is before the step S3, please refer to Figure 3 , Figure 3 The flowchart of step S7 in the target detection method combining computer vision and big data mining provided for another embodiment of the application is shown in FIG. 7, which includes steps S71-S73, and specifically as follows:

[0056] S71: Using big data mining, all target category pairs in a preset image annotation data set are determined to construct a target category pair set.

[0057] In the embodiment, the detection device uses big data mining to determine all target category pairs in a preset image annotation data set to construct a target category pair set, wherein the target category pair set includes a plurality of target category pairs; the target category pair includes a dominant target and an associated target, and each of the target category pairs is associated with a plurality of initial images in the image annotation data set that contain both the dominant target and the associated target.

[0058] Specifically, the detection device processes the image annotation data set through big data mining technology, and the detection device traverses all labeled images in the image annotation data set, generates a set containing all target category labels appearing in each image for each image, thereby converting the entire data set into a category appearance record library that can be used for frequency statistics, and enumerating all possible two-by-two target category combinations in the set of target category labels to form an initial target category pair set.

[0059] S72: Obtain the co-occurrence frequency of each target category pair in the target category pair set; according to the co-occurrence frequency of each of the target category pairs and a preset frequency threshold, the target category pairs with a co-occurrence frequency higher than the frequency threshold are taken as high-frequency co-occurrence pairs, and a plurality of high-frequency co-occurrence pairs of the image annotation data set are obtained.

[0060] In this embodiment, the detection device obtains the co-occurrence frequency of each target class pair in the set of target class pairs, specifically, the detection device calculates the co-occurrence frequency based on the number of images containing both target classes in the class pair in the set of target class pairs, and calculates the ratio of the number to the total number of images in the image annotation dataset, thereby obtaining a normalized co-occurrence frequency measure.

[0061] The detection device obtains several high-frequency co-occurrence pairs of the image annotation dataset according to the co-occurrence frequency of each target class pair and a preset frequency threshold, and the target class pair with a co-occurrence frequency higher than the frequency threshold is a high-frequency co-occurrence pair.

[0062] S73: Constructing a template library according to the image annotation dataset and the several high-frequency co-occurrence pairs of the image annotation dataset, obtaining the context-enhanced template library and the context-suppressed template library.

[0063] In this embodiment, the detection device constructs a template library according to the image annotation dataset and the several high-frequency co-occurrence pairs of the image annotation dataset, and obtains the context-enhanced template library and the context-suppressed template library.

[0064] Please refer to Figure 4 , Figure 4 The flowchart of S73 in the target detection method combining computer vision and big data mining provided by an embodiment of the present application is shown in FIG. 7, which includes steps S7301-S7302, and the details are as follows.

[0065] S7301: According to the target position annotation in the image annotation dataset, performing associated target region recognition and associated target region visual feature extraction on each image associated with the high-frequency co-occurrence pair, and obtaining the visual feature vectors of the associated target regions of each initial image associated with each high-frequency co-occurrence pair.

[0066] In this embodiment, the detection device performs associated target region recognition and associated target region visual feature extraction on each image associated with the high-frequency co-occurrence pair according to the target position annotation in the image annotation dataset, and obtains the visual feature vectors of the associated target regions of each initial image associated with each high-frequency co-occurrence pair.

[0067] Specifically, for each high-frequency co-occurrence pair, the detection device needs to traverse all image instances associated with the high-frequency co-occurrence pair in the image annotation dataset, crop the image region where the associated target is located according to the accurate target position annotation, and perform forward propagation on the region using a pre-trained feature extraction network to extract the high-dimensional visual feature vector thereof.

[0068] S7302: aggregate the visual feature vectors of the associated target regions of the same initial image to obtain an enhanced visual template of each initial image of each high-frequency co-occurrence pair, and construct the context-enhanced template library.

[0069] In this embodiment, the detection device aggregates the visual feature vectors of the associated target regions of the same initial image to obtain an enhanced visual template of each initial image of each high-frequency co-occurrence pair, and constructs the context-enhanced template library.

[0070] Specifically, the detection device aggregates the visual feature vectors of all image instances extracted under the same high-frequency co-occurrence pair through an average pooling operation to generate a visual template representing the typical visual features of the associated targets in the high-frequency co-occurrence pair, and finally stores all high-frequency co-occurrence pairs and their corresponding visual templates to form the context-enhanced template library.

[0071] Please refer to Figure 5 , Figure 5 The flowchart of S73 of the target detection method combining computer vision and big data mining provided for another embodiment of the present application is shown in FIG. 7B, which includes steps S7311-S7314, and specifically as follows.

[0072] S7311: according to the target position annotation in the image annotation dataset, taking the annotation box of the dominant target as the reference system, calculating the spatial relationship vector between the associated target and the dominant target in each initial image associated with each high-frequency co-occurrence pair to obtain the spatial relationship vector of each initial image associated with each high-frequency co-occurrence pair; and according to the spatial relationship vector of each initial image associated with each high-frequency co-occurrence pair and a preset Gaussian mixture model, performing probability density estimation to construct a spatial distribution probability model of each high-frequency co-occurrence pair.

[0073] In this embodiment, the detection device calculates the spatial relationship vector between the associated target and the dominant target in each initial image associated with each high-frequency co-occurrence pair according to the target position annotation in the image annotation dataset, taking the annotation box of the dominant target as the reference system.

[0074] The detection device performs probability density estimation according to the spatial relationship vector of each initial image associated with each high-frequency co-occurrence pair and a preset Gaussian mixture model to construct a spatial distribution probability model of each high-frequency co-occurrence pair.

[0075] S7312: Calculate the likelihood of the spatial relationship vector and the spatial distribution probability model of each initial image associated with the same high-frequency co-occurrence pair to obtain the likelihood of each initial image associated with each high-frequency co-occurrence pair. According to the likelihood of each initial image associated with each high-frequency co-occurrence pair and the preset likelihood threshold, the image with a likelihood lower than the preset likelihood threshold is determined as a spatial anomaly image, and a plurality of spatial anomaly images of each high-frequency co-occurrence pair are obtained.

[0076] In this embodiment, the detection device calculates the likelihood of the spatial relationship vector and the spatial distribution probability model of each initial image associated with the same high-frequency co-occurrence pair to obtain the likelihood of each initial image associated with each high-frequency co-occurrence pair.

[0077] The detection device calculates the likelihood of the spatial relationship vector and the spatial distribution probability model of each initial image associated with the same high-frequency co-occurrence pair to obtain the likelihood of each initial image associated with each high-frequency co-occurrence pair.

[0078] S7313: According to each spatial anomaly image of each high-frequency co-occurrence pair and the target position annotation in the image annotation data set, the associated target region recognition and the visual feature extraction of the associated target region are performed on the spatial anomaly image to obtain a plurality of visual feature vectors of the associated target region of each spatial anomaly image of each high-frequency co-occurrence pair.

[0079] In this embodiment, the detection device calculates the likelihood of the spatial relationship vector and the spatial distribution probability model of each initial image associated with the same high-frequency co-occurrence pair to obtain the likelihood of each initial image associated with each high-frequency co-occurrence pair.

[0080] S7314: Aggregate the visual feature vectors of each associated target region of the same spatial anomaly image to obtain an abnormal visual template of each spatial anomaly image of each high-frequency co-occurrence pair, and construct the context suppression template library.

[0081] In this embodiment, the detection device calculates the likelihood of the spatial relationship vector and the spatial distribution probability model of each initial image associated with the same high-frequency co-occurrence pair to obtain the likelihood of each initial image associated with each high-frequency co-occurrence pair.

[0082] S4: Perform similarity matching calculations between the target feature extraction map and the enhanced template feature vector and suppressed template feature vector corresponding to each of the initial target detection boxes to obtain the enhanced context attention map and suppressed context attention map corresponding to each of the initial target detection boxes.

[0083] The enhanced contextual attention map highlights regions in the image that are similar to the enhanced template features, i.e., those regions that may contain associated targets and conform to normal contextual relationships.

[0084] The suppressed context attention map highlights regions in the image that are similar to the suppressed template features, i.e., those regions that may contain associated targets but are in an abnormal contextual relationship.

[0085] In this embodiment, the detection device performs similarity matching calculations between the target feature extraction map and the enhanced template feature vector and the suppressed template feature vector corresponding to each of the initial target detection boxes to obtain the enhanced context attention map and the suppressed context attention map corresponding to each of the initial target detection boxes.

[0086] Please see Figure 6 , Figure 6 The flowchart of step S4 in the object detection method combining computer vision and big data mining provided in one embodiment of this application includes steps S41 to S42, as follows:

[0087] S41: The enhanced template feature vector and the suppressed template feature vector are respectively used as input template feature vectors, and the input template feature vectors are reshaped by convolution kernels to obtain convolution kernel reshaped data.

[0088] In order to perform similarity matching calculation, in this embodiment, the detection device takes the enhanced template feature vector and the suppressed template feature vector as input template feature vectors respectively, and performs convolution kernel reshaping on the input template feature vectors to obtain convolution kernel reshaping data.

[0089] Specifically, the detection device adjusts each input template feature vector into a convolutional kernel of size 1x1xC, where C represents the number of channels in the high-resolution feature map, ensuring that the convolutional kernel is aligned with the high-resolution feature map in the channel dimension, thereby obtaining convolutional kernel reshaping data.

[0090] S42: Perform a convolution operation based on the reshaped data of the convolution kernel and the target feature extraction map. Calculate the dot product similarity between the convolution kernel and the feature vector corresponding to the spatial position at each spatial position of the target feature extraction map to obtain the single-channel response map corresponding to the input template feature vector, which serves as the corresponding enhanced context attention map and suppressed context attention map. Obtain the enhanced context attention map and suppressed context attention map corresponding to each of the initial target detection boxes.

[0091] In this embodiment, the detection device performs a convolution operation based on the reconstructed data from the convolution kernel and the target feature extraction map. At each spatial location of the target feature extraction map, the dot product similarity between the convolution kernel and the feature vector corresponding to the spatial location is calculated to obtain a single-channel response map corresponding to the input template feature vector, which serves as the corresponding enhanced context attention map and suppressed context attention map. This process yields the enhanced context attention map and suppressed context attention map corresponding to each initial target detection box.

[0092] Specifically, the detection device performs convolution operations on the reshaped data of the convolution kernel and the target feature extraction map respectively. The convolution operation calculates the dot product similarity between the convolution kernel and the local feature vector at each spatial location of the feature map in a sliding window manner, thereby generating a single-channel response map with the same spatial size as the target feature extraction map, which serves as the corresponding enhanced context attention map and suppressed context attention map.

[0093] S5: Fuse the enhanced context attention maps corresponding to each of the initial target detection boxes to obtain a fused enhanced feature map; fuse the suppressed context attention maps corresponding to each of the initial target detection boxes to obtain a fused suppressed feature map.

[0094] Based on the enhanced contextual attention map and the suppressed contextual attention map for each of the initial target detection boxes, these attention maps are consistent with the target feature extraction map in terms of spatial size, and the value of each pixel represents the degree of matching between the spatial location and the corresponding context template.

[0095] In order to fuse all the scattered context attention maps into a unified comprehensive enhancement map, in this embodiment, the detection device fuses the enhanced context attention maps corresponding to each of the initial target detection boxes to obtain a fused enhanced feature map; and fuses the suppressed context attention maps corresponding to each of the initial target detection boxes to obtain a fused suppressed feature map.

[0096] Please see Figure 7 , Figure 7The flowchart of step S5 in the object detection method combining computer vision and big data mining provided in one embodiment of this application includes steps S51 to S52, as follows:

[0097] S51: The enhanced context attention map and the suppressed context attention map are used as input context attention maps respectively; the maximum value operation method is used to traverse the pixel intensity value of each pixel position of each input context attention map to obtain the pixel intensity value of each pixel position of each input context attention map.

[0098] In this embodiment, the detection device uses the enhanced context attention map and the suppressed context attention map as input context attention maps, respectively.

[0099] The detection device uses a maximum value operation method to traverse the pixel intensity values ​​at each pixel position of each input context attention map to obtain the pixel intensity values ​​at each pixel position of each input context attention map.

[0100] S52: Take the maximum value of the pixel intensity values ​​of each input context attention map at the same pixel position to obtain the maximum pixel intensity value at each pixel position, construct the fusion feature map corresponding to the input context attention map, and use it as the corresponding fusion enhancement feature map and fusion suppression feature map to obtain the fusion enhancement feature map and fusion suppression feature map.

[0101] In this embodiment, the detection device takes the maximum pixel intensity value of each input context attention map at the same pixel location, constructs the fusion feature map corresponding to the input context attention map, and uses it as the corresponding fusion enhancement feature map and fusion suppression feature map. This operation can effectively retain the most significant feature response in the input context attention map, thereby highlighting the key areas in the image that are highly matched with any positive contextual clues and avoiding response dilution caused by averaging.

[0102] S6: Based on the fusion enhancement feature map, fusion suppression feature map, and general scene enhancement map, the target feature extraction map is bidirectionally modulated to obtain a target modulation feature map; the target modulation feature map is then subjected to target recognition processing to obtain the target detection result of the image to be detected.

[0103] In this embodiment, the detection device performs bidirectional modulation on the target feature extraction map based on the fusion enhancement feature map, the fusion suppression feature map, and the general scene enhancement map to obtain a target modulation feature map.

[0104] The detection device performs target recognition processing on the target modulation feature map to obtain the target detection result of the image to be detected.

[0105] Specifically, the detection device inputs the target modulation feature map into a pre-defined lightweight target detection head. This target detection head includes a convolutional layer, a classifier, and a bounding box regressor. Through this head, target category prediction and precise location regression are performed on the target modulation feature map, outputting a detection result containing category labels and bounding box coordinates, which serves as the target detection result for the image to be detected. The entire bidirectional modulation process balances enhancement and suppression effects, utilizing pre-computed contextual knowledge to optimize feature representation, thereby improving the accuracy and robustness of target detection.

[0106] Please see Figure 8 , Figure 8 The flowchart of step S6 in the object detection method combining computer vision and big data mining provided in one embodiment of this application includes steps S61 to S62, as follows:

[0107] S61: Multiply the fused enhanced feature map and the general scene enhanced map by preset enhancement modulation coefficients respectively, and then add them together to obtain a composite enhanced weight map; add the composite enhanced weight map to a unit value to obtain an enhanced modulation weight map; multiply the enhanced modulation weight map element-wise with the target feature extraction map to obtain an enhanced modulation feature map.

[0108] In this embodiment, the detection device multiplies the fused enhanced feature map and the general scene enhanced map by preset enhancement modulation coefficients and then adds them together to obtain a composite enhanced weight map. The enhancement modulation coefficients are adjustable parameters used to control the relative intensity of the enhancement. The composite enhanced weight map integrates enhanced signals from specific contexts and the global scene.

[0109] The detection device adds the composite enhancement weight map to the unit value to obtain the enhancement modulation weight map. Specifically, based on the composite enhancement weight map and the unit value in the preset unit matrix, the detection device adds element-wise a unit matrix with all elements equal to 1 and the same spatial size as the composite enhancement weight map to obtain the enhancement modulation weight map. This operation ensures that the enhancement weight baseline value is 1, thereby avoiding feature attenuation, and allows the weight of the enhancement region to be greater than 1 to amplify the feature.

[0110] The detection device performs element-wise multiplication of the enhanced modulation weight map and the target feature extraction map to obtain an enhanced modulation feature map. The high-value regions of the enhanced modulation feature map are significantly amplified, which improves the salience of the target region and the detection probability.

[0111] S62: Multiply the fused suppression feature map by a preset suppression modulation coefficient to obtain a basic suppression weight map; subtract the basic suppression weight map from the unit value to obtain a suppression modulation weight map; multiply the suppression modulation weight map element-wise with the enhanced modulation feature map to obtain a target modulation feature map.

[0112] In this embodiment, the detection device multiplies the fused suppression feature map by a preset suppression modulation coefficient to obtain a basic suppression weight map; the detection device subtracts the basic suppression weight map from the unit value to obtain a suppression modulation weight map.

[0113] Specifically, the detection device subtracts the basic suppression weight map from all matrices with element values ​​of 1 and the same size as the basic suppression weight map, based on the basic suppression weight map and the unit values ​​in the identity matrix, to obtain the suppression modulation weight map. This operation converts the suppression weights into values ​​between 0 and 1, making the suppression region weights less than 1, thereby attenuating the characteristic response.

[0114] The detection device multiplies the suppressed modulation weight map and the enhanced modulation feature map element by element to obtain the target modulation feature map. This step ensures that the feature value is reduced in the suppressed region, thereby reducing false detections and context inconsistencies.

[0115] The detection device performs deep mining on a large-scale image annotation dataset to construct a context enhancement template library and a context suppression template library. These two knowledge bases encapsulate the statistical laws and visual patterns of co-occurrence and mutual exclusion between targets, providing strong prior knowledge support for the subsequent detection process. In the inference stage, the detection device can retrieve relevant enhancement and suppression templates from the knowledge base based on the preliminary detection results, and generate spatial attention maps through similarity matching calculations. These attention maps indicate the feature responses in different regions of the image that need to be enhanced or suppressed. Furthermore, the detection device obtains a comprehensive enhancement map and a comprehensive suppression map through fusion operations, and combines them with a general scene enhancement map generated from scene features to perform fine bidirectional modulation on the original high-resolution feature map. This modulation process can adaptively enhance feature expressions that conform to the target context while weakening interference signals that violate logical relationships, thereby effectively improving the detection capability of occluded targets, small targets, and rare targets in complex scenes, reducing false detections and false negatives, and achieving the goal of target detection by combining computer vision and big data mining.

[0116] Please refer to Figure 9 , Figure 9 This is a schematic diagram of a target detection device combining computer vision and big data mining according to an embodiment of this application. The device can be implemented in whole or in part through software, hardware, or a combination of both. The target detection device 9 combining computer vision and big data mining includes:

[0117] The feature extraction module 91 is used to obtain the image to be detected; to perform feature extraction on the image to be detected, and to obtain the target feature extraction map and the general scene enhancement map of the image to be detected.

[0118] The detection box extraction module 92 is used to extract target detection boxes based on the target feature extraction map, and obtain a number of initial target detection boxes of the image to be detected and the category label corresponding to each initial target detection box;

[0119] The template feature query module 93 is used to query the corresponding template feature vectors from the pre-built context enhancement template library and context suppression template library according to each initial target detection box and the category label corresponding to each initial target detection box, so as to obtain the enhancement template feature vector and suppression template feature vector corresponding to each initial target detection box.

[0120] The context attention extraction module 94 is used to perform similarity matching calculations between the target feature extraction map and the enhanced template feature vector and the suppressed template feature vector corresponding to each of the initial target detection boxes, so as to obtain the enhanced context attention map and the suppressed context attention map corresponding to each of the initial target detection boxes.

[0121] The feature fusion module 95 is used to fuse the enhanced context attention maps corresponding to each of the initial target detection boxes to obtain a fused enhanced feature map; and to fuse the suppressed context attention maps corresponding to each of the initial target detection boxes to obtain a fused suppressed feature map.

[0122] The target detection module 96 is used to perform bidirectional modulation on the target feature extraction map based on the fusion enhancement feature map, the fusion suppression feature map and the general scene enhancement map to obtain a target modulation feature map; and to perform target recognition processing on the target modulation feature map to obtain the target detection result of the image to be detected.

[0123] In this embodiment, a feature extraction module obtains an image to be detected; feature extraction is performed on the image to be detected to obtain a target feature extraction map and a general scene enhancement map; a detection box extraction module extracts target detection boxes based on the target feature extraction map to obtain several initial target detection boxes and category labels corresponding to each initial target detection box; a template feature query module queries corresponding template feature vectors from a pre-built context enhancement template library and a context suppression template library according to each initial target detection box and its corresponding category label to obtain enhancement template feature vectors and suppression template feature vectors corresponding to each initial target detection box; and a context attention extraction module extracts the target feature vectors. The extracted feature map is matched with the enhanced template feature vector and the suppressed template feature vector corresponding to each of the initial target detection boxes to obtain the enhanced context attention map and the suppressed context attention map corresponding to each of the initial target detection boxes. The feature fusion module fuses the enhanced context attention maps corresponding to each of the initial target detection boxes to obtain a fused enhanced feature map. The suppressed context attention maps corresponding to each of the initial target detection boxes are also fused to obtain a fused suppressed feature map. The target detection module performs bidirectional modulation on the extracted target feature map based on the fused enhanced feature map, the fused suppressed feature map, and the general scene enhancement map to obtain a target modulation feature map. The target modulation feature map is then subjected to target recognition processing to obtain the target detection result of the image to be detected. Based on preliminary detection results, relevant enhancement and suppression templates are retrieved from pre-constructed context enhancement and suppression template libraries. Spatial attention maps are generated through similarity matching, and fused enhancement and suppression feature maps are obtained through fusion operations. Combined with a general scene enhancement map generated from scene features, the target feature extraction map extracted from the image to be detected is finely modulated bidirectionally. This can adaptively enhance feature expressions that conform to the target context while weakening interference signals that violate logical relationships. Thus, it can effectively improve the detection capability of occluded targets, small targets, and rare targets in complex scenes, reduce false detections and false negatives, and improve the accuracy and efficiency of target detection.

[0124] Please refer to Figure 10 , Figure 10 This is a schematic diagram of the structure of a computer device provided in one embodiment of this application. The computer device 10 includes: a processor 101, a memory 102, and a computer program 103 stored in the memory 102 and executable on the processor 101. The computer device can store multiple instructions, which are adapted to be loaded and executed by the processor 101. Figures 1 to 8 For the method steps and specific execution process, please refer to [link / reference]. Figures 1 to 8Specific details will not be elaborated here.

[0125] The processor 101 may include one or more processing cores. The processor 101 connects to various parts of the server using various interfaces and lines, and executes various functions and processes data of the target detection device 9 combining computer vision and big data mining by running or executing instructions, programs, code sets, or instruction sets stored in memory 102, and by calling data stored in memory 102. Optionally, the processor 101 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 101 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required to be displayed on the touch screen; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 101 and may be implemented as a separate chip.

[0126] The memory 102 may include random access memory (RAM) or read-only memory. Optionally, the memory 102 may include a non-transitory computer-readable storage medium. The memory 102 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 102 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch instructions), instructions for implementing the various method embodiments described above, etc.; the data storage area may store data involved in the various method embodiments described above, etc. Optionally, the memory 102 may also be at least one storage device located remotely from the aforementioned processor 101.

[0127] This application embodiment also provides a storage medium that can store multiple instructions, which are adapted to be loaded and executed by a processor as described above. Figures 1 to 8 For the method steps and specific execution process, please refer to [link / reference]. Figures 1 to 8Specific details will not be elaborated here.

[0128] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0129] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0130] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the algorithm. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0131] In the embodiments provided by this invention, it should be understood that the disclosed apparatus / terminal devices and methods can be implemented in other ways. For example, the apparatus / terminal device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0132] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0133] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0134] If the integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms.

[0135] This invention is not limited to the above-described embodiments. If any modifications or variations to this invention do not depart from the spirit and scope of this invention, and if such modifications and variations fall within the scope of the claims and equivalent technologies of this invention, then this invention also intends to include such modifications and variations.

Claims

1. A target detection method combining computer vision and big data mining, characterized in that, Includes the following steps: Obtain the image to be detected; perform feature extraction on the image to be detected to obtain the target feature extraction map and the general scene enhancement map of the image to be detected; Based on the target feature extraction map, target detection boxes are extracted to obtain several initial target detection boxes of the image to be detected and the category label corresponding to each initial target detection box; Based on each initial target detection box and the category label corresponding to each initial target detection box, the corresponding template feature vectors are queried from the pre-built context enhancement template library and context suppression template library to obtain the enhancement template feature vector and suppression template feature vector corresponding to each initial target detection box. The target feature extraction map is matched with the enhanced template feature vector and the suppressed template feature vector corresponding to each of the initial target detection boxes to obtain the enhanced context attention map and the suppressed context attention map corresponding to each of the initial target detection boxes. The enhanced context attention maps corresponding to each initial target detection box are fused to obtain a fused enhanced feature map. The initial target detection boxes are fused together to obtain a fused suppression feature map. The fused enhanced feature map and the general scene enhanced map are multiplied by preset enhancement modulation coefficients and then added together to obtain a composite enhanced weight map; the composite enhanced weight map is added to a unit value to obtain an enhanced modulation weight map; the enhanced modulation weight map is multiplied element-wise with the target feature extraction map to obtain an enhanced modulation feature map. Multiply the fused suppression feature map by a preset suppression modulation coefficient to obtain a basic suppression weight map; subtract the basic suppression weight map from the unit value to obtain a suppression modulation weight map; multiply the suppression modulation weight map element-wise with the enhanced modulation feature map to obtain a target modulation feature map; The target modulation feature map is subjected to target recognition processing to obtain the target detection result of the image to be detected.

2. The target detection method combining computer vision and big data mining according to claim 1, characterized in that, The step of extracting features from the image to be detected to obtain a target feature extraction map and a general scene enhancement map of the image to be detected includes the following steps: The image to be detected is input into a preset convolutional neural network for multi-level feature extraction to obtain an intermediate-level feature extraction map and a target-level feature extraction map. The intermediate-level feature extraction map is used as the target feature extraction map, wherein the target-level feature extraction map is a feature extraction map at a deeper level than the intermediate-level feature extraction map. A global average pooling operation is performed on the feature extraction map of the target level to obtain a global average pooling feature map, which serves as the global scene feature vector. The global scene feature vector is then input into a preset deconvolutional network for upsampling and nonlinear mapping to obtain a general scene enhancement map.

3. The target detection method combining computer vision and big data mining according to claim 1, characterized in that, Before obtaining the enhanced template feature vector and the suppressed template feature vector corresponding to each initial target detection box by querying the pre-built context enhancement template library and context suppression template library respectively based on each initial target detection box and the category label corresponding to each initial target detection box, the method further includes the following steps: Big data mining is used to determine all target category pairs in a pre-defined image annotation dataset and construct a target category pair set. The target category pair set includes several target category pairs. Each target category pair includes a dominant target and associated targets. Each target category pair is associated with several initial images in the image annotation dataset that simultaneously contain the dominant target and associated targets. Obtain the co-occurrence frequency of each target category pair in the target category pair set; based on the co-occurrence frequency of each target category pair and a preset frequency threshold, target category pairs with a co-occurrence frequency higher than the frequency threshold are taken as high-frequency co-occurrence pairs, thereby obtaining several high-frequency co-occurrence pairs in the image annotation dataset; Template libraries are constructed based on the image annotation dataset and several high-frequency co-occurrence pairs in the image annotation dataset to obtain the context enhancement template library and the context suppression template library.

4. The target detection method combining computer vision and big data mining according to claim 3, characterized in that, The step of constructing a template library based on the image annotation dataset and several high-frequency co-occurrence pairs in the image annotation dataset to obtain the context enhancement template library and the context suppression template library includes the following steps: Based on the target location annotations in the image annotation dataset, the associated target regions are identified and the visual features of the associated target regions are extracted for each image associated with the high-frequency co-occurrence pairs, thereby obtaining visual feature vectors of several associated target regions for each initial image associated with each high-frequency co-occurrence pair. The visual feature vectors of each associated target region of the same initial image are aggregated to obtain the enhanced visual templates of each initial image for each high-frequency co-occurrence pair, and the context enhancement template library is constructed.

5. The target detection method combining computer vision and big data mining according to claim 3, characterized in that, The step of constructing a template library based on the image annotation dataset and several high-frequency co-occurrence pairs in the image annotation dataset to obtain the context enhancement template library and the context suppression template library includes the following steps: Based on the target location annotations in the image annotation dataset, and taking the bounding box of the dominant target as a reference, the spatial relationship vectors of the associated targets and the dominant target in each initial image associated with each high-frequency co-occurrence pair are calculated to obtain the spatial relationship vectors of each initial image associated with each high-frequency co-occurrence pair; based on the spatial relationship vectors of each initial image associated with each high-frequency co-occurrence pair and a preset Gaussian mixture model, probability density estimation is performed to construct the spatial distribution probability model of each high-frequency co-occurrence pair. Likelihood is calculated based on the spatial relationship vector and spatial distribution probability model of each initial image associated with the same high-frequency co-occurrence pair, and the likelihood of each initial image associated with each high-frequency co-occurrence pair is obtained. Based on the likelihood of each initial image associated with each high-frequency co-occurrence pair and a preset likelihood threshold, images with a likelihood lower than the preset likelihood threshold are identified as spatially anomalous images, and several spatially anomalous images of each high-frequency co-occurrence pair are obtained. Based on the spatial anomaly images of each high-frequency co-occurrence pair and the target location annotations in the image annotation dataset, the spatial anomaly images are used to identify associated target regions and extract visual features of associated target regions to obtain visual feature vectors of several associated target regions of each spatial anomaly image of each high-frequency co-occurrence pair. The visual feature vectors of each associated target region of the same spatial anomaly image are aggregated to obtain the abnormal visual templates of each spatial anomaly image for each high-frequency co-occurrence pair, and the context suppression template library is constructed.

6. The target detection method combining computer vision and big data mining according to claim 1, characterized in that, The step of performing similarity matching calculations between the extracted target feature map and the enhanced template feature vector and suppressed template feature vector corresponding to each of the initial target detection boxes to obtain the enhanced context attention map and suppressed context attention map corresponding to each of the initial target detection boxes includes the following steps: The enhanced template feature vector and the suppressed template feature vector are respectively used as input template feature vectors, and the input template feature vectors are reshaped by convolution kernels to obtain reshaped convolution kernel data. Convolution operations are performed based on the reconstructed data from the convolution kernel and the target feature extraction map. The dot product similarity between the convolution kernel and the feature vector corresponding to the spatial location is calculated at each spatial location of the target feature extraction map. The single-channel response map corresponding to the input template feature vector is obtained as the corresponding enhanced context attention map and suppressed context attention map. The enhanced context attention map and suppressed context attention map corresponding to each initial target detection box are obtained.

7. The target detection method combining computer vision and big data mining according to claim 3, characterized in that, The steps include fusing the enhanced context attention maps corresponding to each of the initial target detection boxes to obtain a fused enhanced feature map, and fusing the suppressed context attention maps corresponding to each of the initial target detection boxes to obtain a fused suppressed feature map. The enhanced context attention map and the suppressed context attention map are used as input context attention maps, respectively; The maximum value operation method is used to traverse the pixel intensity value of each pixel position in each input context attention map to obtain the pixel intensity value of each pixel position in each input context attention map. The maximum pixel intensity value is obtained by taking the maximum value of the pixel intensity values ​​of each input context attention map at the same pixel position. The fusion feature map corresponding to the input context attention map is constructed as the corresponding fusion enhancement feature map and fusion suppression feature map.

8. A target detection device combining computer vision and big data mining, characterized in that, include: The feature extraction module is used to obtain the image to be detected; Feature extraction is performed on the image to be detected to obtain the target feature extraction map and the general scene enhancement map of the image to be detected; The detection box extraction module is used to extract target detection boxes based on the target feature extraction map, and obtain several initial target detection boxes of the image to be detected and the category label corresponding to each initial target detection box; The template feature query module is used to query the corresponding template feature vectors from the pre-built context enhancement template library and context suppression template library according to each initial target detection box and the category label corresponding to each initial target detection box, so as to obtain the enhancement template feature vector and suppression template feature vector corresponding to each initial target detection box. The context attention extraction module is used to perform similarity matching calculations between the target feature extraction map and the enhanced template feature vector and the suppressed template feature vector corresponding to each of the initial target detection boxes, so as to obtain the enhanced context attention map and the suppressed context attention map corresponding to each of the initial target detection boxes. The feature fusion module is used to fuse the enhanced context attention maps corresponding to each of the initial target detection boxes to obtain a fused enhanced feature map. The initial target detection boxes are fused together to obtain a fused suppression feature map. The target detection module is used to multiply the fused enhanced feature map and the general scene enhanced map by preset enhancement modulation coefficients and then add them together to obtain a composite enhanced weight map; add the composite enhanced weight map to a unit value to obtain an enhanced modulation weight map; and multiply the enhanced modulation weight map element-wise with the target feature extraction map to obtain an enhanced modulation feature map. Multiply the fused suppression feature map by a preset suppression modulation coefficient to obtain a basic suppression weight map; subtract the basic suppression weight map from the unit value to obtain a suppression modulation weight map; multiply the suppression modulation weight map element-wise with the enhanced modulation feature map to obtain a target modulation feature map; The target modulation feature map is subjected to target recognition processing to obtain the target detection result of the image to be detected.

9. A computer device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the object detection method combining computer vision and big data mining as described in any one of claims 1 to 7.

Images