Target detection method and device based on multi-scale reverse bottle-neck structure

Abstract

The scheme relates to a target detection method and device based on a multi-scale reverse bottle neck structure, a computer device and a storage medium. The method comprises: obtaining an original image and performing down-sampling processing to obtain a down-sampled image; performing a reverse bottle neck operation on the down-sampled image and generating a feature pyramid network according to a backbone network; obtaining different scale feature maps through the feature pyramid network, and performing multi-level prediction on different size targets on different scale feature maps by using anchor-free pixel-level detection; and performing multi-level prediction and post-processing on the original image after model training to obtain a target detection frame. Through the combination of the multi-scale reverse bottle neck structure and the anchor-free detection, multi-scale detection of the image is realized, information loss is avoided, and information effectiveness is improved, so as to solve the difficult problem of target detection caused by image noise and image distortion in the current target detection field and improve the effect of target detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning target detection, in particular to a target detection method and device based on a multi-scale reverse bottle neck structure, a computer device and a storage medium. BACKGROUND

[0002] With the rapid development of artificial intelligence, target detection technology based on deep learning has gradually become the mainstream technology method in the field of image target detection. The existing target detection method has the shortcomings of inaccurate recognition or inability to recognize in images with distortion, serious noise and low resolution. Especially in the medical field, ultrasonic images have the advantages of real-time, low cost, non-invasiveness and easy operation, and are the most popular medical imaging technology, widely used in disease analysis and clinical diagnosis, and are the preferred method for early detection and diagnosis of thyroid nodules. As a widely used medical imaging technology, ultrasonic technology can also be used for computer-aided diagnosis. However, ultrasonic images have the shortcomings of serious noise, poor imaging quality and low image resolution, and ultrasonic imaging is highly dependent on the technical level of doctors, and the feature description and diagnosis report lack a unified standardized process and specification. These defects result in doctors being unable to accurately and quickly manually distinguish lesions, bringing great challenges to the field of medical image recognition.

[0003] Therefore, in the traditional target detection process, information is easily lost, resulting in image distortion, serious noise and low resolution. That is, the traditional target detection method has the problem of poor detection effect. SUMMARY

[0004] Therefore, in order to solve the above technical problems, a target detection method and device based on a multi-scale reverse bottle neck structure, a computer device and a storage medium are provided, which can improve the effect of target detection.

[0005] A target detection method based on a multi-scale reverse bottle neck structure, the method comprising:

[0006] An original image is obtained, and the original image is down-sampled to obtain a down-sampled image;

[0007] The down-sampled image is input into a multi-scale reverse bottle neck structure, multi-feature expansion and fusion reverse bottle neck operation are performed on the down-sampled image, and a feature pyramid network is generated according to a backbone network;

[0008] Different scale feature maps are obtained through the feature pyramid network, and multi-level prediction is performed on the different scale feature maps for targets of different sizes using anchor-free pixel-level detection, to obtain multi-level prediction results;

[0009] Model training is performed on the multi-level prediction result, and multi-level prediction and post-processing are performed on the original image through the trained model to obtain a target detection frame.

[0010] In one of the embodiments, the downsampling processing on the original image to obtain a downsampled image comprises:

[0011] The original image is stretched to obtain a stretched image.

[0012] The stretched image is subjected to a non-repetitive covering two-dimensional patch convolution to obtain a downsampled image.

[0013] In one of the embodiments, the multi-feature expansion and fusion reverse bottleneck operation on the downsampled image and the generation of a feature pyramid network according to a backbone network comprise:

[0014] The downsampled image is subjected to a large-core spatial information convolution, and after layer normalization, an image feature under different receptive fields is obtained through a hole convolution.

[0015] The image feature is spliced to obtain a high-dimensional feature vector.

[0016] A low-dimensional feature vector is obtained according to the high-dimensional feature vector and the downsampled image, and an input of a next layer reverse bottleneck structure is obtained after the downsampled image is subjected to a downsampling operation by using an independent inverse sub-pixel convolution, and a large-core spatial information convolution is performed again until a backbone feature map is generated.

[0017] In one of the embodiments, the multi-scale reverse bottleneck structure comprises a plurality of reverse bottleneck blocks.

[0018] The first layer of the reverse bottleneck block is the large-core spatial information convolution, the second layer is a multi-class fusion convolution, and the third layer is a channel convolution multiplied by an input image probability.

[0019] Each layer of the backbone network is provided with an independent inverse sub-pixel convolution layer as a downsampling layer.

[0020] In one of the embodiments, the generation of a feature pyramid network according to a backbone network comprises:

[0021] A feature map with high semantic information is obtained through the inverse sub-pixel convolution of the backbone network.

[0022] The layers of the backbone network are upsampled and linearly processed to obtain a feature map with low semantic information.

[0023] A feature pyramid network is generated according to the feature map with high semantic information and the feature map with low semantic information.

[0024] In one of the embodiments, the pixel-level detection without anchor box is used to perform multi-level prediction on the different scale feature maps for targets of different sizes, including:

[0025] Each pixel point in the different scale feature maps is enlarged proportionally, corresponding to the center point of the corresponding block on the original image.

[0026] The prediction points and the prediction boxes on the different scale feature maps are mapped on the original image.

[0027] In one of the embodiments, the multi-level prediction results are trained by a model, and the original image is predicted by the trained model and post-processed to obtain a target detection box, including:

[0028] The multi-level prediction results are trained by a model using distant-ness and classification loss and regression loss.

[0029] The multi-level prediction results are classified, regressed, and threshold judged by a center score by a target parameter, and post-processed by non-maximum suppression to obtain a target detection box.

[0030] A target detection device based on a multi-scale reverse bottle neck structure, the device comprising:

[0031] A down-sampling processing module is configured to acquire an original image and perform down-sampling processing on the original image to obtain a down-sampled image.

[0032] A reverse bottle neck operation module is configured to input the down-sampled image into a multi-scale reverse bottle neck structure, perform multi-feature expansion and fusion reverse bottle neck operation on the down-sampled image, and generate a feature pyramid network according to a backbone network.

[0033] A prediction module is configured to obtain different scale feature maps through the feature pyramid network, and perform multi-level prediction on the different scale feature maps for targets of different sizes by pixel-level detection without anchor box to obtain multi-level prediction results.

[0034] A detection box acquisition module is configured to train the multi-level prediction results by a model, and perform multi-level prediction on the original image by the trained model and post-process to obtain a target detection box.

[0035] A computer device comprising a memory and a processor, the memory storing a computer program, and the processor implementing the following steps when executing the computer program:

[0036] An original image is acquired and down-sampled to obtain a down-sampled image.

[0037] input the down-sampling image into a multi-scale reverse bottle-neck structure, perform multi-feature expansion and fusion reverse bottle-neck operation on the down-sampling image, and generate a feature pyramid network according to a backbone network;

[0038] obtain different scale feature maps through the feature pyramid network, and perform multi-level prediction on different size targets on the different scale feature maps by using anchor-free pixel-level detection to obtain multi-level prediction results;

[0039] perform model training on the multi-level prediction results, and perform multi-level prediction and post-processing on the original image through the trained model to obtain a target detection frame.

[0040] A computer readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the following steps:

[0041] obtain an original image, and perform down-sampling processing on the original image to obtain a down-sampling image;

[0042] input the down-sampling image into a multi-scale reverse bottle-neck structure, perform multi-feature expansion and fusion reverse bottle-neck operation on the down-sampling image, and generate a feature pyramid network according to a backbone network;

[0043] obtain different scale feature maps through the feature pyramid network, and perform multi-level prediction on different size targets on the different scale feature maps by using anchor-free pixel-level detection to obtain multi-level prediction results;

[0044] perform model training on the multi-level prediction results, and perform multi-level prediction and post-processing on the original image through the trained model to obtain a target detection frame.

[0045] The target detection method, device, computer equipment and storage medium based on the multi-scale reverse bottle-neck structure, by obtaining an original image, and performing down-sampling processing on the original image to obtain a down-sampling image; inputting the down-sampling image into a multi-scale reverse bottle-neck structure, performing multi-feature expansion and fusion reverse bottle-neck operation on the down-sampling image, and generating a feature pyramid network according to a backbone network; obtaining different scale feature maps through the feature pyramid network, and performing multi-level prediction on different size targets on the different scale feature maps by using anchor-free pixel-level detection to obtain multi-level prediction results; performing model training on the multi-level prediction results, and performing multi-level prediction and post-processing on the original image through the trained model to obtain a target detection frame. Through the combination of multi-scale reverse bottle-neck structure and anchor-free detection, multi-scale detection of the image is realized, which can avoid information loss and improve information effectiveness, so as to solve the problem of target detection difficulty caused by image noise and image distortion in the current target detection field, and improve the effect of target detection. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 An application environment diagram of a target detection method based on a multi-scale reverse bottle-neck structure in an embodiment;

[0047] Figure 2 A flowchart of a target detection method based on a multi-scale reverse bottle-neck structure in an embodiment;

[0048] Figure 3 A schematic diagram of a multi-scale reverse bottle-neck structure block in an embodiment;

[0049] Figure 4 A schematic diagram of a feature network structure in an embodiment;

[0050] Figure 5 A flowchart of a post-processing of target detection in an embodiment;

[0051] Figure 6 A structure block diagram of a target detection device based on a multi-scale reverse bottle-neck structure in an embodiment;

[0052] Figure 7 An internal structure diagram of a computer device in an embodiment. DETAILED DESCRIPTION

[0053] In order to make the objects, technical solutions and advantages of the present application clearer, further detailed description will be made to the present application in combination with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0054] The target detection method based on a multi-scale reverse bottle-neck structure provided by the embodiments of the present application can be applied in an application environment as shown in Figure 1 As shown in Figure 1 The application environment includes a computer device 110. The computer device 110 can obtain an original image, and perform down-sampling processing on the original image to obtain a down-sampled image. The computer device 110 can input the down-sampled image into a multi-scale reverse bottle-neck structure, perform multi-feature expansion and fusion reverse bottle-neck operation on the down-sampled image, and generate a feature pyramid network according to a backbone network. The computer device 110 can obtain different scale feature maps through the feature pyramid network, and perform multi-level prediction on different scale feature maps for targets of different sizes by using anchor-free pixel-level detection to obtain multi-level prediction results. The computer device 110 can perform model training on the multi-level prediction results, and perform multi-level prediction and post-processing on the original image through the trained model to obtain a target detection frame. The computer device 110 can be, but is not limited to, various personal computers, notebook computers, smart phones, robots, unmanned aerial vehicles, tablet computers and the like.

[0055] In one embodiment, as shown in Figure 2 a target detection method based on a multi-scale reverse bottle-neck structure is provided, comprising the following steps:

[0056] Step 202, obtaining an original image and performing down-sampling processing on the original image to obtain a down-sampled image.

[0057] The computer device can obtain an original image, which can be used to represent an image that needs to be detected. The computer device can perform down-sampling processing on the original image, specifically, stretch the original image to a uniform image size and perform patch convolution to obtain a down-sampled image.

[0058] Step 204, inputting the down-sampled image into a multi-scale reverse bottle-neck structure, performing multi-feature expansion and fusion reverse bottle-neck operation on the down-sampled image, and generating a feature pyramid network according to a backbone network.

[0059] The computer device can perform multi-scale reverse bottle-neck operation on the down-sampled image. In the multi-scale reverse bottle-neck structure, each layer of reverse bottle-neck operation first performs multi-class convolution feature expansion and then performs channel convolution feature fusion, and after independent inverse sub-pixel convolution down-sampling, a multi-scale backbone network is obtained. Then, the computer device can generate a pyramid type 5-layer feature network according to the backbone network.

[0060] Step 206, obtaining different scale feature maps through the feature pyramid network, and using anchor-free pixel-level detection to perform multi-level prediction on different scale feature maps for targets of different sizes to obtain multi-level prediction results.

[0061] The computer device can perform multi-level prediction of 5 scales. Specifically, after obtaining different scale feature maps through the feature pyramid network, anchor-free pixel-level detection can be used to perform pixel-by-pixel target frame, class, and center degree prediction on different scale feature maps for targets of different sizes, thereby obtaining multi-level prediction results.

[0062] Step 208, model training on the multi-level prediction results, and multi-level prediction and post-processing on the original image through the trained model to obtain a target detection frame.

[0063] In the embodiment, the image is down-sampled by patch convolution, and then put into the backbone network of the multi-scale reverse bottle-neck structure, and through multi-class convolutional feature expansion and channel convolutional feature compression fusion reverse bottle-neck operation of the image by multi-class fusion convolution and inverse sub-pixel convolution, the feature pyramid network is generated according to the reverse bottle-neck backbone network, and the high-quality detection frame is obtained through multi-level prediction and post-processing of different sizes of targets on different scale feature maps by using anchor-free pixel-level detection. That is, the multi-scale detection of the image is realized by combining the multi-scale reverse bottle-neck structure with the anchor-free detection, so as to avoid information loss and improve information effectiveness, thereby solving the target detection difficulty problem caused by image noise and image distortion in the current target detection field.

[0064] In one embodiment, the target detection method based on the multi-scale reverse bottle-neck structure further includes a down-sampling process, and the specific process includes: stretching the original image to obtain a stretched image; and performing non-repeated patch convolution on the stretched image to obtain a down-sampled image.

[0065] The computer device can stretch the original image into an image with a height and a width of 224, and then perform two-dimensional patch convolution on the image by using a convolution kernel of 4 and a step of 4, so as to obtain a non-repeated convolution image, and then obtain a down-sampled image with a size of 56*56.

[0066] In one embodiment, the target detection method based on the multi-scale reverse bottle-neck structure further includes a process of generating a backbone feature map, and the specific process includes: performing large kernel space information convolution on the down-sampled image, performing layer normalization, and then obtaining image features under different receptive fields by using a dilated convolution; splicing the image features to obtain a high-dimensional feature vector; obtaining a low-dimensional feature vector according to the high-dimensional feature vector and the down-sampled image, performing down-sampling operation on the down-sampled image by using an independent inverse sub-pixel convolution to obtain the input of the next layer of the reverse bottle-neck structure, and then performing large kernel space information convolution again until the backbone feature map is generated.

[0067] As shown in FIG. 1, Figure 3 The computer device can first perform large kernel space information convolution on the down-sampled image by using a convolution kernel of 7, and then perform layer normalization, and then obtain image features under different receptive fields by using a dilated convolution, a 1*1 convolution and a 3*3 convolution, and splice the features to obtain a high-dimensional feature vector, so as to increase the width of the model and avoid information loss.

[0068] The computer device can perform channel convolution fusion on the high-dimensional feature vector, multiply the down-sampled image by the high-dimensional feature vector to obtain an effective low-dimensional feature vector, and then perform down-sampling operation on the down-sampled image by using an independent inverse sub-pixel convolution to obtain the input of the next layer of the reverse bottle-neck structure.

[0069] Next, the computer equipment can combine the above operations into an anti-bottleneck operation, and generate four layers of backbone feature maps in sequence: C2 (56*56), C3 (28*28), C4 (14*14), and C5 (7*7).

[0070] In one embodiment, the multi-scale anti-bottleneck block uses a large-kernel spatial convolution block as the first layer, a multi-scale fusion convolution as the second layer, and a channel convolution multiplied by probability as the third layer. These three layers constitute the basic anti-bottleneck block. The number of blocks in each layer of the backbone network is (3, 3, 9, 3). Each layer of the backbone network has an independent inverse sub-pixel convolution layer as a downsampling layer to generate the backbone feature map.

[0071] In one embodiment, a target detection method based on a multi-scale anti-bottleneck structure may further include a process of generating a feature pyramid network. The specific process includes: performing inverse sub-pixel convolution on the backbone network to obtain a feature map with high semantic information; upsampling and linear processing on the layers of the backbone network to obtain a feature map with low semantic information; and generating a feature pyramid network based on the feature maps with high semantic information and the feature maps with low semantic information.

[0072] like Figure 4 As shown, the computer device can perform 1x, 2x, and 4x inverse subpixel convolution on layer C5 to obtain feature maps P5, P6, and P7 with high semantic information. Then, C5 is upsampled and linearly added to C3 and C4 to obtain P3 and P4 with low semantic information, thus generating a five-layer pyramid feature network.

[0073] Next, the computer device can perform pixel-by-pixel (t, b, l, r) four-dimensional spatial vector, class vector, and centrality prediction on the five-layer feature network. The class vector is obtained through three 3*3 convolutional layers and one classification convolutional layer with the number of output channels equal to the number of classes. The spatial vector and centrality are obtained through three shared 3*3 convolutional layers and convolutional layers with 4 and 1 output channels, respectively.

[0074] In one embodiment, the target detection method based on a multi-scale anti-bottleneck structure may further include a multi-level prediction process, specifically including: scaling up each pixel in the feature maps at different scales proportionally to correspond to the center point of the corresponding block in the original image; wherein the predicted points and prediction boxes on the feature maps at different scales are all mapped onto the original image.

[0075] For multi-scale feature maps, a pixel-by-pixel multi-level prediction method can be used, each layer predicting a target of different size, the high semantic information layer predicting a larger target, and the low semantic information layer predicting a smaller target, reducing the number of ambiguous samples and avoiding the problem of special size targets being difficult to detect caused by single scale fusion features, effectively detecting multi-level targets, and simultaneously mapping the prediction points on the multi-level feature maps and the prediction boxes to the original image according to the relative positions of the pixels, i.e. each pixel point in the feature map corresponds to the center point of the corresponding block on the original image in a proportional magnification manner.

[0076] In one embodiment, the target detection method based on the multi-scale reverse bottle neck structure further includes a model training process, and the specific process includes: using distant-ness and classification loss, regression loss to train the multi-level prediction results; using target parameters to classify, regress, and threshold judge the center score of the multi-level prediction results, and using non-maximum suppression for post-processing to obtain the target detection box.

[0077] In one embodiment, the target detection method based on the multi-scale reverse bottle neck structure further includes a model training process, and the specific process includes: using distant-ness and classification loss, regression loss to train the multi-level prediction results; using target parameters to classify, regress, and threshold judge the center score of the multi-level prediction results, and using non-maximum suppression for post-processing to obtain the target detection box. 0.5 .

[0078] In this embodiment, as shown in Figure 5 the specific process of post-processing is: using given diou_threshold, distant-ness_threshold, and cls_threshold parameters to classify, regress, and threshold judge the center score of the multi-level prediction results, and using non-maximum suppression for post-processing to select high-quality target detection boxes that meet the threshold.

[0079] It should be understood that although each step in the above flowchart is displayed in sequence according to the direction of the arrow, these steps are not necessarily executed in sequence according to the direction of the arrow. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other sequences. Moreover, at least part of the steps in the above flowchart can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution sequence of these sub-steps or stages is not necessarily sequential, but can be executed in rotation or alternation with at least part of other steps or sub-steps or stages of other steps.

[0080] In one embodiment, as shown in Figure 6As shown, a target detection device based on a multi-scale reverse bottle neck structure is provided, comprising: a down-sampling processing module 610, a reverse bottle neck operation module 620, a prediction module 630 and a detection frame acquisition module 640, wherein:

[0081] The down-sampling processing module 610 is configured to acquire an original image and perform down-sampling processing on the original image to obtain a down-sampled image.

[0082] The reverse bottle neck operation module 620 is configured to input the down-sampled image into the multi-scale reverse bottle neck structure, perform multi-feature expansion and fusion reverse bottle neck operation on the down-sampled image, and generate a feature pyramid network according to a backbone network.

[0083] The prediction module 630 is configured to obtain different scale feature maps through the feature pyramid network, and perform multi-level prediction on different size targets on different scale feature maps by using anchor-free pixel-level detection to obtain multi-level prediction results.

[0084] The detection frame acquisition module 640 is configured to perform model training on the multi-level prediction results, and perform multi-level prediction and post-processing on the original image through the trained model to obtain a target detection frame.

[0085] In one embodiment, the down-sampling processing module 610 is further configured to perform stretching processing on the original image to obtain a stretched image, and perform non-repeated covering two-dimensional patch convolution on the stretched image to obtain the down-sampled image.

[0086] In one embodiment, the reverse bottle neck operation module 620 is further configured to perform large core space information convolution on the down-sampled image, perform layer normalization, and obtain image features under different receptive fields through a hollow convolution; splice the image features to obtain a high-dimensional feature vector; obtain a low-dimensional feature vector according to the high-dimensional feature vector and the down-sampled image, and perform down-sampling operation on the down-sampled image by using an independent inverse sub-pixel convolution to obtain the input of the next layer of the reverse bottle neck structure, and perform large core space information convolution again until the backbone feature map is generated.

[0087] In one embodiment, the multi-scale reverse bottle neck structure comprises a plurality of reverse bottle neck blocks; the first layer of the reverse bottle neck block is large core space information convolution; the second layer is multi-class fusion convolution; the third layer is channel convolution multiplied by input image probability; and each layer of the backbone network is provided with an independent inverse sub-pixel convolution layer as a down-sampling layer.

[0088] In one embodiment, the reverse bottle neck operation module 620 is further configured to obtain a feature map with high semantic information by performing inverse sub-pixel convolution through the backbone network; perform up-sampling and linear processing on the level of the backbone network to obtain a feature map with low semantic information; and generate a feature pyramid network according to the feature map with high semantic information and the feature map with low semantic information.

[0089] In one embodiment, the prediction module 630 is further configured to scale up each pixel point in the different scale feature map by an equal scale, and map a center point of a corresponding block on the original image; wherein the prediction points and the prediction boxes on the different scale feature map are mapped on the original image.

[0090] In one embodiment, the detection box obtaining module 640 is further configured to train the multi-level prediction result by using the distant-ness and the classification loss and the regression loss; perform threshold judgment on the multi-level prediction result by using the target parameters respectively in classification, regression and center degree score, and perform post-processing by using non-maximum suppression to obtain the target detection box.

[0091] In one embodiment, a computer device is provided, which can be a terminal, and an internal structure diagram of the computer device can be as shown in Figure 7 The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The network interface of the computer device is configured to communicate with external terminals through network connection. The computer program is executed by the processor to implement a target detection method based on a multi-scale reverse bottle-neck structure. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0092] Those skilled in the art can understand that Figure 7 the structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0093] In one embodiment, a computer device is provided, which includes a memory and a processor, and the memory stores a computer program. The processor executes the computer program to implement the following steps:

[0094] obtain an original image, and perform down-sampling processing on the original image to obtain a down-sampled image;

[0095] The down-sampling image is input into the multi-scale reverse bottle-neck structure, multi-feature expansion and fusion reverse bottle-neck operation are performed on the down-sampling image, and a feature pyramid network is generated according to the backbone network;

[0096] Different scale feature maps are obtained through the feature pyramid network, and multi-level prediction is performed on different scale feature maps for targets of different sizes by using anchor-free pixel-level detection, to obtain multi-level prediction results.

[0097] The multi-level prediction results are trained by a model, and the original image is subjected to multi-level prediction and post-processing by the trained model, to obtain a target detection frame.

[0098] In one embodiment, the processor, when executing the computer program, also implements the following steps: stretching the original image to obtain a stretched image; and performing non-repetitive two-dimensional patch convolution on the stretched image to obtain a down-sampling image.

[0099] In one embodiment, the processor, when executing the computer program, also implements the following steps: performing large-core space information convolution on the down-sampling image, obtaining image features under different receptive fields through layer normalization and hole convolution; splicing the image features to obtain a high-dimensional feature vector; obtaining a low-dimensional feature vector from the high-dimensional feature vector and the down-sampling image, and performing down-sampling operation on the down-sampling image by using an independent inverse sub-pixel convolution to obtain the input of the next layer of reverse bottle-neck structure, and performing large-core space information convolution again until the backbone feature map is generated.

[0100] In one embodiment, the processor, when executing the computer program, also implements the following steps: obtaining a feature map with high semantic information by inverse sub-pixel convolution through the backbone network; upsampling and linearly processing the levels of the backbone network to obtain a feature map with low semantic information; and generating a feature pyramid network according to the feature map with high semantic information and the feature map with low semantic information.

[0101] In one embodiment, the processor, when executing the computer program, also implements the following steps: each pixel point in the different scale feature map is enlarged by the same ratio to correspond to the center point of the corresponding block on the original image; wherein the prediction points and the prediction frame on the different scale feature map are mapped on the original image.

[0102] In one embodiment, the processor, when executing the computer program, also implements the following steps: training the multi-level prediction results by using distance and classification loss and regression loss; performing classification, regression, and threshold judgment of the central degree score on the multi-level prediction results by using target parameters, and performing post-processing by using non-maximum suppression to obtain a target detection frame.

[0103] In one embodiment, a computer readable storage medium is provided, having stored thereon a computer program which, when executed by a processor, implements the following steps:

[0104] An original image is acquired, and a down-sampling process is performed on the original image to obtain a down-sampled image;

[0105] The down-sampled image is input into a multi-scale reverse bottle-neck structure, a multi-feature expansion and fusion reverse bottle-neck operation is performed on the down-sampled image, and a backbone network is used to generate a feature pyramid network;

[0106] Different scale feature maps are obtained through the feature pyramid network, and multi-level prediction is performed on different size targets on different scale feature maps using anchor-free pixel-level detection to obtain multi-level prediction results;

[0107] The multi-level prediction results are subjected to model training, and the trained model is used to perform multi-level prediction and post-processing on the original image to obtain a target detection frame.

[0108] In one embodiment, the computer program, when executed by the processor, further implements the following steps: stretching the original image to obtain a stretched image; and performing a non-repetitive covering two-dimensional patch convolution on the stretched image to obtain the down-sampled image.

[0109] In one embodiment, the computer program, when executed by the processor, further implements the following steps: performing a large-core spatial information convolution on the down-sampled image, normalizing the layer, and obtaining image features under different receptive fields through a hole convolution; splicing the image features to obtain a high-dimensional feature vector; obtaining a low-dimensional feature vector from the high-dimensional feature vector and the down-sampled image, and performing a down-sampling operation on the down-sampled image using an independent inverse sub-pixel convolution to obtain the input of the next layer of the reverse bottle-neck structure, and performing a large-core spatial information convolution again until the backbone feature map is generated.

[0110] In one embodiment, the computer program, when executed by the processor, further implements the following steps: performing an inverse sub-pixel convolution through the backbone network to obtain a feature map with high semantic information; upsampling and linearly processing the levels of the backbone network to obtain a feature map with low semantic information; and generating a feature pyramid network from the feature map with high semantic information and the feature map with low semantic information.

[0111] In one embodiment, the computer program, when executed by the processor, further implements the following steps: each pixel point in the different scale feature map is enlarged by the same ratio to correspond to the center point of the corresponding block on the original image; wherein the prediction points and the prediction frames on the different scale feature maps are mapped on the original image.

[0112] In one embodiment, the computer program, when executed by the processor, further implements the following steps: model training on the multi-level prediction results by using the distant-ness and classification loss and regression loss; threshold judgment of the multi-level prediction results by the target parameters respectively in classification, regression and central degree score, and post-processing by using non-maximum suppression to obtain the target detection frame.

[0113] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. Any reference to memory, storage, database or other medium used in the embodiments provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM) and memory bus dynamic RAM (RDRAM) and the like.

[0114] The technical features of the above embodiments can be combined in any manner. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not exist contradictions, they should be considered as the scope of the present application.

[0115] The above embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, some modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of the patent of the present application should be subject to the appended claims.

Claims

1. A target detection method based on a multi-scale inverse bottle-neck structure, characterized in that, The method comprises: obtaining an original image and performing down-sampling processing on the original image to obtain a down-sampled image; inputting the down-sampled image into a multi-scale reverse bottle-neck structure, performing multi-feature expansion and fusion reverse bottle-neck operation on the down-sampled image, and generating a feature pyramid network according to a backbone network, comprising: performing large-core spatial information convolution on the down-sampled image, obtaining image features under different receptive fields through hole convolution after layer normalization; splicing the image features to obtain a high-dimensional feature vector; obtaining a low-dimensional feature vector according to the high-dimensional feature vector and the down-sampled image, and performing down-sampling operation on the down-sampled image by using an independent inverse sub-pixel convolution to obtain the input of the next layer reverse bottle-neck structure, and then performing large-core spatial information convolution again until the backbone feature map is generated; the multi-scale reverse bottle-neck structure comprises a plurality of reverse bottle-neck blocks; the first layer of the reverse bottle-neck block is the large-core spatial information convolution; the second layer is a multi-class fusion convolution; and the third layer is a channel convolution multiplied by an input image probability; each layer of the backbone network is provided with an independent inverse sub-pixel convolution layer as a down-sampling layer; obtaining different scale feature maps through the feature pyramid network, and performing multi-level prediction on different size targets on the different scale feature maps by using an anchor-free pixel-level detection to obtain multi-level prediction results; model training is performed on the multi-level prediction results, and multi-level prediction and post-processing are performed on the original image by using the trained model to obtain a target detection frame. 2.The multi-scale reverse-bottleneck structure based object detection method of claim 1, wherein, The method comprises: stretching the original image to obtain a stretched image; performing non-repetitive coverage two-dimensional patch convolution on the stretched image to obtain a down-sampled image. 3.The multi-scale reverse bottle-neck structure based object detection method of claim 1, wherein, The method comprises: performing inverse sub-pixel convolution on the backbone network to obtain a feature map with high semantic information; performing up-sampling and linear processing on the level of the backbone network to obtain a feature map with low semantic information; generating a feature pyramid network according to the feature map with high semantic information and the feature map with low semantic information. 4.The target detection method based on multi-scale reverse bottle-neck structure according to claim 1, characterized in that, The method comprises: scaling each pixel point in the different scale feature map proportionally to correspond to the center point of the corresponding block on the original image; wherein the prediction points and the prediction frame on the different scale feature map are mapped on the original image.

5. The method of claim 1, wherein the method further comprises: The method comprises: training a model on the multi-level prediction results by using distant-ness and classification loss and regression loss; performing classification, regression, and threshold judgment of center degree score on the multi-level prediction results by using target parameters, and performing post-processing by using non-maximum suppression to obtain a target detection frame.

6. A target detection device based on a multi-scale anti-bottleneck structure, characterized in that, The device comprises: The downsampling processing module is configured to acquire an original image and perform downsampling processing on the original image to obtain a down-sampled image. The anti-bottleneck operation module is configured to input the down-sampled image into a multi-scale anti-bottleneck structure, perform multi-feature expansion and fusion anti-bottleneck operation on the down-sampled image, and generate a feature pyramid network according to a backbone network, including: performing large-core spatial information convolution on the down-sampled image, obtaining image features under different receptive fields through hole convolution after layer normalization; splicing the image features to obtain a high-dimensional feature vector; obtaining a low-dimensional feature vector according to the high-dimensional feature vector and the down-sampled image, and performing downsampling operation on the down-sampled image by using an independent inverse sub-pixel convolution to obtain the input of the next layer of anti-bottleneck structure, and then performing large-core spatial information convolution again until the backbone feature map is generated; the multi-scale anti-bottleneck structure includes a plurality of anti-bottleneck blocks; the first layer of the anti-bottleneck block is the large-core spatial information convolution; the second layer is a multi-class fusion convolution; the third layer is a channel convolution multiplied by an input image probability; each layer of the backbone network is provided with an independent inverse sub-pixel convolution layer as a downsampling layer; The prediction module is configured to obtain different scale feature maps through the feature pyramid network, and perform multi-level prediction on different size targets on the different scale feature maps by using anchor-free pixel-level detection to obtain multi-level prediction results. The detection box acquisition module is configured to perform model training on the multi-level prediction results, and perform multi-level prediction and post-processing on the original image through the trained model to obtain a target detection box. 7.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-6 when the computer program is executed by the processor. The processor executes the computer program to realize the steps of the method in any one of claims 1 to 5.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the method in any one of claims 1 to 5.

Images