An information processing method, apparatus and device

Abstract

The present application provides an information processing method, device and equipment, wherein the information processing method comprises: determining a target feature map corresponding to a target object according to an image size of the target object and first information; obtaining target coordinates of each pixel on the target feature map according to original coordinates of each pixel on the target mask and an input picture; obtaining a center coordinate of the target object according to the target coordinates; obtaining a target weight corresponding to each pixel according to the target coordinates and the center coordinate; training an object detection model according to the target weight corresponding to each pixel; wherein the first information comprises a mapping relationship between an object image size and a feature map on a feature pyramid; the target feature map is a feature map of a target layer on the feature pyramid; and the target object corresponds to the target mask. The present application improves the model precision obtained by the information processing scheme for training the target detection model in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information processing, and in particular to an information processing method, device and equipment. BACKGROUND

[0002] Target detection is an important and active research direction in the field of computer vision, and it is the basis for other complex vision tasks. As the cornerstone of image understanding and computer vision, target detection is a necessary prerequisite for solving higher-level vision tasks such as instance segmentation, scene understanding, target tracking, action recognition, image description, and event detection. Target detection in computer vision is defined as using image processing and machine learning techniques to automatically locate the subject of interest in an image or video, and telling users the "where" and "what" information in the form of the coordinates of the bounding box of the subject.

[0003] The target detection algorithm based on deep learning mainly includes two technical routes: anchor-based and anchor-free methods. The existing anchor-free algorithms (such as CenterNet, FCOS, FSAS, FoveaBox, etc.) still cannot match the anchor-based methods in terms of positioning evaluation indicators (such as the average precision mean Map) on public data sets.

[0004] Taking CenterNet, a typical representative of anchor-free, as an example, the calculation process of CenterNet classification loss (loss value) is analyzed, and the following three shortcomings can be found:

[0005] 1. CenterNet takes the mapping point of the center of the real object bounding box as the center point, and takes the width and height of the object bounding box as the long and short axes, and then generates a single Gaussian heatmap with a fixed shape for each labeled object. For objects close to regular elliptical shape (such as standing pedestrians), the heatmap has less deviation from the actual shape of the object, and can better weight the loss predicted for each point on the object. However, when the target object is irregular in shape, especially when there is a hollow area in the object (such as a circular ring object), since the weight on the heatmap is only related to the distance of the center point, and the distribution shape of the single Gaussian heatmap is a real circle or ellipse with the center point as the peak value and gradually decreasing towards the four corners, the heatmap cannot well weight the loss predicted for each point on the feature map, and background noise is easily introduced.

[0006] 2. CenterNet does not consider the size of the object, and uniformly maps all objects to a single feature map for processing.

[0007] 3. In the CenterNet training process, each target is mapped to a feature map, and a single Gaussian heatmap is used to calculate the loss, which is adjusted by weighting. Essentially, this is a way of mapping each target to many dense points, which can introduce noise points and affect the training of the model and the final performance. SUMMARY

[0008] The present application aims to provide an information processing method, device and equipment to improve the accuracy of the model obtained by the information processing scheme for training the target detection model in the prior art.

[0009] To achieve the above-mentioned purpose, the embodiments of the present application provide an information processing method, comprising:

[0010] According to the image size of the target object and the first information, the target feature map corresponding to the target object is determined;

[0011] According to the original coordinates of each pixel on the target mask and the target feature map on the input picture, the target coordinates of each pixel on the target feature map are obtained;

[0012] According to the target coordinates, the center coordinates of the target object are obtained;

[0013] According to the target coordinates and the center coordinates, the target weight corresponding to each pixel is obtained;

[0014] According to the target weight corresponding to each pixel, the object detection model is trained;

[0015] The first information includes the mapping relationship between the object image size and the feature map on the feature pyramid; the target feature map is the feature map of the target layer on the feature pyramid.

[0016] The target object corresponds to the target mask.

[0017] Optionally, the first information further includes a reference size for the object image and a mapping reference layer for the feature pyramid.

[0018] According to the image size of the target object and the first information, the target feature map corresponding to the target object is determined, comprising:

[0019] According to the image size of the target object, the mapping relationship, the reference size and the mapping reference layer, the target feature map corresponding to the target object is determined.

[0020] Optionally, the obtaining the target coordinates of each pixel on the target feature map according to the target feature map and the original coordinates of each pixel on the target mask on the input picture comprises:

[0021] obtaining a scaling ratio between the input picture and the target feature map;

[0022] obtaining the target coordinates of each pixel on the target feature map according to the scaling ratio and the original coordinates of each pixel on the target mask on the input picture.

[0023] Optionally, the obtaining the target weight corresponding to each pixel according to the target coordinates and the center coordinates comprises:

[0024] obtaining a diagonal line length of an image obtained by mapping each pixel on the target feature map according to the target coordinates;

[0025] obtaining an initial weight corresponding to each pixel according to the diagonal line length, the target coordinates and the center coordinates;

[0026] obtaining the target weight corresponding to each pixel according to the initial weight.

[0027] Optionally, the method further comprises:

[0028] obtaining a foreground loss value and a background loss value corresponding to each pixel after mapping each pixel on the target feature map;

[0029] selecting a target pixel from each pixel according to the foreground loss value;

[0030] the training the object detection model according to the target weight corresponding to each pixel comprises:

[0031] training the object detection model according to the background loss value, the foreground loss value corresponding to the target pixel and the target weight;

[0032] wherein the foreground loss value comprises a classification loss value and a regression loss value, and the background loss value comprises the classification loss value.

[0033] The embodiment of the application further provides an information processing method, comprising:

[0034] obtaining a foreground loss value, a background loss value and a target weight corresponding to each pixel of a target object after mapping each pixel on a feature map;

[0035] selecting a target pixel from each pixel according to the foreground loss value;

[0036] According to the background loss value, and a foreground loss value and a target weight corresponding to the target pixel, the object detection model is trained.

[0037] The foreground loss value comprises a classification loss value and a regression loss value, and the background loss value comprises a classification loss value.

[0038] Embodiments of the present application also provide an information processing device, comprising:

[0039] A first determining module is configured to determine a target feature map corresponding to a target object according to an image size of the target object and first information.

[0040] A first processing module is configured to obtain target coordinates of each pixel on the target feature map according to the target feature map and original coordinates of the each pixel on an input picture on a target mask.

[0041] A first obtaining module is configured to obtain a center coordinate of the target object according to the target coordinates.

[0042] A second obtaining module is configured to obtain a target weight corresponding to the each pixel according to the target coordinates and the center coordinate.

[0043] A first training module is configured to train an object detection model according to the target weight corresponding to the each pixel.

[0044] The first information comprises a mapping relationship between an object image size and a feature map on a feature pyramid, and the target feature map is a feature map of a target layer on the feature pyramid.

[0045] The target object corresponds to the target mask.

[0046] Optionally, the first information further comprises a reference size for an object image and a mapping reference layer for the feature pyramid.

[0047] The determination of the target feature map corresponding to the target object according to the image size of the target object and the first information comprises:

[0048] The target feature map corresponding to the target object is determined according to the image size of the target object, the mapping relationship, the reference size and the mapping reference layer.

[0049] Optionally, the obtaining of the target coordinates of the each pixel on the target feature map according to the target feature map and the original coordinates of the each pixel on the input picture on the target mask comprises:

[0050] A scaling ratio between the input picture and the target feature map is obtained.

[0051] According to the scaling ratio and original coordinates of each pixel on the input picture, target coordinates of the each pixel on the target feature map are obtained.

[0052] Optionally, the obtaining the target weight corresponding to the each pixel according to the target coordinates and the center coordinates comprises:

[0053] According to the target coordinates, a diagonal line length of an image obtained by mapping the each pixel on the target feature map is obtained.

[0054] According to the diagonal line length, the target coordinates and the center coordinates, an initial weight corresponding to the each pixel is obtained.

[0055] According to the initial weight, the target weight corresponding to the each pixel is obtained.

[0056] Optionally, the method further comprises:

[0057] The third obtaining module is configured to obtain foreground loss values and background loss values corresponding to the each pixel after the each pixel is mapped on the target feature map.

[0058] The first selecting module is configured to select a target pixel from the each pixel according to the foreground loss values.

[0059] The training the object detection model according to the target weight corresponding to the each pixel comprises:

[0060] The second training module is configured to train the object detection model according to the background loss values, the foreground loss values corresponding to the target pixel and the target weight.

[0061] The foreground loss values comprise classification loss values and regression loss values, and the background loss values comprise classification loss values.

[0062] Embodiments of the present application also provide an information processing device, comprising:

[0063] The fourth obtaining module is configured to obtain foreground loss values, background loss values and target weights corresponding to the each pixel of the target object after the each pixel is mapped on the feature map.

[0064] The second selecting module is configured to select a target pixel from the each pixel according to the foreground loss values.

[0065] The second training module is configured to train the object detection model according to the background loss values, the foreground loss values corresponding to the target pixel and the target weight.

[0066] The foreground loss values comprise classification loss values and regression loss values, and the background loss values comprise classification loss values.

[0067] The embodiment of the present application further provides an information processing device, comprising a processor and a transceiver;

[0068] The processor is configured to determine a target feature map corresponding to a target object according to an image size of the target object and first information.

[0069] According to the original coordinates of each pixel on the target mask in the input picture, the target coordinates of the each pixel on the target feature map are obtained.

[0070] According to the target coordinates, the center coordinates of the target object are obtained.

[0071] According to the target coordinates and the center coordinates, the target weight corresponding to the each pixel is obtained.

[0072] According to the target weight corresponding to the each pixel, a target detection model is trained.

[0073] The first information comprises a mapping relationship between an object image size and a feature map on a feature pyramid; and the target feature map is a feature map of a target layer on the feature pyramid.

[0074] The target object corresponds to the target mask.

[0075] Optionally, the first information further comprises a reference size of an object image and a mapping reference layer of the feature pyramid.

[0076] The processor is configured to determine the target feature map corresponding to the target object according to the image size of the target object and the first information.

[0077] The processor is configured to determine the target feature map corresponding to the target object according to the image size of the target object, the mapping relationship, the reference size and the mapping reference layer.

[0078] Optionally, the processor is configured to obtain the target coordinates of the each pixel on the target feature map according to the original coordinates of each pixel on the target mask in the input picture.

[0079] The processor is configured to obtain a scaling ratio between the input picture and the target feature map.

[0080] The processor is configured to obtain the target coordinates of the each pixel on the target feature map according to the scaling ratio and the original coordinates of each pixel on the target mask in the input picture.

[0081] Optionally, the processor is configured to obtain the target weight corresponding to the each pixel according to the target coordinates and the center coordinates.

[0082] According to the target coordinates, a diagonal line length of an image obtained by mapping each pixel on the target feature map is acquired;

[0083] According to the diagonal line length, target coordinates and center coordinates, an initial weight corresponding to each pixel is acquired;

[0084] According to the initial weight, a target weight corresponding to each pixel is acquired.

[0085] Optionally, the processor is further configured to:

[0086] After acquiring the foreground loss value and the background loss value corresponding to each pixel after mapping on the target feature map;

[0087] According to the foreground loss value, a target pixel is selected from each pixel;

[0088] According to the target weight corresponding to each pixel, the object detection model is trained, including:

[0089] According to the background loss value, and the foreground loss value and the target weight corresponding to the target pixel, the object detection model is trained;

[0090] The foreground loss value includes a classification loss value and a regression loss value; and the background loss value includes a classification loss value.

[0091] Embodiments of the present application also provide an information processing device, including a processor and a transceiver;

[0092] The processor is configured to acquire a foreground loss value, a background loss value and a target weight corresponding to each pixel of a target object after mapping on a feature map;

[0093] According to the foreground loss value, a target pixel is selected from each pixel;

[0094] According to the background loss value, and the foreground loss value and the target weight corresponding to the target pixel, the object detection model is trained;

[0095] The foreground loss value includes a classification loss value and a regression loss value; and the background loss value includes a classification loss value.

[0096] Embodiments of the present application also provide an information processing device, including a memory, a processor and a program stored in the memory and executable on the processor; the processor executes the program to implement the information processing method.

[0097] The embodiment of the present application further provides a readable storage medium, which has a program stored thereon, and the program is executed by a processor to realize the steps in the information processing method.

[0098] The beneficial effects of the above technical solutions of the present application are as follows:

[0099] In the above scheme, the information processing method determines a target feature map corresponding to a target object according to an image size of the target object and first information; obtains target coordinates of each pixel on the target feature map according to original coordinates of the each pixel on the target mask in an input picture; acquires a center coordinate of the target object according to the target coordinates; acquires a target weight corresponding to the each pixel according to the target coordinates and the center coordinate; trains an object detection model according to the target weight corresponding to the each pixel; wherein the first information includes a mapping relationship between an object image size and a feature map on a feature pyramid; the target feature map is a feature map of a target layer on the feature pyramid; the target object corresponds to the target mask; the model training is realized by introducing a mask, so that the shape information of the target object is reserved on the mapped feature map, the supervision signal for the model training is more accurate, and the model precision is further improved. BRIEF DESCRIPTION OF DRAWINGS

[0100] Figure 1 The information processing method flowchart of the embodiment of the present application Figure One ;

[0101] Figure 2 The information processing method flowchart of the embodiment of the present application Figure Two ;

[0102] Figure 3 The information processing method implementation framework diagram of the embodiment of the present application

[0103] Figure 4 The object shape diagram of the embodiment of the present application

[0104] Figure 5 The information processing device structure diagram of the embodiment of the present application Figure One ;

[0105] Figure 6 The information processing device structure diagram of the embodiment of the present application Figure Two ;

[0106] Figure 7 The information processing equipment structure diagram of the embodiment of the present application Figure One ;

[0107] Figure 8 The information processing equipment structure diagram of the embodiment of the present applicationFigure Two . DETAILED DESCRIPTION

[0108] In order to make the technical problems, technical solutions and advantages of the present application clearer, specific embodiments will be described in detail below with reference to the accompanying drawings.

[0109] The present application is to improve the model precision obtained by the information processing scheme for target detection model training in the prior art, and provides an information processing method, as shown in Figure 1 , comprising:

[0110] Step 11: determining the target feature map corresponding to the target object according to the image size of the target object and the first information;

[0111] Step 12: obtaining the target coordinates of each pixel on the target feature map according to the original coordinates of each pixel on the target mask in the input picture;

[0112] Step 13: obtaining the center coordinates of the target object according to the target coordinates;

[0113] Step 14: obtaining the target weight corresponding to each pixel according to the target coordinates and the center coordinates;

[0114] Step 15: training the object detection model according to the target weight corresponding to each pixel; wherein the first information includes the mapping relationship between the object image size and the feature map on the feature pyramid; the target feature map is the feature map of the target layer on the feature pyramid; the target object corresponds to the target mask.

[0115] Step 13 can specifically include: obtaining the center coordinates of the target object according to the target coordinates and the number of pixels. Regarding the mapping relationship, it can be that the larger the image size, the higher the corresponding layer of the feature map; but it is not limited thereto. The center coordinates can be specifically the centroid coordinates.

[0116] The information processing method provided by the embodiment of the present application determines the target feature map corresponding to the target object according to the image size of the target object and first information, obtains target coordinates of each pixel on the target feature map according to the target feature map and original coordinates of each pixel on the target mask on an input picture, acquires the center coordinates of the target object according to the target coordinates, acquires the target weight corresponding to each pixel according to the target coordinates and the center coordinates, and trains an object detection model according to the target weight corresponding to each pixel, wherein the first information includes a mapping relationship between the object image size and a feature map on a feature pyramid, the target feature map is a feature map of a target layer on the feature pyramid, and the target object corresponds to the target mask. The model training is performed by introducing a mask, so that the shape information of the target object is retained on the mapped feature map, the supervision signal for the model training is more accurate, the model precision is further improved, and the problem of low model precision obtained by the existing solution is solved.

[0117] Further, the first information further includes a reference size of an object image and a mapping reference layer of the feature pyramid, and the determination of the target feature map corresponding to the target object according to the image size of the target object and the first information includes the determination of the target feature map corresponding to the target object according to the image size of the target object, the mapping relationship, the reference size and the mapping reference layer.

[0118] In this way, the target feature map can be more accurately and quickly determined.

[0119] The obtaining of the target coordinates of each pixel on the target feature map according to the target feature map and the original coordinates of each pixel on the target mask on the input picture includes the acquisition of a scaling ratio between the input picture and the target feature map and the obtaining of the target coordinates of each pixel on the target feature map according to the scaling ratio and the original coordinates of each pixel on the target mask on the input picture.

[0120] In this way, the target coordinates of each pixel can be accurately obtained.

[0121] In the embodiment of the present application, the acquisition of the target weight corresponding to each pixel according to the target coordinates and the center coordinates includes the acquisition of the diagonal line length of the image obtained by the mapping of each pixel on the target feature map according to the target coordinates, the acquisition of the initial weight corresponding to each pixel according to the diagonal line length, the target coordinates and the center coordinates, and the acquisition of the target weight corresponding to each pixel according to the initial weight.

[0122] In this way, a more balanced weight can be ensured. In the step of obtaining a diagonal line length of an image obtained by mapping each pixel on the target feature map according to the target coordinates, the following steps can be further included: obtaining a maximum value in a vertical direction, a maximum value in a horizontal direction, a minimum value in the vertical direction, and a minimum value in the horizontal direction; determining a first vertex and a second vertex according to the maximum value in the vertical direction, the maximum value in the horizontal direction, the minimum value in the vertical direction, and the minimum value in the horizontal direction; and obtaining a distance between the first vertex and the second vertex as the diagonal line length of the image obtained by mapping each pixel on the target feature map, wherein the horizontal and vertical coordinates of the first vertex are the minimum value in the horizontal direction and the maximum value in the vertical direction, and the horizontal and vertical coordinates of the second vertex are the maximum value in the horizontal direction and the minimum value in the vertical direction; or the horizontal and vertical coordinates of the first vertex are the minimum value in the horizontal direction and the minimum value in the vertical direction, and the horizontal and vertical coordinates of the second vertex are the maximum value in the horizontal direction and the maximum value in the vertical direction. Alternatively, the following steps can be further included: obtaining a bounding rectangle of the image obtained by mapping each pixel on the target feature map; and obtaining a diagonal line of the bounding rectangle as the diagonal line length of the image obtained by mapping each pixel on the target feature map. The two methods are essentially the same, and are not limited herein.

[0123] In the step of obtaining a target weight corresponding to each pixel according to the initial weight, the following steps can be further included: performing normalization again according to a maximum value and a minimum value in the initial weight to obtain the target weight corresponding to each pixel. For details, refer to the formula described below.

[0124] Further, the information processing method further includes: obtaining foreground loss values and background loss values corresponding to each pixel after the pixel is mapped on the target feature map; selecting a target pixel from the pixels according to the foreground loss values; and training the object detection model according to the background loss values, the foreground loss values corresponding to the target pixel, and the target weight. The foreground loss values include classification loss values and regression loss values, and the background loss values include classification loss values.

[0125] In this way, the target key points in the sample and the difficult sample points can be selected for targeted training, so that the model can learn the representation of the target object more effectively, and the precision of the obtained model can be further improved. The background loss value here refers to the sum of loss values corresponding to all background points, and can also be the sum of loss values corresponding to selected background points, and does not necessarily refer to the sum of loss values corresponding to all background points. The background points are points on the target feature map that are not mapped.

[0126] The embodiment of the present application also provides an information processing method, as shown inFigure 2 As shown in the figure, comprising:

[0127] Step 21: obtaining the foreground loss value, the background loss value and the target weight corresponding to each pixel of the target object after mapping on the feature map;

[0128] Step 22: selecting a target pixel from the pixels according to the foreground loss value;

[0129] Step 23: training the object detection model according to the background loss value, the foreground loss value corresponding to the target pixel and the target weight; wherein the foreground loss value includes a classification loss value and a regression loss value; the background loss value includes a classification loss value.

[0130] In step 21, the foreground loss value, the background loss value and the target weight corresponding to each pixel after mapping on the target feature map can be obtained. The background loss value is consistent for each pixel, that is, the loss value corresponding to all points not mapped on the target feature map.

[0131] The information processing method provided by the embodiment of the application can obtain the foreground loss value, the background loss value and the target weight corresponding to each pixel of the target object after mapping on the feature map; select a target pixel from the pixels according to the foreground loss value; and train the object detection model according to the background loss value, the foreground loss value corresponding to the target pixel and the target weight; wherein the foreground loss value includes a classification loss value and a regression loss value; the background loss value includes a classification loss value. The method can realize targeted training of target key points and difficult sample points in the sample, make the model more effectively learn the representation of the target object, and thus improve the precision of the obtained model, and can solve the problem of low model precision in the prior art.

[0132] The information processing method provided by the embodiment of the application will be described below by taking an object as an example.

[0133] To achieve the above object, and to effectively design an automatic abnormal point sample screening mechanism, the embodiment of the application provides an information processing method, which can be specifically implemented as a target detection method based on a hierarchical semantic heat map, and the general framework can be as shown in the figure. Figure 3 As shown in the figure,

[0134] According to the ground truth mask, each object mask is matched to different feature map levels (also understood as different layers of feature maps), to obtain a low-level mask, a middle-level mask, and a high-level mask, and the neural network model (backbone) is trained according to the masks;

[0135] Subsequently, an image can be input into the neural network model (backbone) for detection, and a shape-aware heatmap can be obtained, which can indicate that the detected object contains an irregular object with an internal cavity (see the RGB image in the figure); in the shape-aware heatmap, the mass center is the center point. In addition, in the neural network model, the low-level mask corresponds to a low-level bounding box loss (Low level Bbox loss, corresponding to Figure 3 A) and a low-level heatmap loss (Low level Heatmap loss, corresponding to Figure 3 B), the middle-level mask corresponds to a middle-level bounding box loss (Middle level Bbox loss, corresponding to Figure 3 C) and a middle-level heatmap loss (Middle level Heatmap loss, corresponding to Figure 3 D), and the high-level mask corresponds to a high-level bounding box loss (High level Bbox loss, corresponding to Figure 3 E) and a high-level heatmap loss (High level Heatmap loss, corresponding to Figure 3 F).

[0136] Specifically, the present scheme includes model training based on multi scale soft shape-aware heatmap (multi-scale fine shape-aware heatmap) supervision signal and sample point sampling based on online representative point selection and outlier removal mechanism. The following will be described in detail respectively:

[0137] 1. Model training based on multi-scale soft shape-aware heatmap supervision signal

[0138] For the training sample set for instance segmentation, in addition to the bounding box annotation of the target object, there is also mask annotation information of the target object. Compared with the bounding box annotation of the target object, the mask information of the target object can more accurately reflect the image and contour area belonging to the target object. Therefore, unlike the existing method, which only uses the bounding box of the target object to represent the position of the target, the present scheme projects all (pixel) points on the mask of the target object onto the target feature map. Here, the target feature map is not unique, and a feature pyramid similar to FPN is adopted. According to the size of the target object (corresponding to the image size described above), the masks of target objects of different scales are mapped to feature maps of different levels. Specifically, the masks of large (size) target objects can be mapped to high-level feature maps of the network, and the masks of small (size) target objects can be mapped to low-level feature maps of the network. The layer (corresponding to the target layer described above) to which the mask of the target object is mapped can be determined using the following formula:

[0139]

[0140] wherein l' represents the target layer, represents the floor function, l0 represents the mapping reference layer, wh represents the image size of the target object (which can be the length multiplied by the width, i.e., the area of the bounding box), and 224 is an empirical parameter equal to the size of the pre-trained model sample (corresponding to the reference size described above). It can be adjusted according to specific circumstances, and l0 can be understood as the case where the side length of the square root of the area of the bounding box of the target object is 224, and the mask is mapped to the feature map of the layer. Here, it is assumed that l0 = 5. As shown in Figure 3 As shown, taking the detection of contraband in a security inspection machine x-ray image as an example, the ground truth mask (true target mask, i.e., the original mask) represents the mask annotation of each target object on the input image, and the high-level mask, middle-level mask, and low-level mask represent, respectively, that according to the size of the object in the image (corresponding to the image size described above), the mask of a large (size) target object is mapped to the highest layer feature map of the network, the mask of a medium-sized target object is mapped to a middle layer feature map, and the mask of a small (size) target object is mapped to a low layer feature map.

[0141] Since the mask annotation of the target object can accurately reflect the shape of the object, in order to maintain the shape invariability of the object, the following formula can be used to map the mask of the target object to the feature map:

[0142]

[0143] Here, represents the coordinate value (i.e., the original coordinate) of the kth point on the mask of the ith target object on the input picture (i.e., the original picture), N obj represents the total number of objects in the input picture, represents the area (specifically, the number of pixels) of the mask of the ith target object, s l represents the cumulative step length (i.e., the scaling ratio, which can also be understood as the step length from the size on the input picture to the target layer) of mapping the target object to the lth (here, the letter l is explained) layer feature map, and L is equal to the total number of layers (corresponding to the layer where the target feature map is located) minus 1. represents the coordinate of the mapped pixel point on the feature map (corresponding to the target coordinate). It should be noted that when the objects in the input picture are relatively dense, since the mapped feature map is generally smaller than the size of the input picture, there is a case that the masks of different categories of objects are mapped to the same point, and therefore the classification loss in the training can adopt a c two-classification manner (about c, it can be determined according to the number of categories to be determined, such as whether it is a two-classification to detect whether it is a person).

[0144] In addition, in the existing method, the center of the object mapping to the feature map is generally selected as the center of the object bounding rectangle, such as the center point selected by centernet as the center of the target single Gaussian heatmap. When the object is irregularly shaped (as shown in Figure 4 ), the center of the bounding rectangle cannot accurately represent the physical center of the object, and therefore subsequent other operations (such as the generation of the heatmap weight map, the label assignment of the point on the feature map, etc.) based on the center point will also be affected.

[0145] Therefore, the scheme adopts the centroid of the mapped mask as the physical center of the target object, as shown in the following formula:

[0146]

[0147] The formula indicates that the centroid of the ith target object = the sum of the target coordinates divided by the number of pixels of the target object. The related parameter meanings can be referred to the above.

[0148] Based on the above analysis, we propose a soft shape-aware heatmap to more accurately weight the loss calculated for all points on the projected feature map according to the shape of the object. The specific calculation process of the weights can be expressed as follows:

[0149]

[0150]

[0151] in, This represents the initial weight of the k-th pixel of the i-th object. This represents the length of the diagonal of the rectangle corresponding to the i-th object after mapping. This represents the target weight of the k-th pixel of the i-th object. and These represent the points (pixels) of the i-th object. The minimum and maximum values ​​of the initial and target weights are calculated using a normalization operation (this avoids excessive concentration of weights and achieves a balanced distribution). The meanings of other parameters can be found above.

[0152] A computational example for the above scheme (training model) can be found as follows: Figure 3 The right-middle section shows the shape-aware heatmap calculated from an X-ray image of a pair of scissors. The heatmap in this figure accurately preserves the shape information of the scissors, and the distribution of weights better reflects the importance of the loss at different locations for object recognition and localization.

[0153] 2. Sampling based on online representative point selection and outlier removal;

[0154] Since noise may be introduced during object masking, and rounding operations in the feature map coordinate mapping also introduce certain systematic errors, this paper proposes an online method for selecting object keypoints based on loss, thereby avoiding the introduction of outliers. The specific calculation method can be expressed as follows:

[0155]

[0156] Wherein, BpTrain represents back propagation (for updating, training model), Nobj represents total number of target objects, SelectTopKLeast represents selecting K points with minimum loss from Nmask points (pixels) mapped on mask of a certain target object for back propagation gradient update; loss of the i-th target object (corresponding to the above foreground loss value) (i.e. classification loss + regression loss), i.e. a pixel point mapped as a foreground point on the feature map will correspond to a classification loss (i.e. loss of whether it is a certain class) and a regression loss (i.e. loss of size), Nmask represents number of (pixels) points, and loss bgi represents classification loss of a certain background point (sum of the loss corresponds to the above background loss value, i.e. sum of background loss values of all background points on the entire target feature map). Of course, background points can also be selected, not necessarily all.

[0157] In the present scheme, it is considered that for a set of mapped feature points on a certain target object, each point plays a different role in the identification and positioning of the target object. Generally, the more representative the region is, the more accurately the object can be positioned and identified, such as logo image feature regions on different brands of similar category goods. Therefore, the present scheme here uses the method of online loss ranking to mine key point feature information of the identified target, and can also exclude the interference of outliers introduced by mislabeling and other factors.

[0158] In addition, it should be noted that: in addition to using dynamic selection of topK (i.e. top K) points with minimum loss to train the model, topR (i.e. top R) points with maximum loss can also be selected together for training after outlier analysis. The topR points with maximum loss can be regarded as belonging to the target object, but are difficult samples that are easily misidentified by the current detector, i.e. the present scheme can use the following formula for sampling:

[0159]

[0160] Wherein, SelectTopRLagest represents selecting R points with maximum loss from Nmask points (pixels) mapped on mask of a certain target object for back propagation gradient update.

[0161] Based on the above, subsequent model training can be performed according to the background loss value, and the weights of the selected pixel points and the foreground loss value; the parameters of the model are adjusted (updated) in reverse.

[0162] It is explained that the above-mentioned part 1 (model training) and 2 (sampling) involved can be used in combination, or used separately, here is not limited.

[0163] From the above, in the embodiments of the application:

[0164] 1. A target detection method based on hierarchical semantic heat map is proposed. In the model training stage, a multi-scale soft shape-aware semantic heatmap can be generated for each target object in each picture, the correspondence between the features in the best feature pyramid and the target image information is established (corresponding to the content of determining the target layer), and the shape information of the target object is preserved on the mapped feature map, so that the supervision signal for network training is more accurate. In addition, the scheme proposes to use an online object key point sampling method to mine training samples. By sorting the loss of all response positions on the heatmap corresponding to each target object in each training round, and selecting part of the pixel points according to the size of the loss to participate in the training process of the model, the object key point mining, difficult sample mining and outlier removal due to incorrect labeling are realized, so that the performance of the model obtained by the final training is improved.

[0165] 2. The provided scheme mainly includes two core operations: generation of multi-scale soft shape-aware heatmap and sampling based on online representative point selection and outlier removal.

[0166] 3. Regarding the generation of multi-scale soft shape-aware heatmap: first, according to the size of the target object in the training image (corresponding to the input image), the target object is mapped to different layers of the feature pyramid. Large size target objects are mapped to high layers of the feature pyramid, and small size target objects are mapped to bottom layers of the pyramid. Unlike existing schemes that only map the center point of the bounding rectangle of each target object as the center of the target on the feature map, the present scheme proposes to map all the labeled mask points (all mask points) of each target object, and then calculate the centroid of all the mapped mask points as the center of the target object on the feature map. Finally, the heatmap retaining the shape information of the target can be generated according to the centroid for the training of the model.

[0167] 4. Sampling based on online representative point selection and outlier removal: Select a certain proportion of loss for each target object, such as the top K smallest and the top R largest, and do loss back propagation at the corresponding position to update the model parameters, and select the key points of the target object and mine difficult samples to improve the final training effect of the model.

[0168] In summary, in this scheme, a model training method based on multi-scale soft shape-aware heatmap supervision signal and a sampling method based on online representative point selection and outlier removal are proposed to improve the shortcomings of existing anchor-free based detection technology. Specifically, by mapping objects of different sizes to different layers of the feature pyramid according to the size of the object, and projecting all mask points of the object to form a shape-aware heatmap for each labeled object, the shape information of the object is maximized, and the object centroid is selected as the center according to the projected shape information, and then a more refined training model supervision signal is generated to improve the accuracy of the obtained model. In addition, an online object key point sampling method is used to mine training samples, by sorting the loss of all response positions on the heatmap corresponding to each target object in each training round, and selecting part of the points to participate in the training of the model according to the size of the loss, the object key point and difficult sample mining and the elimination of outliers introduced by incorrect labeling are realized, so that the performance of the model obtained by the final training is improved.

[0169] In summary, unlike existing solutions: In training, the midpoint of the target bounding rectangle is mapped to the feature map to generate a fixed shape single Gaussian heatmap, which cannot generate an accurate supervision signal that preserves the shape information of the target. This scheme proposes a multi-scale shape-aware semantic heatmap based scheme to generate the supervision signal of the training model; by mapping all mask annotation points of each target object to the feature map, calculating the centroid of all mask mapping points of the target object as the mapping center (i.e. the center point) of the target object, and then generating a heatmap according to all target mask mapping points as the supervision signal for model training, the generated supervision signal is more accurate than the existing solution, thereby improving the performance of the final model.

[0170] In addition, the scheme proposes to use an online representative point selection-based sampling technology to sample all mask mapping points of each target according to loss, select some (pixel) points with large and small loss to mine target key points and difficult sample points for targeted training, so that the network can learn the representation of the target more effectively, thereby further improving the positioning and recognition ability of the detector.

[0171] The embodiment of the application further provides an information processing device, as shown in the figure, comprising: Figure 5

[0172] The first determining module 51 is configured to determine a target feature map corresponding to the target object according to the image size of the target object and first information.

[0173] The first processing module 52 is configured to obtain target coordinates of each pixel on the target feature map according to the target feature map and original coordinates of each pixel on the target mask on an input picture.

[0174] The first obtaining module 53 is configured to obtain the center coordinates of the target object according to the target coordinates.

[0175] The second obtaining module 54 is configured to obtain target weights corresponding to each pixel according to the target coordinates and the center coordinates.

[0176] The first training module 55 is configured to train an object detection model according to the target weights corresponding to each pixel.

[0177] The first information comprises a mapping relationship between an object image size and a feature map on a feature pyramid; and the target feature map is a feature map of a target layer on the feature pyramid.

[0178] The target object corresponds to the target mask.

[0179] ​The information processing device provided in the embodiment of the present application determines a target feature map corresponding to a target object according to an image size of the target object and first information, obtains target coordinates of each pixel on the target feature map according to original coordinates of each pixel on a target mask in an input picture and the target feature map, acquires a center coordinate of the target object according to the target coordinates, acquires a target weight corresponding to each pixel according to the target coordinates and the center coordinate, and trains an object detection model according to the target weight corresponding to each pixel, wherein the first information includes a mapping relationship between an object image size and a feature map on a feature pyramid, the target feature map is a feature map of a target layer on the feature pyramid, and the target object corresponds to the target mask. The model training is performed by introducing a mask, so that the shape information of the target object is reserved on the mapped feature map, the supervision signal for the model training is more accurate, and the model precision is improved.

[0180] Further, the first information further includes a reference size of an object image and a mapping reference layer of the feature pyramid, and the determination of the target feature map corresponding to the target object according to the image size of the target object and the first information includes the determination of the target feature map corresponding to the target object according to the image size of the target object, the mapping relationship, the reference size and the mapping reference layer.

[0181] The obtaining of the target coordinates of each pixel on the target feature map according to the original coordinates of each pixel on the target mask in the input picture and the target feature map includes the acquisition of a scaling ratio between the input picture and the target feature map and the obtaining of the target coordinates of each pixel on the target feature map according to the scaling ratio and the original coordinates of each pixel on the input picture.

[0182] In the embodiment of the present application, the acquisition of the target weight corresponding to each pixel according to the target coordinates includes the acquisition of a diagonal line length of an image obtained by mapping each pixel on the target feature map according to the target coordinates, the acquisition of an initial weight corresponding to each pixel according to the diagonal line length, the target coordinates and the center coordinate, and the acquisition of the target weight corresponding to each pixel according to the initial weight.

[0183] Further, the information processing device also includes: a third acquisition module, configured to acquire foreground loss values and background loss values corresponding to each pixel after the each pixel is mapped on the target feature map; a first selection module, configured to select a target pixel from the each pixel according to the foreground loss values; and the training of the object detection model according to the target weights corresponding to the each pixel includes: training of the object detection model according to the background loss values, and the foreground loss values and the target weights corresponding to the target pixel; wherein the foreground loss values include classification loss values and regression loss values; and the background loss values include classification loss values.

[0184] wherein the above Figure 1 The implementation embodiments of the information processing method corresponding to the embodiments are applicable to the embodiments of the information processing device, and can achieve the same technical effects.

[0185] The embodiments of the information processing device also provide an information processing method, as shown in the following. Figure 6 The information processing method includes:

[0186] A fourth acquisition module 61 is configured to acquire foreground loss values, background loss values and target weights corresponding to each pixel of a target object after the each pixel is mapped on a feature map;

[0187] A second selection module 62 is configured to select a target pixel from the each pixel according to the foreground loss values;

[0188] A second training module 63 is configured to train an object detection model according to the background loss values, and the foreground loss values and the target weights corresponding to the target pixel;

[0189] Wherein the foreground loss values include classification loss values and regression loss values; and the background loss values include classification loss values.

[0190] The information processing device provided by the embodiments of the information processing device can acquire foreground loss values, background loss values and target weights corresponding to each pixel of a target object after the each pixel is mapped on a feature map; select a target pixel from the each pixel according to the foreground loss values; and train an object detection model according to the background loss values, and the foreground loss values and the target weights corresponding to the target pixel; wherein the foreground loss values include classification loss values and regression loss values; and the background loss values include classification loss values; so that the target key points and the difficult sample points in the sample can be selected for targeted training, the model can learn the representation of the target object more effectively, and the precision of the obtained model can be improved.

[0191] wherein the above Figure 2 The implementation embodiments of the information processing method corresponding to the embodiments are applicable to the embodiments of the information processing device, and can achieve the same technical effects.

[0192] The embodiment of the present application also provides an information processing device, which comprises a processor 71 and a transceiver 72. Figure 7 As shown in the figure, the processor 71 and the transceiver 72 are connected.

[0193] The processor 71 is configured to determine a target feature map corresponding to a target object according to an image size of the target object and first information.

[0194] According to original coordinates of each pixel on the target feature map and a target mask in an input picture, target coordinates of the each pixel on the target feature map are obtained.

[0195] According to the target coordinates, a center coordinate of the target object is obtained.

[0196] According to the target coordinates and the center coordinate, a target weight corresponding to the each pixel is obtained.

[0197] According to the target weight corresponding to the each pixel, an object detection model is trained.

[0198] The first information comprises a mapping relationship between an object image size and a feature map on a feature pyramid; and the target feature map is a feature map of a target layer on the feature pyramid.

[0199] The target object corresponds to the target mask.

[0200] The transceiver 72 can communicate with the processor 71.

[0201] The information processing device provided by the embodiment of the present application determines a target feature map corresponding to a target object according to an image size of the target object and first information, obtains target coordinates of each pixel on the target feature map according to original coordinates of the each pixel on the target feature map and a target mask in an input picture, obtains a center coordinate of the target object according to the target coordinates, obtains a target weight corresponding to the each pixel according to the target coordinates and the center coordinate, and trains an object detection model according to the target weight corresponding to the each pixel, wherein the first information comprises a mapping relationship between an object image size and a feature map on a feature pyramid; the target feature map is a feature map of a target layer on the feature pyramid; and the target object corresponds to the target mask. The model training is performed by introducing a mask, so that shape information of the target object is reserved on the mapped feature map, the supervision signal for the model training is more accurate, and the model precision is improved.

[0202] Furthermore, the first information also includes: a reference size for the object image and a mapping reference layer for the feature pyramid; determining the target feature map corresponding to the target object based on the image size of the target object and the first information includes: determining the target feature map corresponding to the target object based on the image size of the target object, the mapping relationship, the reference size, and the mapping reference layer.

[0203] The step of obtaining the target coordinates of each pixel on the target feature map based on the original coordinates of each pixel on the target mask on the input image includes: obtaining the scaling ratio between the input image and the target feature map; and obtaining the target coordinates of each pixel on the target feature map based on the scaling ratio and the original coordinates of each pixel on the target mask on the input image.

[0204] In this embodiment of the invention, obtaining the target weight corresponding to each pixel based on the target coordinates and the center coordinates includes: obtaining the diagonal length of the image obtained by mapping each pixel onto the target feature map based on the target coordinates; obtaining the initial weight corresponding to each pixel based on the diagonal length, the target coordinates, and the center coordinates; and obtaining the target weight corresponding to each pixel based on the initial weights.

[0205] Furthermore, the processor is also configured to: obtain the foreground loss value and background loss value corresponding to each pixel after mapping onto the target feature map; select a target pixel from the pixels according to the foreground loss value; the step of training the object detection model according to the target weights corresponding to each pixel includes: training the object detection model according to the background loss value, the foreground loss value corresponding to the target pixel, and the target weights; wherein the foreground loss value includes: a classification loss value and a regression loss value; the background loss value includes: a classification loss value.

[0206] Among them, the above Figure 1 The implementation embodiments of the information processing method described in the corresponding embodiments are all applicable to the embodiments of the information processing device and can achieve the same technical effect.

[0207] This invention also provides an information processing device, such as... Figure 8 As shown, it includes: a processor 81 and a transceiver 82;

[0208] The processor 81 is used to obtain the foreground loss value, background loss value, and target weight corresponding to each pixel of the target object after it is mapped onto the feature map;

[0209] Based on the foreground loss value, a target pixel is selected from each of the pixels;

[0210] training the object detection model according to the background loss value, and the foreground loss value and the target weight corresponding to the target pixel;

[0211] The foreground loss value comprises a classification loss value and a regression loss value, and the background loss value comprises a classification loss value.

[0212] The transceiver 82 can communicate with the processor 81.

[0213] The information processing device provided by the embodiment of the present application can realize the following technical effects. The information processing device provided by the embodiment of the present application can obtain the foreground loss value, the background loss value and the target weight corresponding to each pixel of the target object after the pixel is mapped on the feature map; select a target pixel from the pixels according to the foreground loss value; and train the object detection model according to the background loss value, and the foreground loss value and the target weight corresponding to the target pixel. The foreground loss value comprises a classification loss value and a regression loss value, and the background loss value comprises a classification loss value. The information processing device can realize targeted training of the target key points and the difficult sample points in the sample, so that the model can effectively learn the representation of the target object, thereby improving the precision of the obtained model.

[0214] The implementation embodiments of the information processing method corresponding to the above embodiments are all applicable to the embodiments of the information processing device, and can achieve the same technical effects. Figure 2 The implementation embodiments of the information processing method corresponding to the above embodiments are all applicable to the embodiments of the information processing device, and can achieve the same technical effects.

[0215] The embodiment of the present application further provides an information processing device, which comprises a memory, a processor and a program stored in the memory and executable on the processor; and the processor implements the information processing method (including the information processing method Figure 1 or Figure 2 corresponding to the above embodiments) when executing the program.

[0216] The implementation embodiments of the information processing method corresponding to the above embodiments are all applicable to the embodiments of the information processing device, and can achieve the same technical effects.

[0217] The embodiment of the present application further provides a readable storage medium, which stores a program; and the program is executable on a processor to implement the steps of the information processing method (including the steps of the information processing method Figure 1 or Figure 2 corresponding to the above embodiments) when executed on the processor.

[0218] The implementation embodiments of the information processing method corresponding to the above embodiments are all applicable to the embodiments of the information processing device, and can achieve the same technical effects.

[0219] It should be noted that many of the functional components described in this specification are referred to as modules to more particularly emphasize their implementational independence.

[0220] In embodiments of the present application, modules can be implemented in software for execution by various types of processors. An identified module of executable code may, for instance, comprise one or more physical or logical blocks of computer instructions which may, for instance, be organized as an object, procedure or function. Nevertheless, the executables of an identified module need not be physically located together, but can comprise disparate instructions stored in different locations which, when joined logically together, comprise the module and achieve the stated purpose for the module.

[0221] Indeed, a module of executable code can be a single instruction, or many instructions, and can even be distributed over several different code segments, among different programs, and across several memory devices. Similarly, operational data can be identified and illustrated herein within modules, and can be embodied in any suitable form and organized within any suitable type of data structure. The operational data can be collected as a single data set, or can be distributed over different locations including over different storage devices, and can exist, at least partially, merely as electronic signals on a system or network.

[0222] Where a module can be implemented in software, the software implementation can be stored in any suitable storage medium, for instance a storage disc or other storage medium, as desired. Given the level of detail of material involved, those skilled in the art will appreciate that the module implemented in software can be embodied in machine executable code (e.g., encoded instructions for all or a portion of this specification) that, when executed by a machine (e.g., processor) will cause the machine to perform a process described herein. Accordingly, such machine executable code is also an aspect of the present application. Such machine executable code can be stored on one or more of the storage devices, such as a storage disc, or other storage medium, as desired.

[0223] The preferred embodiments of the present application have been described above with the understanding that variations and modifications will occur to those skilled in the art in view of these teachings. Such variations and modifications are intended to be within the scope of the present application.

Claims

1. An information processing method, characterized in that, include: Based on the image size of the target object and the first information, the target feature map corresponding to the target object is determined; Based on the target feature map and the original coordinates of each pixel on the target mask in the input image, the target coordinates of each pixel on the target feature map are obtained; Based on the target coordinates, obtain the center coordinates of the target object; Based on the target coordinates and center coordinates, obtain the target weight corresponding to each pixel; Train the object detection model based on the target weights corresponding to each pixel; The first information includes: the mapping relationship between the object image size and the feature map on the feature pyramid; the target feature map is the feature map of the target layer on the feature pyramid; The target object corresponds to the target mask; The step of obtaining the target weight corresponding to each pixel based on the target coordinates and center coordinates includes: Based on the target coordinates, obtain the diagonal length of the image obtained by mapping each pixel onto the target feature map; Based on the diagonal length, target coordinates, and center coordinates, obtain the initial weights corresponding to each pixel; Based on the initial weights, the target weights corresponding to each pixel are obtained.

2. The information processing method according to claim 1, characterized in that, The first information also includes: a reference size for the object image and a mapping reference layer for the feature pyramid; The step of determining the target feature map corresponding to the target object based on the image size of the target object and the first information includes: The target feature map corresponding to the target object is determined based on the image size of the target object, the mapping relationship, the reference size, and the mapping reference layer.

3. The information processing method according to claim 1, characterized in that, The step of obtaining the target coordinates of each pixel on the target feature map based on the original coordinates of each pixel on the target mask in the input image includes: Obtain the scaling ratio between the input image and the target feature map; Based on the scaling ratio and the original coordinates of each pixel on the target mask in the input image, the target coordinates of each pixel on the target feature map are obtained.

4. The information processing method according to claim 1, characterized in that, Also includes: Obtain the foreground loss value and background loss value corresponding to each pixel after it is mapped onto the target feature map; Based on the foreground loss value, a target pixel is selected from each of the pixels; The step of training the object detection model based on the target weights corresponding to each pixel includes: The object detection model is trained based on the background loss value, the foreground loss value corresponding to the target pixel, and the target weight. The foreground loss value includes: classification loss value and regression loss value; the background loss value includes: classification loss value.

5. An information processing device, characterized in that, include: The first determining module is used to determine the target feature map corresponding to the target object based on the image size of the target object and the first information; The first processing module is used to obtain the target coordinates of each pixel on the target feature map based on the target feature map and the original coordinates of each pixel on the target mask on the input image; The first acquisition module is used to acquire the center coordinates of the target object based on the target coordinates; The second acquisition module is used to acquire the target weight corresponding to each pixel based on the target coordinates and the center coordinates; The first training module is used to train the object detection model according to the target weights corresponding to each pixel. The first information includes: the mapping relationship between the object image size and the feature map on the feature pyramid; the target feature map is the feature map of the target layer on the feature pyramid; The target object corresponds to the target mask; The step of obtaining the target weight corresponding to each pixel based on the target coordinates and center coordinates includes: Based on the target coordinates, obtain the diagonal length of the image obtained by mapping each pixel onto the target feature map; Based on the diagonal length, target coordinates, and center coordinates, obtain the initial weights corresponding to each pixel; Based on the initial weights, the target weights corresponding to each pixel are obtained.

6. The information processing apparatus according to claim 5, characterized in that, The first information also includes: a reference size for the object image and a mapping reference layer for the feature pyramid; The step of determining the target feature map corresponding to the target object based on the image size of the target object and the first information includes: The target feature map corresponding to the target object is determined based on the image size of the target object, the mapping relationship, the reference size, and the mapping reference layer.

7. The information processing apparatus according to claim 5, characterized in that, The step of obtaining the target coordinates of each pixel on the target feature map based on the original coordinates of each pixel on the target mask in the input image includes: Obtain the scaling ratio between the input image and the target feature map; Based on the scaling ratio and the original coordinates of each pixel on the target mask in the input image, the target coordinates of each pixel on the target feature map are obtained.

8. The information processing apparatus according to claim 5, characterized in that, Also includes: The third acquisition module is used to acquire the foreground loss value and background loss value corresponding to each pixel after it is mapped onto the target feature map; The first selection module is used to select a target pixel from the pixels based on the foreground loss value; The step of training the object detection model based on the target weights corresponding to each pixel includes: The object detection model is trained based on the background loss value, the foreground loss value corresponding to the target pixel, and the target weight. The foreground loss value includes: classification loss value and regression loss value; the background loss value includes: classification loss value.

9. An information processing device, characterized in that, include: Processor and transceiver; The processor is configured to determine the target feature map corresponding to the target object based on the image size of the target object and the first information; Based on the target feature map and the original coordinates of each pixel on the target mask in the input image, the target coordinates of each pixel on the target feature map are obtained; Based on the target coordinates, obtain the center coordinates of the target object; Based on the target coordinates and center coordinates, obtain the target weight corresponding to each pixel; Train the object detection model based on the target weights corresponding to each pixel; The first information includes: the mapping relationship between the object image size and the feature map on the feature pyramid; the target feature map is the feature map of the target layer on the feature pyramid; The target object corresponds to the target mask; The step of obtaining the target weight corresponding to each pixel based on the target coordinates and center coordinates includes: Based on the target coordinates, obtain the diagonal length of the image obtained by mapping each pixel onto the target feature map; Based on the diagonal length, target coordinates, and center coordinates, obtain the initial weights corresponding to each pixel; Based on the initial weights, the target weights corresponding to each pixel are obtained.

10. An information processing device, comprising a memory, a processor, and a program stored in the memory and executable on the processor; characterized in that, When the processor executes the program, it implements the information processing method as described in any one of claims 1 to 4.

11. A readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the information processing method as described in any one of claims 1 to 4.

Images