A target detection method based on high-efficiency pole model

Abstract

The application discloses a target detection method based on a high-efficiency pole model, which comprises the following steps: S1, detecting the resolution of an image to be detected, and supplementing the image to be detected into a square image by adding a gray pixel frame on the upper and lower or left and right of the image to be detected; S2, extending or compressing the square image, so that the square image is matched with the resolution in the high-efficiency pole model, and a matched image is obtained; S3, predicting the matched image through the high-efficiency pole model, so that a square image containing a target detection frame and a target type is obtained; and S4, restoring the resolution of the square image containing the target detection frame and the target type to the resolution of the image to be detected. The application solves the problems of a large background area range in a detection frame, low target detection precision and waste of computing resources which are commonly existed in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image recognition, and particularly relates to a target detection method based on a high-efficiency pole model. BACKGROUND

[0002] Target detection is a common and key visual recognition problem in computer vision, which has been extensively studied in the past few decades. The purpose of visual target detection is to find objects with precise positioning and specific target categories in a given image, and to assign a corresponding category label to each object instance. In recent years, scientists around the world have conducted extensive and in-depth research on deep learning algorithms, and the research results of applying them to image classification, target detection and other fields are increasing, showing a good development prospect.

[0003] In the detection process, there may be multiple targets in the same image, and the same target has different appearances, shapes and postures. In addition, users have different requirements for detection algorithms and time complexity, so target detection has always been one of the most challenging problems in the field of computer vision. The continuous optimization of deep learning makes the target detection model more accurate and shorter in detection time. For example: the EfficientDet (high efficiency) detection model with a complexity of 326B Flops can achieve an average detection accuracy (mAP) of 51% on the COCO dataset.

[0004] Researchers at home and abroad have carried out a lot of work in the research of target detection and have made great research progress, but there are still the following problems:

[0005] First: the regular rectangular target detection frame contains less target information. Traditional first-order target detection converts the target detection problem into a rectangular region classification problem by setting anchor frames of different sizes. Since most objects are not axisymmetric and square targets, the rectangular bounding box is not the optimal representation of the natural target. The rectangular detection frame lists a large number of possible frame positions through anchor points, without truly understanding the object's own combination visual grammar, so it often contains meaningless background pixels.

[0006] Second: the average detection accuracy is low when the calculation complexity is low. When the calculation complexity is low, the target feature extraction model has fewer iterations, and the anchor point of the target feature is greatly different from the actual target position. As can be seen from the formula for calculating the average detection accuracy, the smaller the overlap range between the predicted frame and the real frame, the lower the target detection accuracy.

[0007] Thirdly, the same detection algorithm is used for input images with different resolutions. For high-resolution images, a low-complexity model detects by compressing image quality, which is prone to missed detection. For low-resolution images, a high-complexity model has precision saturation, resulting in waste of computing resources. The existing detection model cannot use different detection algorithms for input images with different resolutions. SUMMARY

[0008] In view of the above problems in the prior art, the target detection method based on the high-efficiency pole model provided by the present application solves the problems of large background area in the detection frame, low target detection accuracy and waste of computing resources in the existing target detection technology.

[0009] To achieve the above-mentioned purposes, the technical scheme adopted by the present application is as follows: a target detection method based on a high-efficiency pole model, comprising the following steps:

[0010] S1, detecting the resolution of the image to be detected, and supplementing it into a square image by adding a gray pixel frame above, below, left or right;

[0011] S2, extending or compressing the square image to match the resolution in the high-efficiency pole model, to obtain a matching image;

[0012] S3, predicting the matching image by the high-efficiency pole model to obtain a square image containing a target detection frame and a target type;

[0013] S4, restoring the resolution of the square image containing the target detection frame and the target type to the resolution of the image to be detected.

[0014] Further, the high-efficiency pole model in step S2 comprises: an EffD0+ExtNet unit, an EffD1+ExtNet unit, an EffD2+ExtNet unit, an EffD3+ExtNet unit, an EffD4 unit, an EffD5 unit, an EffD6 unit and an EffD7 unit.

[0015] Further, step S2 includes the following cases:

[0016] S21, if the edge pixels of the square image are less than 512, the edge pixels of the square image are all expanded to 512 to obtain a 512*512 matching image;

[0017] S22, if the edge pixels of the square image are greater than or equal to 512 and less than 640, the edge pixels of the square image are all expanded to 640 to obtain a 640*640 matching image;

[0018] S23, if the edge pixel of the square image is greater than or equal to 640 and less than 768, the edge pixel of the square image is expanded to 768, and a matching image of 768*768 is obtained;

[0019] S24, if the edge pixel of the square image is greater than or equal to 768 and less than 896, the edge pixel of the square image is expanded to 896, and a matching image of 896*896 is obtained;

[0020] S25, if the edge pixel of the square image is greater than or equal to 896 and less than 1024, the edge pixel of the square image is expanded to 1024, and a matching image of 1024*1024 is obtained;

[0021] S26, if the edge pixel of the square image is greater than or equal to 1024 and less than 1280, the edge pixel of the square image is expanded to 1280, and a matching image of 1280*1280 is obtained;

[0022] S27, if the edge pixel of the square image is greater than or equal to 1280 and less than 1408, the edge pixel of the square image is expanded to 1408, and a matching image of 1408*1408 is obtained;

[0023] S28, if the edge pixel of the square image is greater than 1408, the edge pixel of the square image is expanded or compressed to 1536, and a matching image of 1536*1536 is obtained.

[0024] Further, the step S3 includes the following cases:

[0025] S31, the EffD0+ExtNet unit is used to predict the matching image of 512*512, and a square image containing a target detection frame and a target type is obtained;

[0026] S32, the EffD1+ExtNet unit is used to predict the matching image of 640*640, and a square image containing a target detection frame and a target type is obtained;

[0027] S33, the EffD2+ExtNet unit is used to predict the matching image of 768*768, and a square image containing a target detection frame and a target type is obtained;

[0028] S34, the EffD3+ExtNet unit is used to predict the matching image of 896*896, and a square image containing a target detection frame and a target type is obtained;

[0029] S35, the EffD4 unit is used to predict the matching image of 1024*1024, and a square image containing a target detection frame and a target type is obtained;

[0030] S36, using the EffD5 unit to predict the matching image of 1280*1280 to obtain a square image containing a target detection frame and a target type;

[0031] S37, using the EffD6 unit to predict the matching image of 1408*1408 to obtain a square image containing a target detection frame and a target type;

[0032] S38, using the EffD7 unit to predict the matching image of 1536*1536 to obtain a square image containing a target detection frame and a target type.

[0033] Further, the structures of the EffD0+ExtNet unit, the EffD1+ExtNet unit, the EffD2+ExtNet unit and the EffD3+ExtNet unit each include: an EfficientNet subunit, a feature pyramid subunit, a class prediction subunit, a pole prediction subunit and a pole processing subunit;

[0034] The input end of the EfficientNet subunit serves as the input end of the EffD0+ExtNet unit, the EffD1+ExtNet unit, the EffD2+ExtNet unit or the EffD3+ExtNet unit;

[0035] The output end of the EfficientNet subunit is connected with the input end of the feature pyramid subunit; the output end of the feature pyramid subunit is respectively connected with the input end of the class prediction subunit and the input end of the pole prediction subunit;

[0036] The output end of the class prediction subunit serves as the output end of the predicted target type of the EffD0+ExtNet unit, the EffD1+ExtNet unit, the EffD2+ExtNet unit or the EffD3+ExtNet unit;

[0037] The input end of the pole processing subunit is connected with the output end of the pole prediction subunit, and the output end thereof serves as the output end of the irregular target detection frame of the EffD0+ExtNet unit, the EffD1+ExtNet unit, the EffD2+ExtNet unit or the EffD3+ExtNet unit.

[0038] Further, the pole processing subunit is used for processing the target center point and the pole output by the pole prediction subunit, so as to obtain an irregular target detection frame, and the specific method process is as follows:

[0039] A1, randomly selecting four poles output by the pole prediction subunit to form a group, to obtain an initial pole group;

[0040] A2, calculate the geometric center of each group of pole points;

[0041] A3, calculate the distance between the geometric center and the target center point;

[0042] A4, discard the pole point group with a distance greater than 3 pixel units, and obtain the pole point group with a distance less than or equal to 3 pixel units;

[0043] A5, calculate the pixel area covered by the pole point group with a distance less than or equal to 3 pixel units, and the real area in the corresponding label image;

[0044] A6, calculate the overlap degree of the pixel area and the real area;

[0045] A7, discard the pole point group with an overlap degree less than 75% in the remaining pole point group of step A4, and obtain the pole point group with an overlap degree greater than or equal to 75%, that is, the irregular target detection frame.

[0046] Further, the pole point prediction subunit comprises a first depth separable convolution layer conv1 and an hourglass network connected in sequence.

[0047] The beneficial effects of the above further scheme are: in the pole point prediction subunit, the feature channel number input into the hourglass network is adjusted through the first depth separable convolution layer conv1, and the purpose of the hourglass network is to regress the input features into the target center point and the upper and lower and left and right pole points.

[0048] The combination of the EfficientDet model and the pole point processing replaces the rectangular target detection frame with the irregular polygon target detection frame under the condition of ensuring the detection efficiency, outputs the semantic of the target object itself, and has a higher overlap degree between the target detection frame and the target real frame, thereby improving the model detection precision.

[0049] Further, the category prediction subunit comprises a second depth separable convolution layer conv2 and a third depth separable convolution layer conv3 connected in sequence.

[0050] In summary, the beneficial effects of the present application are:

[0051] 1. When detecting a target, a regular rectangular frame will cause a large background area in the detection frame, especially when the center of gravity of the target object is outside the object, which makes it difficult to select a rectangular detection frame; at the same time, the anchor frame cannot understand the visual syntax of the object itself whether it is labeled or predicted. Therefore, the present application replaces the rectangular target detection frame with the irregular target detection frame, which can reduce the range of the background area in the detection frame, and the output position frame contains the posture and shape of the object itself.

[0052] 2、In the condition of low anchor point precision caused by low calculation complexity, the detection frame obtained according to the traditional anchor point regression is easy to be different from the real target position, so that the target detection precision is low. Therefore, the method for extracting the pole point and the target object center point is adopted to replace the method for extracting the anchor point, so as to improve the detection efficiency; meanwhile, the extracted pole point and the center point are compared, and a series of algorithms are used to obtain the irregular target detection frame, and the irregular target detection frame has a higher overlapping degree with the real frame.

[0053] 3、The traditional target detection adopts a single algorithm to adapt to different resolution input images, and a complex model is used for a low resolution image, and a compressed mode is used for a high resolution image, so that the precision loss or calculation saturation condition is easy to be caused. Therefore, the resolution is classified before the to-be-detected image is transmitted into the model for calculation, and then the to-be-detected image is transmitted into the corresponding model, so that the detection efficiency in the condition of low calculation complexity can be ensured, and the image resolution in the condition of high calculation complexity is not lost, so that the problem of waste of calculation resources is solved. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 It is a flowchart of a target detection method based on a high-efficiency pole point model.

[0055] Figure 2 It is a structural schematic view of an EffD0+ExtNet unit, an EffD1+ExtNet unit, an EffD2+ExtNet unit and an EffD3+ExtNet unit. DETAILED DESCRIPTION

[0056] The specific embodiments of the present application are described below to facilitate the understanding of the present application by those skilled in the art, but it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, it is obvious that various changes are within the spirit and scope of the present application defined and limited by the appended claims, and all the application and creation utilizing the concept of the present application are within the scope of protection.

[0057] As shown in the figure, a target detection method based on a high-efficiency pole point model comprises the following steps: Figure 1

[0058] S1, the resolution of the to-be-detected image is detected, and the to-be-detected image is supplemented into a square image by adding a gray pixel frame on the top and bottom or left and right thereof;

[0059] When the width of the to-be-detected image is greater than the height, the height of the to-be-detected image should be expanded by adding a frame on the top and bottom, and the to-be-detected image is supplemented into a square image; when the height of the to-be-detected image is greater than the width, the width of the to-be-detected image should be expanded on the left and right, and the to-be-detected image is supplemented into a square image.

[0060] ​In step S1, the added gray pixel box has pixels of (128, 128, 128), because the gray pixel (128, 128, 128) has the least influence on the detection target among the (0, 0, 0) to (255, 255, 255) random pixels.

[0061] S2, expand or compress the square image to match the resolution in the high-efficiency pole model to obtain a matching image.

[0062] The high-efficiency pole model includes: EffD0+ExtNet unit, EffD1+ExtNet unit, EffD2+ExtNet unit, EffD3+ExtNet unit, EffD4 unit, EffD5 unit, EffD6 unit and EffD7 unit.

[0063] Step S2 includes the following cases:

[0064] S21, if the edge pixels of the square image are less than 512, then expand the edge pixels of the square image to 512 to obtain a 512*512 matching image;

[0065] S22, if the edge pixels of the square image are greater than or equal to 512 and less than 640, then expand the edge pixels of the square image to 640 to obtain a 640*640 matching image;

[0066] S23, if the edge pixels of the square image are greater than or equal to 640 and less than 768, then expand the edge pixels of the square image to 768 to obtain a 768*768 matching image;

[0067] S24, if the edge pixels of the square image are greater than or equal to 768 and less than 896, then expand the edge pixels of the square image to 896 to obtain a 896*896 matching image;

[0068] S25, if the edge pixels of the square image are greater than or equal to 896 and less than 1024, then expand the edge pixels of the square image to 1024 to obtain a 1024*1024 matching image;

[0069] S26, if the edge pixels of the square image are greater than or equal to 1024 and less than 1280, then expand the edge pixels of the square image to 1280 to obtain a 1280*1280 matching image;

[0070] S27, if the edge pixels of the square image are greater than or equal to 1280 and less than 1408, then expand the edge pixels of the square image to 1408 to obtain a 1408*1408 matching image;

[0071] S28, if the edge pixel of the square image is greater than 1408, then the edge pixel of the square image is expanded or compressed to 1536, obtaining a matching image of 1536*1536.

[0072] S3, predicting the matching image by using the high-efficiency pole model to obtain a square image containing a target detection frame and a target type.

[0073] Step S3 includes the following cases:

[0074] S31, predicting the matching image of 512*512 by using the EffD0+ExtNet unit to obtain a square image containing a target detection frame and a target type.

[0075] S32, predicting the matching image of 640*640 by using the EffD1+ExtNet unit to obtain a square image containing a target detection frame and a target type.

[0076] S33, predicting the matching image of 768*768 by using the EffD2+ExtNet unit to obtain a square image containing a target detection frame and a target type.

[0077] S34, predicting the matching image of 896*896 by using the EffD3+ExtNet unit to obtain a square image containing a target detection frame and a target type.

[0078] S35, predicting the matching image of 1024*1024 by using the EffD4 unit to obtain a square image containing a target detection frame and a target type.

[0079] S36, predicting the matching image of 1280*1280 by using the EffD5 unit to obtain a square image containing a target detection frame and a target type.

[0080] S37, predicting the matching image of 1408*1408 by using the EffD6 unit to obtain a square image containing a target detection frame and a target type.

[0081] S38, predicting the matching image of 1536*1536 by using the EffD7 unit to obtain a square image containing a target detection frame and a target type.

[0082] S4, restoring the resolution of the square image containing the target detection frame and the target type to the resolution of the image to be detected.

[0083] The high-efficiency pole model in step S3 is trained, and the training process of the high-efficiency pole model is as follows:

[0084] The first step is to convert the anchor box into a pole box according to the labels corresponding to all images in the VOC2007 dataset obtained;

[0085] The second step is to divide 16552 images and their labels in the dataset into a training set and a test set according to a 9:1 ratio, wherein the training set has 14896 images and the test set has 1656 images;

[0086] The third step is to add gray pixels to the top and bottom (or left and right) of the image to make it a square image;

[0087] The fourth step is to convert all images into a high-efficiency pole model corresponding resolution, wherein the EffD0+ExtNet unit, the EffD1+ExtNet unit, the EffD2+ExtNet unit, the EffD3+ExtNet unit, the EffD4 unit, the EffD5 unit, the EffD6 unit and the EffD7 unit correspond to 512*512, 640*640, 768*768, 896*896, 1024*1024, 1280*1280, 1408*1408 and 1536*1536 resolutions respectively. Correspondingly, the target pole coordinates in the label also change with the image compression;

[0088] The fifth step is to input the image preprocessed in the fourth step into the high-efficiency pole model for training.

[0089] Each of the EffD0+ExtNet unit, the EffD1+ExtNet unit, the EffD2+ExtNet unit, the EffD3+ExtNet unit, the EffD4 unit, the EffD5 unit, the EffD6 unit and the EffD7 unit compresses 16552 images to the corresponding resolution to ensure the sufficiency of the training samples.

[0090] As shown in Figure 2 The structure of the EffD0+ExtNet unit, the EffD1+ExtNet unit, the EffD2+ExtNet unit and the EffD3+ExtNet unit includes an EfficientNet subunit, a feature pyramid subunit, a class prediction subunit, a pole prediction subunit and a pole processing subunit;

[0091] The input end of the EfficientNet subunit serves as the input end of the EffD0+ExtNet unit, the EffD1+ExtNet unit, the EffD2+ExtNet unit or the EffD3+ExtNet unit;

[0092] The output end of the EfficientNet subunit is connected with the input end of the feature pyramid subunit; the output end of the feature pyramid subunit is connected with the input end of the category prediction subunit and the input end of the pole prediction subunit respectively;

[0093] The output end of the category prediction subunit is the output end of the predicted target type of the EffD0+ExtNet unit, the EffD1+ExtNet unit, the EffD2+ExtNet unit or the EffD3+ExtNet unit.

[0094] The input end of the pole processing subunit is connected with the output end of the pole prediction subunit, and the output end thereof is the output end of the irregular target detection frame of the EffD0+ExtNet unit, the EffD1+ExtNet unit, the EffD2+ExtNet unit or the EffD3+ExtNet unit.

[0095] The EffD4 unit, the EffD5 unit, the EffD6 unit and the EffD7 unit are EfficientDet model D4 to D7.

[0096] The pole processing subunit is used for processing the target center point and the pole output by the pole prediction subunit, so as to obtain the irregular target detection frame, and the specific method process is as follows:

[0097] A1, randomly selecting four poles output by the pole prediction subunit to form a group, to obtain an initial pole group;

[0098] A2, calculating the geometric center of each pole group;

[0099] A3, calculating the distance between the geometric center and the target center point;

[0100] A4, discarding the pole group with a distance greater than 3 pixel units, to obtain a pole group with a distance less than or equal to 3 pixel units;

[0101] A5, calculating the pixel area covered by the pole group with a distance less than or equal to 3 pixel units, and the real area in the corresponding label image;

[0102] A6, calculating the overlap degree of the pixel area and the real area;

[0103] A7, discarding the pole group with an overlap degree less than 75% in the remaining pole group of step A4, to obtain a pole group with an overlap degree greater than or equal to 75%, which is the irregular target detection frame.

[0104] The pole prediction subunit comprises a first depth separable convolution layer conv1 and an hourglass network connected in sequence.

[0105] The category prediction subunit comprises a second deep separable convolution layer conv2 and a third deep separable convolution layer conv3 connected in sequence.

[0106] In step S3, if the matching image is processed by the EffD0+ExtNet unit, the EffD1+ExtNet unit, the EffD2+ExtNet unit or the EffD3+ExtNet unit, the image is restored through step S4, the target is framed by an irregular target detection frame on the restored image, and the predicted target type is labeled. If the matching image is processed by the EffD4 unit, the EffD5 unit, the EffD6 unit and the EffD7 unit, the image is restored through step S4, the target is framed by a rectangular target detection frame on the restored image, and the predicted target type is labeled.

[0107] In step S3, different sizes of images are adapted to different models, for example: the image of 636*450 pixels is converted into an image of 636*636 pixels by adding a gray pixel frame, is upward compatible, and the EffD1+ExtNet unit with an input of 640*640 pixels is selected; when the picture resolution exceeds the EffD6 unit (1408*1408 pixels), the image is uniformly compressed (or expanded) to 1536*1536 pixels, and the EffD7 unit is used for detection. This method not only ensures the sufficiency of detection, but also reduces the waste of computing resources.

[0108] In step S4, the square image containing the target detection frame and the target type is restored, which includes two parts: compression and de-gray frame, which correspond to expansion and gray frame respectively.

[0109] For high-resolution images, the target is framed by a rectangular detection frame, and for low-resolution images, the target is framed by an irregular detection frame.

Claims

1. A target detection method based on a high-efficiency pole model, characterized in that, Includes the following steps: S1. Detect the resolution of the image to be detected, and fill it into a square image by adding gray pixel boxes above and below or to the left and right of it. S2. Expand or compress the square image to match the resolution in the high-efficiency pole model to obtain a matching image. The high-efficiency pole model in step S2 includes: EffD0+ExtNet unit, EffD1+ExtNet unit, EffD2+ExtNet unit, EffD3+ExtNet unit, EffD4 unit, EffD5 unit, EffD6 unit and EffD7 unit; The structures of the EffD0+ExtNet unit, EffD1+ExtNet unit, EffD2+ExtNet unit, and EffD3+ExtNet unit all include: an EfficientNet subunit, a feature pyramid subunit, a category prediction subunit, an extreme point prediction subunit, and an extreme point processing subunit. The input terminal of the EfficientNet sub-unit serves as the input terminal of the EffD0+ExtNet unit, EffD1+ExtNet unit, EffD2+ExtNet unit, or EffD3+ExtNet unit. The output of the EfficientNet subunit is connected to the input of the feature pyramid subunit; the output of the feature pyramid subunit is connected to the input of the category prediction subunit and the input of the pole prediction subunit, respectively. The output of the category prediction subunit serves as the output of the predicted target type for the EffD0+ExtNet unit, EffD1+ExtNet unit, EffD2+ExtNet unit, or EffD3+ExtNet unit. The input of the pole processing subunit is connected to the output of the pole prediction subunit, and its output serves as the output of the irregular target detection box of the EffD0+ExtNet unit, EffD1+ExtNet unit, EffD2+ExtNet unit or EffD3+ExtNet unit. The pole processing subunit is used to process the target center point and poles output by the pole prediction subunit, thereby obtaining the irregular target detection box. The specific method process is as follows: A1. Randomly select four poles from the output of the pole prediction subunit to form a group, thus obtaining the initial pole group; A2. Calculate the geometric center of each set of poles; A3. Calculate the distance between the geometric center and the target center point; A4. Discard the pole groups with a distance greater than 3 pixels and obtain the pole groups with a distance less than or equal to 3 pixels. A5. Calculate the pixel area covered by the pole group with a distance of less than or equal to 3 pixel units, and the actual area in the corresponding label image; A6. Calculate the degree of overlap between the pixel area and the actual area; A7. In the remaining pole groups in step A4, discard the pole groups with an overlap of less than 75% to obtain the pole groups with an overlap of greater than or equal to 75%, which are the irregular target detection boxes. S3. The matching image is predicted using a high-efficiency pole model to obtain a square image containing the target detection box and the target type; S4. Restore the resolution of the square image containing the target detection box and target type to the resolution of the image to be detected.

2. The target detection method based on the high-efficiency pole model according to claim 1, characterized in that, Step S2 includes the following situations: S21. If the edge pixels of the square image are less than 512, then the edge pixels of the square image are all expanded to 512 to obtain a 512*512 matching image. S22. If the edge pixels of the square image are greater than or equal to 512 and less than 640, then the edge pixels of the square image are all expanded to 640 to obtain a matching image of 640*640. S23. If the edge pixels of the square image are greater than or equal to 640 and less than 768, then the edge pixels of the square image are all expanded to 768 to obtain a matching image of 768*768. S24. If the edge pixels of the square image are greater than or equal to 768 and less than 896, then the edge pixels of the square image are all expanded to 896 to obtain a matching image of 896*896. S25. If the edge pixels of the square image are greater than or equal to 896 and less than 1024, then the edge pixels of the square image are all expanded to 1024 to obtain a 1024*1024 matching image. S26. If the edge pixels of the square image are greater than or equal to 1024 and less than 1280, then the edge pixels of the square image are all expanded to 1280 to obtain a matching image of 1280*1280. S27. If the edge pixels of the square image are greater than or equal to 1280 and less than 1408, then the edge pixels of the square image are all expanded to 1408 to obtain a matching image of 1408*1408. S28. If the edge pixels of the square image are greater than 1408, then the edge pixels of the square image are expanded or compressed to 1536 to obtain a matching image of 1536*1536.

3. The target detection method based on the high-efficiency pole model according to claim 2, characterized in that, Step S3 includes the following situations: S31. Use EffD0+ExtNet units to predict the 512*512 matching image to obtain a square image containing the target detection box and the target type; S32. Use EffD1+ExtNet units to predict the 640*640 matching image to obtain a square image containing the target detection box and the target type; S33. Use EffD2+ExtNet units to predict the 768*768 matching image to obtain a square image containing the target detection box and the target type; S34. Use EffD3+ExtNet units to predict the 896*896 matching image to obtain a square image containing the target detection box and the target type; S35. The EffD4 unit is used to predict the 1024*1024 matching image to obtain a square image containing the target detection box and the target type; S36. The EffD5 unit is used to predict the 1280*1280 matching image to obtain a square image containing the target detection box and the target type; S37. Use the EffD6 unit to predict the 1408*1408 matching image to obtain a square image containing the target detection box and the target type; S38. The EffD7 unit is used to predict the 1536*1536 matching image to obtain a square image containing the target detection box and the target type.

4. The target detection method based on the high-efficiency pole model according to claim 1, characterized in that, The pole prediction subunit comprises a first depthwise separable convolutional layer conv1 and an hourglass network connected in sequence.

5. The target detection method based on the high-efficiency pole model according to claim 1, characterized in that, The category prediction subunit includes a second depthwise separable convolutional layer conv2 and a third depthwise separable convolutional layer conv3 connected in sequence.

Images