KR20220050960A - Target detection method, apparatus, electronic device and storage medium - Google Patents

Abstract

The present invention relates to a target detection method, apparatus, electronic device, and storage medium, the method comprising: acquiring a detection image to be processed; based on the image characteristic of the detection image, a size characteristic and a corner corresponding to the target detection object specifying a feature, extracting a target feature corresponding to the target detection target from the image feature based on the size feature and the corner feature, and specifying a category of the target detection target based on the target feature It relates to a target detection method comprising:

Description

Target detection method, apparatus, electronic device and storage medium

The present invention claims the priority of the Chinese patent application with the title of invention No. 201911038042.3 "target detection method and apparatus, electronic device, and storage medium", filed with the Chinese Intellectual Property Office on October 29, 2019, the entire contents of which is incorporated herein by reference.

The present invention relates to the field of computer vision technology, and more particularly to a target detection method and apparatus, an electronic device and a storage medium.

Computer vision is a technology for simulating a living body vision using a computer and related devices, and by processing the collected image or video, it is possible to acquire three-dimensional information of a corresponding scene. As one application of computer vision, target detection may be performed using a collected image or video, and the category of the target detection object and its location in the image may be specified.

Currently, in target detection technology, a detection frame for determining a category and a position of a target detection target can be directly specified using a neural network.

The present invention proposes a technical solution of target detection.

According to one aspect of the present invention, acquiring a detection image as a processing target, specifying a size characteristic and a corner characteristic corresponding to a target detection target based on an image characteristic of the detection image, the size characteristic and the corner characteristic to provide a target detection method, comprising extracting a target feature corresponding to the target detection target from the image feature based on , and specifying a category of the target detection target based on the target feature.

In one possible embodiment, specifying a size feature and a corner feature corresponding to a target detection object based on the image feature of the detection image performs one or more steps of convolution processing on the detection image, and the image of the detection image obtaining a feature, and performing corner pulling processing on the image feature of the detection image, and obtaining a size feature and a corner feature corresponding to a target detection object.

In one possible embodiment, the convolution processing includes up-sampling processing and down-sampling processing, performing one or more stages of convolution processing on the detected image, and obtaining image characteristics of the detection image is performed on the detection image. to the second stage after one or more stages of upsampling processing, based on the first feature map after one or more stages of downsampling processing, and obtaining a first feature map after one or more stages of downsampling processing obtaining a feature map; and obtaining image features of the detected image based on the first feature map after the one or more stages of downsampling processing and the second feature map after the one or more stages of upsampling processing.

In one possible embodiment, one first feature map is output after the downsampling process of each stage, and one second feature image is output after the upsampling process of each stage, and the one or more stages of downsampling processing Based on the subsequent first feature map, obtaining a second feature map after one or more stages of upsampling processing is the last stage of the first stage downsampling processing for the first stage upsampling processing among the one or more stages of upsampling processing using the first feature map after the downsampling process as an input to the first-stage upsampling process; obtaining a second feature map output after the first-stage upsampling process; A first feature matching the second feature map output after the up-sampling process immediately before the N-th up-sampling process and the second feature map output after the N-th up-sampling process with respect to the N-th up-sampling process using a map as an input to the N-th up-sampling process, and obtaining a second feature map output by the N-th up-sampling process, where N is a positive integer greater than 1.

In one possible embodiment, a first feature map matching the second feature map output after the up-sampling process immediately before the N-th stage up-sampling process and the second feature map output after the N-th stage up-sampling process is described above The input to the N-th up-sampling process is a second feature map output after the up-sampling process immediately before the N-th up-sampling process, and a second feature map output after the N-th up-sampling process. The first feature map is subjected to feature fusion to obtain an input for the N-th step up-sampling process.

In one possible embodiment, corner pulling processing is performed on the image feature of the detection image, and obtaining the size feature and corner feature corresponding to the target detection object is performed by performing corner pulling processing on the image feature of the detection image, resulting in the processing obtaining a size feature corresponding to a target detection target by performing convolution processing on the processing result using a first branch network, and using a second branch network having a different number of channels from the first branch network. performing convolution processing on the processing result, and obtaining a corner feature corresponding to the target detection target.

In one possible embodiment, extracting a target feature corresponding to the target detection object in the image feature based on the size feature and the corner feature comprises: based on the size feature and the corner feature, and specifying a feature region having a mapping relationship with an image region of the target detection target, and extracting a target feature corresponding to the target detection target from the feature region of the image feature.

In one possible embodiment, the corner feature corresponding to the target detection object comprises at least a first corner feature and a second corner feature corresponding to the target detection object, and the size feature corresponding to the target detection object is the target and a length feature and a width feature corresponding to the first corner feature of the detection object, and a length feature and a width feature corresponding to a second corner feature of the target detection object.

In one possible embodiment, the detection frame of the target detection object in the detection image is determined based on a length characteristic and a width characteristic corresponding to the first corner characteristic and a length characteristic and a width characteristic corresponding to the second corner characteristic. specifying, specifying an intersection-over-union (IoU) between any two overlapping detection frames, and when the intersection-over-union is greater than a preset threshold, the overlapping arbitrary two detection frames and merging into one detection frame.

In one possible embodiment, specifying the category of the target detection target based on the target characteristic includes performing one or more stages of convolution processing on the target characteristic, and the probability that the target detection target belongs to one or more preset categories and specifying a category of the target detection object in the preset category based on a probability that the target detection object belongs to one or more preset categories.

According to one aspect of the present invention, an acquisition module for acquiring a detection image as a processing target, a specific module for specifying size characteristics and corner characteristics corresponding to a target detection target based on image characteristics of the detection image, the size characteristic and an extraction module configured to extract a target feature corresponding to the target detection target from the image feature based on the corner feature, and a classification module configured to specify a category of the target detection target based on the target feature. provide the device.

In one possible embodiment, the specific module specifically performs one or more steps of convolution processing on the detection image, obtaining image features of the detection image, and performing corner pooling processing on the image features of the detection image, A size feature and a corner feature corresponding to the target detection object are obtained.

In one possible embodiment, the convolution process includes an up-sampling process and a down-sampling process, and the specific module specifically performs one or more stages of down-sampling processing on the detected image, and the second stage after the one or more stages of down-sampling processing. Obtain one feature map, and on the basis of the first feature map after one or more stages of downsampling processing, obtain a second feature map after one or more stages of upsampling processing, the first feature map after one or more stages of downsampling processing and the above Image features of the detected image are obtained based on the second feature map after one or more stages of up-sampling processing.

In one possible embodiment, one first feature map is output after the down-sampling process of each stage, and one second feature image is output after the up-sampling process of each stage, and the specific module is specifically configured to: With respect to the upsampling process of the first stage among the upsampling processing of more than one stage, a first feature map after the downsampling process of the final stage among the downsampling processing of the stage more than one stage is set as an input to the upsampling process of the first stage; A second feature map output after the first-stage upsampling process is obtained, and for the N-th up-sampling process in the first or more-stage upsampling processing, the second outputted after the upsampling process immediately before the N-th up-sampling process is obtained. A first feature map matching the feature map and the second feature map output after the N-th up-sampling process is input to the N-th up-sampling process, and is output by the N-th up-sampling process A second feature map is obtained, where N is a positive integer greater than 1.

In one possible embodiment, the specific module specifically matches the second feature map output after the up-sampling process immediately before the N-th stage upsampling process and the second feature map output after the N-th stage up-sampling process The input to the N-th step upsampling process is obtained by feature fusion of the first feature map.

In one possible embodiment, the specific module specifically performs corner pooling processing on the image features of the detection image, obtaining a processing result, and performing convolution processing on the processing result using a first branch network; A size feature corresponding to the target detection object is obtained, and a convolution process is performed on the processing result using a second branch network different in the number of channels from the first branch network to obtain a corner feature corresponding to the target detection object.

In one possible embodiment, the extraction module specifically specifies a feature region having a mapping relationship with an image region of the target detection object in the detection image based on the size feature and the corner feature, A target feature corresponding to the target detection target is extracted from the feature area.

In one possible embodiment, the corner feature corresponding to the target detection object comprises at least a first corner feature and a second corner feature corresponding to the target detection object, and the size feature corresponding to the target detection object is the target and a length feature and a width feature corresponding to the first corner feature of the detection object, and a length feature and a width feature corresponding to a second corner feature of the target detection object.

In one possible embodiment, the detection of the target detection object in the detection image based on a corresponding length feature, a width feature in the first corner feature and a corresponding length feature, a width feature in the second corner feature Merge that specifies a frame, determines a cross-over union between any two overlapping detection frames, and merges the two overlapping detection frames into one detection frame when the cross-over union is greater than a preset threshold Includes additional modules.

In one possible embodiment, the classification module specifically performs one or more steps of convolutional processing on the target feature, obtains a probability that the target detection object belongs to one or more preset categories, and the target detection object includes one or more Based on the probability of belonging to a preset category, the category of the target detection target is specified in the preset category.

According to one aspect of the present invention, there is provided an electronic device comprising a processor and a memory for storing instructions executable by the processor, wherein the processor is configured to execute the target detection method.

According to one aspect of the present invention, there is provided a computer readable storage medium storing computer program instructions, wherein the computer program instructions are executed by a processor to realize the target detection method.

According to one aspect of the present invention, there is provided a computer program comprising computer readable code, wherein when the computer readable code operates in an electronic device, a processor of the electronic device executes instructions for realizing the target detection method do.

In addition, it should be understood that the foregoing schematic description and the following detailed description are illustrative and interpretative only, and not limiting of the present invention.

BRIEF DESCRIPTION OF THE DRAWINGS Other features and aspects of the present invention will become apparent by describing exemplary embodiments in detail below with reference to the drawings.

Here, the drawings included as a part of the present specification are suitable for the embodiment of the present invention, and together with the specification are used for the description of the technical solution of the present invention.
1 is a flowchart of a target detection method according to an embodiment of the present invention.
2 is a block diagram of a detection frame of a target detection target according to an embodiment of the present invention.
3 is a block diagram of obtaining a category of a target detection target using a neural network according to an embodiment of the present invention.
4 is a block diagram of a target detection apparatus according to an embodiment of the present invention.
5 is a block diagram illustrating an example of an electronic device according to an embodiment of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Various exemplary embodiments, features and aspects of the present invention are described in detail below with reference to the drawings. In the drawings, the same reference numerals indicate elements having the same or similar functions. Although the drawings have shown various aspects of the embodiments, the drawings are not necessarily drawn to scale unless otherwise noted.

As used herein, the term “exemplary” means “an example, used as an example, or explanatory.” Here, any embodiment described as “exemplary” should not be understood as being preferable or superior to other embodiments.

The term "and/or" in the present specification describes a relationship with a related object, and indicates that three relationships can exist, for example, when A and/or B are only A, A and B When both exist, you may show three cases like the case where only B exists. In addition, the term "one or more" in the present specification indicates any one of a plurality or any combination of two or more of the plurality, for example, A, B and C including one or more of A, B, and C It may indicate including any one or a plurality of elements selected from the set consisting of.

In addition, in order to explain the present invention more effectively, various specific details are set forth in the following specific embodiments. It should be understood by those skilled in the art that the present invention may be practiced without any specific details. In some embodiments, detailed descriptions of methods, means, elements and circuits known to those skilled in the art are not set forth in order to emphasize the spirit of the present invention.

According to the target detection method according to the embodiment of the present invention, first, a detection image to be processed is acquired, and then size characteristics and corner characteristics corresponding to the target detection object are specified based on the image characteristics of the detection image, and the size characteristic A target feature corresponding to the target detection target may be extracted from the image feature based on and corner features, and then the category of the target detection target may be specified based on the target feature corresponding to the target detection target. Thereby, the detection result of target detection can be obtained. In this way, it is possible to first specify a target characteristic corresponding to the target detection target, and then classify the target detection target based on the corresponding target characteristic. For this reason, in target detection, the target characteristic of the target detection target and the category of the target detection target can be specified non-parallelly, a more accurate detection result can be obtained, and the precision of a detection result can be improved.

In the related art, target detection generally needs to intensively collect anchors that form a preliminary detection frame. However, since there are many invalid anchors in a large number of anchors, processing time and storage space are required. Further, in the target detection process of the related art, since the detection frame and the category of the target detection object are specified in parallel, the information of the detection frame cannot be taken into account when specifying the category of the target detection object, as a result, a sufficiently accurate detection result can't get According to the target detection method according to the embodiment of the present invention, by specifying the corner and size of the target detection target, it is possible to first specify the target characteristics of the target detection target. Reduce waste of time and storage space by collecting a large amount of anchors. Further, the target feature obtained based on the corner and the size can alleviate the difficulty of distinguishing different target detection objects according to the center point specified by the anchor, in the case where two target detection objects overlap. For this reason, according to the target detection method according to the embodiment of the present invention, it is possible to distinguish different target detection targets according to the corner features, it is possible to save time and storage space for collecting anchors, and to obtain the detection result efficiency. This is improved, and a detection result with high accuracy can be obtained.

The information processing method according to the embodiment of the present invention can be applied to any scene requiring target detection. For example, it may be applied to a scene in which target detection is performed with respect to the collected video, and the trajectory of the target detection target in the video may be obtained based on the detection result. For example, it may be applied to a security scene to recognize and track a suspect based on a detection result. Hereinafter, a target detection method according to the present invention will be described by way of example.

1 is a flowchart of a target detection method according to an embodiment of the present invention. This target detection method includes a user equipment (UE), a portable device, a user terminal, a terminal, a cellular phone, a wireless telephone, a personal digital assistant (PDA), a portable device, a computing device, an in-vehicle device, and a wearable device. It may be executed by a terminal device such as a server, or other target detection device. In some possible embodiments, this target detection method may be realized by invoking computer readable instructions stored in a memory by a processor. Hereinafter, a target detection apparatus according to an embodiment of the present invention will be described by taking the target detection method as an execution subject as an example.

As shown in Fig. 1, the target detection method includes the following steps.

Step S11: Acquire a detection image to be processed.

In the embodiment of the present invention, the target detection device may have an image acquisition function capable of photographing the current scene and acquiring a detection image to be processed, or may acquire a detection image to be processed by another device. The detected image may be an image acquired individually or may be an image frame in a video stream.

Here, the acquired detection image to be processed may be, for example, a detection image subjected to pre-processing such as image scaling, image enhancement, and image filter. For example, a preprocessed detection image can be obtained by adjusting the long and short sides of the detection image to appropriate sizes without changing the aspect ratio of the detection image.

Step S12: Based on the image characteristics of the detection image, a size characteristic and a corner characteristic corresponding to the target detection object are specified.

In an embodiment of the present invention, an image feature of a detection image can be extracted using a neural network, and a size feature corresponding to the target detection target is specified based on the image characteristic of the detection image, and a corner corresponding to the target detection target is specified. characteristics can be specified. Here, the target detection target may be a target that needs to be detected in the detection image. For example, it is an image of an object such as a pedestrian, a vehicle, a building, or a mark. The size characteristic of the target detection object may characterize the size characteristic of an image region in which the target detection object is located. For example, when the image region of the target detection object in the detection image is shown as a rectangle, the size characteristic corresponding to the target detection object may be a length characteristic and/or a width characteristic corresponding to the rectangle. The corner feature may characterize position information of a diagonal point of an image region in which the target detection object is located.

In one possible embodiment, the corner feature corresponding to the target detection object comprises at least a first corner feature and a second corner feature corresponding to the target detection object, and the size feature corresponding to the target detection object is the target detection object. and a length feature and a width feature corresponding to a first corner feature of the object, and a length feature and a width feature corresponding to a second corner feature of the target detection object.

In this embodiment, a corner may contain a 1st corner and a 2nd corner, and a pair of diagonal points may be sufficient as a 1st corner and a 2nd corner. Accordingly, the corner feature may include a first corner feature and a second corner feature. In this way, by expressing the position information of the diagonal points of the image region where the target detection target is located as a combination of the first corner feature and the second corner feature, it is possible to reduce the occurrence of a problem in that it is difficult to distinguish different target detection targets. . Accordingly, the size characteristic corresponding to the target detection object may include a corresponding length characteristic in the first corner characteristic, a corresponding length characteristic in the width characteristic and a corresponding length characteristic in the second corner characteristic, and a width characteristic. As such, based on the corresponding size features at different corner features, it is possible to further distinguish different target detection objects. Thereby, a more accurate detection result can be obtained with respect to a target detection object.

Here, the first corner may be an upper left corner or a lower right corner of the target detection target. Accordingly, the second corner may be an upper right corner or a lower left corner corresponding to the target detection object. The first corner and the second corner may be diagonal, that is, the first corner may be the upper-left corner, the second corner may be the lower-right corner, the first corner may be the upper-right corner, and the second corner may be the lower-left corner. You can do it.

Step S13: extracting a target feature corresponding to the target detection target from the image feature based on the size feature and the corner feature.

In an embodiment of the present invention, an image size corresponding to a target detection target may be specified based on a size feature corresponding to the target detection target, and an image size corresponding to the target detection target may be specified in the detection image of the target detection target based on a corner feature corresponding to the target detection target. It is possible to specify the image position, and by combining the size feature and the corner feature corresponding to the target detection object, it is possible to specify the object feature corresponding to the target detection object. The target characteristic may characterize a characteristic corresponding to an image region of the target detection object in the detection image, and the target characteristic may indicate an image position of the target detection object.

In one possible embodiment, when extracting a target feature corresponding to the target detection object, a feature region having a mapping relationship with an image region of the target detection object in the detection image is specified based on the size feature and the corner feature, and then , a target feature corresponding to the target detection target may be extracted from the feature region of the image feature.

In this embodiment, it is possible to perform feature extraction on the detection image to obtain image features of the detection image. This image feature may be represented by a feature map. Based on the size feature and the corner feature corresponding to the target detection object, it is possible to specify a feature area indicated by the size feature and the corner feature that corresponds to the image area of the target detection object in the detection image in the feature map. From the feature map, the image feature of this feature region may be extracted as a target feature corresponding to the target detection target. Based on the target characteristic, an image region of the target detection target in the detection image may be specified.

In one possible embodiment, the detection frame of the target detection object in the detection image is specified based on the corresponding length and width features in the first corner feature and the corresponding length and width features in the second corner feature. and then determine a cross-over union between any two overlapping detection frames, and if the cross-over union is larger than a preset threshold value, merge any two overlapping detection frames into one detection frame.

In this embodiment, the corresponding length feature and width feature in the first corner feature may represent one feature region in the feature map. Corresponding length features and width features in the second corner feature may represent one feature region in the feature map. The characteristic region indicated by the first corner feature and the feature region indicated by the second corner feature may be the same feature region or different feature regions. Based on the mapping relationship between the feature region and the image region of the target detection object in the detection image, the detection frame may be used to wrap the image region of the target detection object in the detection image. The detection frame may be a closed figure, and is, for example, a rectangle such as a square or a rectangle. The detection frame may specify an image region of the target detection object in the detection image, the corner feature may indicate the positions of two diagonal points of the detection frame, and the size characteristic may indicate the length and width of the detection frame there is.

Here, there is a possibility that a plurality of detection frames exist for the same target detection object, and there is a possibility that the plurality of detection frames overlap each other. For this reason, when the detection frames of the target detection object overlap, the intersection over union between arbitrary two overlapping detection frames can be calculated. If the crossover union calculated from the preset threshold is large, the two overlapping detection frames are considered to mark the same target detection object, and the larger detection frame among the two overlapping detection frames is set as the detection frame of the target detection object. and the smaller detection frame may be deleted. Alternatively, one new detection frame may be formed by merging two overlapping detection frames, and a new detection frame including the two detection frames before merging may be used as the detection frame of the target detection object. In this way, the obtained detection frames may be further selected, and detection frames of the same target detection object may be merged so that one target detection object corresponds to one detection frame.

2 is a block diagram of a detection frame of a target detection target according to an embodiment of the present invention. As an example in which the first corner is the upper left corner, the detection frame shown in FIG. 2 may be formed based on the first corner feature, the corresponding length feature in the first corner feature, and the width feature.

Step S14: Specify the category of the target detection target based on the target characteristic.

In an embodiment of the present invention, feature extraction may be additionally performed on the extracted target feature using a neural network. For example, a target detection target category can be obtained by performing convolution processing, normalization processing, or the like on the target feature. For example, the target detection target belongs to a category such as a vehicle, a pedestrian, a building, a public facility, and the like. Thereby, it is possible to obtain the category of the target detection target based on the target characteristic, and realize target detection for the target detection target in the detection image.

In the embodiment of the present invention, first, the size characteristic and the corner characteristic corresponding to the target detection object are specified by the image characteristic of the detection image. Next, a target feature corresponding to the target detection target is extracted from the image feature based on the size feature and the corner feature. Further, the category of the target detection target is specified based on the extracted target feature. In this way, the specification of the target characteristic of the target detection target and the classification of the target detection target are performed non-parallelly. When classifying the target detection target, since the target characteristics of the target detection target can be considered, a more accurate classification result can be obtained, and the precision of target detection can be improved.

In one possible embodiment, one or more stages of convolution processing are performed on the target feature, and a probability that the target detection object belongs to one or more preset categories is obtained. Then, based on the probability that the target detection object belongs to one or more preset categories, the category of the target detection object is specified in the preset category.

In this embodiment, one or more stages of convolution processing may be additionally performed on the extracted target feature using the neural network, and the probability that the target detection target belongs to one or more preset categories may be obtained. For example, the preset category is any category among categories such as pedestrians, vehicles, and buildings. By further performing convolutional processing on the target feature, it is possible to obtain a probability that the target detection target belongs to a plurality of preset categories among pedestrians, vehicles, and buildings, respectively. Then, the preset category with the highest probability can be specified as the category of the target detection object.

According to the target detection method according to an embodiment of the present invention, it is possible to first specify a target feature corresponding to the target detection target, then classify the target detection target using the target feature, and specify the category of the target detection target. there is. In this way, a detection result with high accuracy can be obtained. According to the target detection method according to an embodiment of the present invention, a category of a target detection target may be obtained using a neural network. Hereinafter, a process of obtaining a category of a target detection target using a neural network will be described.

In one possible embodiment, one or more stages of convolution processing are performed on the detection image, image features of the detection image are obtained, and thereafter, corner pooling processing is performed on the image features of the detection image, and Corresponding size features and corner features can be obtained.

In this embodiment, the neural network may include a multi-stage convolutional layer and each point pooling layer. The detected image may be input to the neural network, and multi-stage convolution processing may be performed on the detected image using the neural network to obtain image characteristics of the detected image. Thereafter, corner pooling processing is performed on the image features of the detected image using each point pooling layer of the neural network. A size characteristic and a corner characteristic corresponding to the target detection object can be obtained.

In an example of this embodiment, the convolution processing includes up-sampling processing and down-sampling processing, performing one or more stages of convolution processing on the detection image, and obtaining image features of the detection image is 1 for the detection image Performing one or more stages of downsampling processing to obtain a first feature map after one or more stages of downsampling processing, and based on the first feature map after the one or more stages of downsampling processing, a second feature map after one or more stages of upsampling processing and obtaining image features of the detection image based on the first feature map after the one or more stages of downsampling processing and the second feature map after the one or more stages of upsampling processing.

In this example, the convolution processing may include an up-sampling processing and a down-sampling processing. First, a multi-stage down-sampling processing is performed on the detected image using a neural network, and the first feature map after the down-sampling processing of each stage is obtained. get Next, a multi-stage up-sampling process is performed on the first feature map taken after the last-stage down-sampling process among the multi-stage down-sampling processes, and a second feature map after each stage of the up-sampling process can be obtained. Then, the image features of the detection image can be obtained based on the first feature map after the multi-stage down-sampling process and the second feature map after the multi-stage up-sampling process. For example, the image feature of the detection image can be obtained by fusing the first feature map after the multi-stage down-sampling process and the second feature map after the multi-stage up-sampling process. Here, the upsampling process is performed by the bilinear interpolation method, and an accurate second feature map can be obtained.

In this example, one first feature map is output after the down-sampling process of each stage, and one second feature image is output after the up-sampling process of each stage, and the first stage of the one or more stages of upsampling processing With respect to the upsampling process of , a first feature map after the downsampling process of the last stage among the downsampling processes of the first stage or more is input to the upsampling process of the first stage, and output after the upsampling process of the first stage a second feature map outputted after the up-sampling process immediately before the N-th up-sampling process for the N-th up-sampling process among the one or more up-sampling processes, and the N-th up-sampling process A first feature map matching the second feature map output after the sampling process is taken as an input to the N-th up-sampling process, and a second feature map output by the N-th up-sampling process is obtained, wherein N is a positive integer greater than 1.

In this example, multiple stages of downsampling processing are performed on the detection image, and the first feature map after each stage of downsampling processing can be obtained. With respect to the first feature map taken after the downsampling process of the last stage of the multi-stage downsampling process, the first feature map is upsampled by the upsampling process of the first stage of the multistage upsampling process, and the first stage of upsampling A second feature map after processing can be obtained. Then, based on the 2nd feature map after the upsampling process of the 1st stage, and the 1st feature map matching this 2nd feature map, the input to the upsampling process of the 2nd stage can be obtained. For example, this second feature map is fused with the first feature map to obtain an input for the second-stage upsampling process. Alternatively, after convolution processing is performed on this first feature map, it is fused with this second feature map to obtain an input for the second-stage upsampling process. Here, the first feature map matching the second feature map may be a first feature map matching the image size of the second feature map. The upsampling of the input is performed by the upsampling process of the 2nd stage, and the 2nd feature map after the upsampling process of the 2nd stage can be obtained. Then, based on the 2nd feature map after the upsampling process of the 2nd stage, and the 1st feature map matching this 2nd feature map, the input to the upsampling process of the 3rd stage is acquired. In the following manner, the second feature map after the upsampling process of the N-th stage can be obtained in the same manner. Here, N is a positive integer greater than 1. In this way, in the up-sampling processing process, the image features taken by the down-sampling processing can be taken into account, and more accurate image features are extracted.

In one example of this embodiment, corner pulling processing may be performed on the image feature of the detected image, and a processing result may be obtained. Thereafter, convolution processing is performed on the processing result using the first branch network, a size characteristic corresponding to the target detection object is obtained, and the processing result is obtained using a second branch network different from the first branch network in the number of channels. A convolution process is performed on ? to obtain a corner feature corresponding to the target detection target.

In this example, the neural network may include two branch networks, that is, a first branch network and a second branch network. After extracting the image features of the detection image using the neural network, convolution processing is performed on the image features of the detection image using the first branching network, and a feature map of the first branching network can be obtained. This feature map may correspond to four channels. In this case, one channel corresponds to the length characteristic of the first corner, one channel corresponds to the width characteristic of the first corner, one channel corresponds to the length characteristic of the second corner, and one channel corresponds to the length characteristic of the second corner. 2 Corresponds to the width characteristic of the corner. Accordingly, it is possible to perform convolution processing on the image features of the detected image using the second branch network, and obtain a feature map of the second branch network. This feature map may correspond to two channels. In this case, one channel may correspond to the first corner feature and indicate the position in the detection image of the first corner, and the other channel may correspond to the second corner feature and in the detection image of the second corner. can indicate the location of In this way, it is possible to specify the image area in which the target detection object is located based on the size characteristic and the corner characteristic corresponding to the target detection object, and it is possible to reduce the possibility that different target detection objects cannot be distinguished.

3 is a block diagram of obtaining a detection result of a target detection target using a neural network according to an embodiment of the present invention.

Hereinafter, as an example, a process for obtaining a category of a target detection target using the neural network will be described. Without changing the aspect ratio of the detection image, the long side and the short side of the detection image can be adjusted to an appropriate size, for example, the short side of the detection image can be adjusted to 800 pixels. Then, the detected image after adjustment is input to the neural network. The neural network may include multiple convolutional layers. First, a downsampling process can be performed on a detected image using a neural network. One first feature map can be obtained by the downsampling process of each stage, and the convolution process of four stages can be performed, and the first feature maps of four different sizes can be obtained. They are denoted as C ₂ , C ₃ , C ₄ , and C ₅ , respectively. Here, the long and short sides of C ₂ are all twice that of C ₃ , the long and short sides of C ₃ are all twice that of C ₄ , and the long and short sides of C ₄ are all twice that of C ₅ . Next, a 1*1 convolution kernel is calculated for C ₅ , a new feature map F ₅ is obtained, and the long and short sides of F ₅ are the same as those of C ₅ . _A multistage upsampling process is performed with respect to F5, and a 2nd feature map can be obtained by the upsampling process of each stage. That is, the bilinear interpolation upsampling process can be performed with respect to F ₅ , and a second feature map in which both the long and short sides are enlarged twice can be obtained, and this second feature map can be expressed as F ₅ ′. A 1*1 convolution kernel is calculated for C ₄ , and a new feature map C ₄ ′ can be obtained. C ₄ ' and C ₅ ' are the same size. By adding the two feature maps of C ₄ ' and C ₅ ', an input F ₄ for upsampling processing in the second stage can be obtained. Then, an upsampling process is performed on F ₄ , a second feature map F ₄ ′ in which both the long and short sides are doubled is obtained, and a 1*1 convolution kernel is calculated for C ₃ , and a new feature is obtained. A map (C ₃ ') can be obtained. C ₃ ' and F ₄ ' are the same size. By adding the two feature maps C ₃ ′ and F ₄ ′, an input F ₃ to the upsampling process of the third stage can be obtained. Hereinafter, the _2nd feature map F2 outputted after the upsampling process of the last stage can be obtained similarly by multiple times of upsampling process. The long and short sides of F ₂ are the same as those of C ₂ .

And a corner pulling process is performed with respect to the 2nd feature map F2, and a process result is obtained. A result of this processing may pass through the first branch network and the second branch network, respectively. Each branch network may contain a 3*3 convolution kernel. The first branch network may form a 4-channel feature map location, and the second branch network may form a 2-channel feature map mask. Here, the 2 channels of the feature map mask represent the upper left corner feature and the lower right corner feature, respectively, and the 4 channels of the feature map location respectively correspond to the width feature (dw), the length feature (dh), and the right side corresponding to the upper left corner. The width feature (dw) and the length feature (dh) corresponding to the bottom corner are shown.

One characteristic region may be specified based on the upper left corner feature and the lower right corner feature, the width feature and the length feature corresponding to the upper left corner, and the width feature and the length feature corresponding to the lower right corner. An image feature of a feature region may be extracted from the second feature map F ₂ , and a target feature of a target detection target may be obtained. For example, a corresponding image feature in the feature region of the second feature map F2 can be obtained by the RoI Align layer. Then, the target feature is classified using the 3*3 convolution kernel, and the target detection target category in the detection image can be obtained.

Here, the detection frame of the target detection target may be obtained by dw and dh corresponding to the upper left corner feature, the lower right corner feature, and the upper left corner, and dw and dh corresponding to the lower right corner.

Taking the width of the detection frame as an example, the width of the detection frame is calculated by the following formula (1).

However, w is the image width of the detection frame, and s, β, and α may be mapping parameters, which can be obtained by network parameters. dw is the width characteristic.

When there are a plurality of detection frames of the target detection object, non-polar value suppression processing is performed on the plurality of detection frames of the target detection object, and the plurality of detection frames of the target detection object are merged into one detection frame to obtain a final target detection object. detection results can be obtained.

According to the target detection method according to the embodiment of the present invention, it is possible to more effectively predict the detection frame of the target detection target taken based on the corner. The detection frame can be predicted more accurately, and the problem that the precision of the predicted detection frame is low due to the overlap of the target detection objects can be effectively alleviated. Further, the prediction of the detection frame and the classification of the target detection object are performed non-parallel, that is, the size characteristic and the corner characteristic indicating the position of the detection frame are first obtained, and then, the size characteristic and the target characteristic specified by the corner characteristic are obtained. Based on the classification, the target detection target can be classified, and a more accurate classification result can be obtained.

It should be understood that the embodiments of the respective methods mentioned in the present invention may be combined with each other to form embodiments as long as the principles and logic are not violated. Since there is a limit to the amount, a detailed description is omitted in the present invention.

In addition, the present invention provides a target detection device, an electronic device, a computer-readable storage medium, and a program. All of these can be used to realize any one of the target detection methods according to the present invention. Corresponding technical solutions and descriptions may refer to corresponding descriptions of methods, and detailed descriptions are omitted.

In addition, those skilled in the art will know that in the above method according to a specific embodiment, the description order of each step does not strictly limit the execution order to limit the implementation process, but the execution order of each step is specifically determined by this function and internal logic It must be understood that

4 is a block diagram of a target detection apparatus according to an embodiment of the present invention. As shown in Fig. 4, the target detection device

an acquisition module 41 for acquiring a detection image to be processed;

a specific module (42) for specifying a size feature and a corner feature corresponding to a target detection target based on the image feature of the detection image;

an extraction module (43) for extracting a target feature corresponding to the target detection target from the image feature based on the size feature and the corner feature;

and a classification module (44) for specifying a category of the target detection target based on the target characteristic.

In one possible embodiment, the specific module 42 specifically performs one or more steps of convolutional processing on the detected image, obtaining image features of the detected image, and performing corner pooling processing on the image features of the detected image. and obtains a size characteristic and a corner characteristic corresponding to the target detection object.

In one possible embodiment, the convolution processing includes up-sampling processing and down-sampling processing, and the specific module 42 specifically performs one or more stages of downsampling processing on the detected image, and one or more stages of downsampling processing. Obtain a first feature map after sampling processing,

Based on the first feature map after the one or more stages of downsampling processing, a second feature map after one or more stages of upsampling is obtained, and the first feature map after the one or more stages of downsampling processing and the one or more stages of upsampling processing Based on the later second feature map, image features of the detected image are obtained.

In one possible embodiment, one first feature map is output after the downsampling process of each stage, and one second feature image is output after the upsampling process of each stage, and the specific module 42 is configured to: , with respect to the upsampling process of the first stage in the upsampling process of the first stage or more, the first feature map after the downsampling process of the last stage among the downsampling processes of the first stage or more is input to the upsampling process of the first stage to obtain a second feature map output after the first-stage up-sampling process, and after the up-sampling process immediately before the N-th up-sampling process for the N-th up-sampling process among the first-stage or more up-sampling processes A first feature map matching the output second feature map and the second feature map output after the N-th up-sampling process is used as an input to the N-th up-sampling process, and the N-th up-sampling process is performed. A second feature map output by the process is obtained, where N is a positive integer greater than one.

In one possible embodiment, the specific module 42 specifically includes a second feature map output after the up-sampling process immediately before the N-th stage upsampling process, and a second feature output after the N-th stage upsampling process. A first feature map matching the map is feature-fused to obtain an input for the N-th step upsampling process.

In one possible embodiment, the specific module 42 specifically performs corner pooling processing on the image features of the detected image, obtaining a processing result, and performing convolution processing on the processing result using a first branching network. a size feature corresponding to the target detection target is obtained, and a convolution process is performed on the processing result using a second branch network different from the first branch network in the number of channels to obtain a corner feature corresponding to the target detection target.

In one possible embodiment, the extraction module 43 specifically specifies a feature region having a mapping relationship with an image region of the target detection object in the detection image based on the size feature and the corner feature, A target feature corresponding to the target detection target is extracted from the feature area of the image feature.

In one possible embodiment the corner feature corresponding to the target detection object comprises at least a first corner feature and a second corner feature corresponding to the target detection object, and the size feature corresponding to the target detection object is the target detection and a length feature and a width feature corresponding to a first corner feature of the object, and a length feature and a width feature corresponding to a second corner feature of the target detection object.

In one possible embodiment the device determines the target detection object in the detection image based on a corresponding length feature, a width feature in the first corner feature and a corresponding length feature, a width feature in the second corner feature. A detection frame is specified, a cross-over union between any two overlapping detection frames is determined, and when the cross-over union is greater than a preset threshold, the two overlapping detection frames are merged into one detection frame. It further includes a merge module.

In one possible embodiment, the classification module 44 specifically performs one or more stages of convolutional processing on the target feature, obtains a probability that the target detection target belongs to each preset category, and the target detection target is Based on the probability of belonging to each preset category, the category of the target detection target is specified in the preset category.

In some embodiments, functions or modules provided by the apparatus according to the embodiments of the present invention may be used to implement the methods described in the embodiments of the methods described above, and for specific realization thereof, refer to the description of the embodiments of the methods described above. For reference, a detailed description is omitted here for the sake of simplification.

An embodiment of the present invention further provides an electronic device comprising a processor and a memory for storing instructions executable by the processor, wherein the processor is configured to execute the method.

The electronic device may be provided as a terminal, server, or other type of device.

5 is a block diagram of an electronic device 1900 according to an exemplary embodiment. For example, the electronic device 1900 may be provided as a server. Referring to FIG. 5 , an electronic device 1900 includes a processing component 1922 including one or more processors and a memory 1932 for storing instructions executable by the processing component 1922 , for example, an application program. Includes memory resources represented by . The application program stored in the memory 1932 may include one or more modules each corresponding to one instruction group. Further, processing component 1922 is configured to execute the method by executing instructions.

The electronic device 1900 further includes a power component 1926 configured to perform power management of the electronic device 1900 , a wired or wireless network interface 1950 configured to connect the electronic device 1900 to a network, and input/output (I) /O) interface 1958. The electronic device 1900 may operate based on an operating system stored in the memory 1932 , for example, Windows Server ^TM , Mac OS X ^TM , Unix ^TM , Linux ^TM , FreeBSD ^TM or the like.

In the exemplary embodiment, there is further provided a volatile computer readable storage medium or a nonvolatile computer readable storage medium, for example, a memory 1932 comprising computer program instructions, wherein the computer program instructions are stored in the electronic device 1900 . ), may cause the method to be executed.

The exemplary embodiment further provides a computer program comprising computer readable code, which, when the computer readable code operates in an electronic device, executes instructions for realizing the method in a processor of the electronic device.

The invention may be a system, method and/or computer program product. The computer program product may include a computer readable storage medium having computer readable program instructions for realizing each aspect of the present invention in a processor.

The computer-readable storage medium may be a tangible device capable of storing and storing instructions used in an instruction execution device. The computer-readable storage medium may be, for example, an electrical storage device, a magnetic storage device, an optical storage device, an electronic storage device, a semiconductor storage device, or any suitable combination of the above. More specific examples (non-exhaustive list) of computer-readable storage media include portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory); Static random access memory (SRAM), portable compact disk read-only memory (CD-ROM), digital versatile disk (DVD), memory stick, floppy disk, such as a punched card or slot in which instructions are stored; such mechanical encoding devices, and any suitable combination of the above. As used herein, a computer-readable storage medium is an instantaneous signal itself, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating via waveguides or other transmission media (eg, optical pulses passing through optical fiber cables). ), or an electrical signal transmitted via a wire.

The computer readable program instructions described herein may be downloaded to each computing/processing device from a computer readable storage medium, or may be downloaded to an external computer via a network, for example, the Internet, a local area network, a wide area network and/or a wireless network. Alternatively, it may be downloaded to an external storage device. The network may include copper transport cables, fiber optic transport, wireless transport, routers, firewalls, switchboards, gateway computers and/or edge servers. A network adapter card or network interface in each computing/processing device receives a computer readable program command from the network, transmits the computer readable program command, and stores the computer readable program command in a computer readable storage medium in each computing/processing device.

The computer program instructions for carrying out the operations of the present invention may include assembler instructions, instruction set architecture (ISA) instructions, machine language instructions, machine dependent instructions, microcode, firmware instructions, state setting data or an object-oriented programming language such as Smalltalk, C++, etc.; and source code or target code written in any combination of one or more programming languages including a general procedural programming language such as "C" language or a similar programming language. The computer readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partly on a remote computer, or It may run entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any kind of network, including a local area network (LAN) or a wide area network (WAN), or (eg, an Internet service provider via the Internet) may be connected to an external computer. In some embodiments, state information from computer readable program instructions is used to customize electronic circuitry, such as, for example, a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), the electronic circuitry. Each aspect of the present invention may be realized by executing computer readable program instructions by

Here, although each aspect of the present invention has been described with reference to flowcharts and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the present invention, each block in the flowcharts and/or block diagrams, and flowcharts And/or it should be understood that all combinations of blocks in the block diagram may be realized by computer-readable program instructions.

These computer readable program instructions are provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing device, and when these instructions are executed by the processor of the computer or other programmable data processing device, one or more of the flowcharts and/or block diagrams The device may be manufactured to realize the functions/actions specified in the block. These computer readable program instructions may be stored in a computer readable storage medium, and may cause a computer, a programmable data processing apparatus, and/or other apparatus to operate in a specific manner. Thereby, a computer-readable storage medium having instructions stored thereon includes a product having instructions for realizing each aspect of a function/action specified in one or more blocks of a flowchart and/or a block diagram.

The computer readable program instructions may be loaded into a computer, other programmable data processing device, or other device, and cause the computer, other programmable data processing device, or other device to execute a series of operational steps, thereby creating a process executed by the computer. . In this way, the functions/operations specified in one or more blocks of the flowchart and/or block diagram are realized by instructions executed in a computer, other programmable data processing device, or other device.

Flowcharts and block diagrams in the drawings represent realizable system architectures, functions, and operations of systems, methods, and computer program products according to a plurality of embodiments of the present invention. In this regard, each one or more blocks in the flowchart or block diagram may represent a module, program segment, or portion of an instruction, wherein the module, program segment, or portion of the instruction may represent one or more executions for realizing a specified logical function. Includes possible commands. In some alternative implementations, the functions indicated in the blocks may be realized in a different order from the order in the drawings. For example, two consecutive blocks may be executed substantially in parallel, or may be executed in the reverse order depending on the function involved. In addition, each block in the block diagram and/or flowchart, and a combination of blocks in the block diagram and/or flowchart may be realized by a dedicated system based on hardware for executing specified functions or operations, or dedicated hardware and It should also be noted that it may be realized by a combination of computer instructions.

As mentioned above, although each embodiment of this invention was described, the said description is only exemplary, and is not exhaustive, and is not limited to each disclosed embodiment, either. Various modifications and changes will be apparent to those skilled in the art without departing from the scope and spirit of each described embodiment. The terminology selected herein is intended to preferably interpret the principle, practical application, or improvement of existing techniques of each embodiment, or to enable others skilled in the art to understand each embodiment disclosed herein.

Claims (23)

acquiring a detection image to be processed;
specifying a size characteristic and a corner characteristic corresponding to a target detection target based on the image characteristic of the detection image;
extracting a target feature corresponding to the target detection target from the image feature based on the size feature and the corner feature;
and specifying a category of the target detection target based on the target characteristic. The method of claim 1,
Specifying the size feature and the corner feature corresponding to the target detection target based on the image feature of the detection image
performing one or more steps of convolution processing on the detected image to obtain image characteristics of the detected image;
and performing corner pooling processing on the image features of the detection image, and obtaining size features and corner features corresponding to the target detection object. 3. The method of claim 2,
The convolution process includes an up-sampling process and a down-sampling process,
Performing one or more stages of convolution processing on the detected image, and obtaining image characteristics of the detected image
performing one or more stages of downsampling processing on the detected image, and obtaining a first feature map after one or more stages of downsampling processing;
obtaining a second feature map after one or more stages of upsampling processing on the basis of the first feature map after the one or more stages of downsampling;
and obtaining image features of the detection image based on the first feature map after the one or more stages of downsampling processing and the second feature map after the one or more stages of upsampling processing. 4. The method of claim 3,
outputting one first feature map after the down-sampling process of each stage, and outputting one second feature image after the up-sampling process of each stage;
Based on the first feature map after one or more stages of downsampling processing, obtaining a second feature map after one or more stages of upsampling processing is
With respect to the upsampling process of the first stage among the upsampling processing of the first stage or more, a first feature map after the downsampling process of the last stage among the downsampling processing of the first stage is input to the upsampling process of the first stage and,
obtaining a second feature map output after the first-stage up-sampling process;
A second feature map output after the upsampling process immediately before the N-th up-sampling process with respect to the N-th up-sampling process among the one or more up-sampling processes, and a second output after the N-th up-sampling process using a first feature map matching the feature map as an input to the N-th step upsampling process;
and obtaining a second feature map output by the N-th up-sampling process, wherein N is a positive integer greater than one. 5. The method of claim 4,
A first feature map matching the second feature map output after the upsampling process immediately before the N-th-stage up-sampling process and the second feature map output after the N-th up-sampling process is subjected to upsampling of the N-th stage What to do as input to processing
The second feature map output after the upsampling process immediately before the N-th up-sampling process and the first feature map matching the second feature map output after the N-th up-sampling process are feature-fused, and the N-th stage A target detection method, obtaining an input for the upsampling process of. 3. The method of claim 2,
Performing corner pulling processing on the image features of the detection image, and obtaining the size features and corner features corresponding to the target detection target
performing corner pooling processing on the image features of the detected image, and obtaining a processing result;
performing convolution processing on the processing result using a first branch network to obtain a size characteristic corresponding to a target detection target;
A target detection method comprising: performing convolution processing on the processing result by using a first branching network and a second branching network having a different number of channels, and obtaining a corner feature corresponding to a target detection object. 7. The method according to any one of claims 1 to 6,
extracting a target feature corresponding to the target detection target from the image feature based on the size feature and the corner feature
specifying a feature region having a mapping relationship with an image region of the target detection target in the detection image based on the size feature and the corner feature;
and extracting a target feature corresponding to the target detection target from a feature region of the image feature. 8. The method of claim 7,
The corner feature corresponding to the target detection object includes at least a first corner feature and a second corner feature corresponding to the target detection object,
The size characteristic corresponding to the target detection object includes a length characteristic and a width characteristic corresponding to a first corner characteristic of the target detection object, and a length characteristic and a width characteristic corresponding to a second corner characteristic of the target detection object, Target detection method. 9. The method of claim 8,
specifying a detection frame of the target detection object in the detection image based on a length characteristic and a width characteristic corresponding to the first corner characteristic and a length characteristic and a width characteristic corresponding to the second corner characteristic;
determining a cross over union between any two overlapping detection frames;
When the cross over union is greater than a preset threshold, the method further comprising merging the arbitrary two overlapping detection frames into one detection frame. 10. The method according to any one of claims 1 to 9,
Specifying the category of the target detection target based on the target characteristic includes:
performing one or more stages of convolution processing on the target feature, and obtaining a probability that the target detection target belongs to one or more preset categories;
and specifying a category of the target detection object in the preset category based on a probability that the target detection object belongs to one or more preset categories. an acquisition module for acquiring a detection image to be processed;
a specific module for specifying a size characteristic and a corner characteristic corresponding to a target detection target based on the image characteristic of the detection image;
an extraction module for extracting a target feature corresponding to the target detection target from the image feature based on the size feature and the corner feature;
and a classification module for specifying a category of the target detection target based on the target characteristic. 12. The method of claim 11,
The specific module is specifically,
performing one or more stages of convolution processing on the detected image to obtain image characteristics of the detected image;
A target detection apparatus, wherein a corner pulling process is performed on an image characteristic of the detection image, and a size characteristic and a corner characteristic corresponding to a target detection object are obtained. 13. The method of claim 12,
The convolution process includes an up-sampling process and a down-sampling process,
The specific module is specifically,
performing one or more stages of downsampling processing on the detected image, and obtaining a first feature map after one or more stages of downsampling processing;
obtaining a second feature map after one or more stages of upsampling processing on the basis of the first feature map after one or more stages of downsampling processing;
An image feature of the detection image is obtained based on a first feature map after the one or more stages of downsampling processing and a second feature map after the one or more stages of upsampling processing. 14. The method of claim 13,
outputting one first feature map after the down-sampling process of each stage, and outputting one second feature image after the up-sampling process of each stage;
The specific module is specifically,
With respect to the first-stage upsampling processing among the one or more stages of upsampling processing, a first feature map after the downsampling processing of the last stage among the one or more stages of downsampling processing is used as an input to the first stage upsampling processing, ,
Obtaining a second feature map output after the up-sampling process of the first stage,
A second feature map output after the upsampling process immediately before the N-th up-sampling process with respect to the N-th up-sampling process among the one or more up-sampling processes, and a second output after the N-th up-sampling process using a first feature map matching the feature map as an input to the N-th step upsampling process;
and obtain a second feature map output by the N-th up-sampling process, wherein N is a positive integer greater than 1. 15. The method of claim 14,
The specific module specifically includes a second feature map output after the up-sampling process immediately before the N-th up-sampling process, and a first feature map matching the second feature map output after the N-th up-sampling process. A target detection device for obtaining an input for the N-th up-sampling process by fusion. 13. The method of claim 12,
The specific module is specifically,
performing corner pooling processing on the image features of the detected image to obtain a processing result;
performing convolution processing on the processing result using a first branch network to obtain a size characteristic corresponding to a target detection target;
A target detection apparatus, comprising: performing convolution processing on the processing result using a second branching network different from the first branching network in the number of channels to obtain a corner feature corresponding to a target detection target. 17. The method according to any one of claims 11 to 16,
The extraction module is specifically,
specifying a feature region having a mapping relationship with an image region of the target detection target in the detection image based on the size feature and the corner feature;
and extracting a target feature corresponding to the target detection target from a feature region of the image feature. 18. The method of claim 17,
The corner feature corresponding to the target detection object includes at least a first corner feature and a second corner feature corresponding to the target detection object,
The size characteristic corresponding to the target detection object includes a length characteristic and a width characteristic corresponding to a first corner characteristic of the target detection object, and a length characteristic and a width characteristic corresponding to a second corner characteristic of the target detection object, target detection device. 19. The method of claim 18,
specifying a detection frame of the target detection object in the detection image based on a length characteristic and a width characteristic corresponding to the first corner characteristic and a length characteristic and a width characteristic corresponding to the second corner characteristic; Target detection, further comprising a merging module for determining a cross-over union between detection frames, and merging the two overlapping detection frames into one detection frame when the cross-over union is greater than a preset threshold value Device. 20. The method according to any one of claims 11 to 19,
The classification module is specifically,
performing one or more stages of convolution processing on the target feature, and obtaining a probability that the target detection target belongs to one or more preset categories;
and specifying a category of the target detection object in the preset category based on a probability that the target detection object belongs to one or more preset categories. processor and
a memory for storing instructions executable by the processor;
The electronic device, wherein the processor is configured to execute the method according to any one of claims 1 to 10 by invoking an instruction stored in the memory. A computer readable storage medium having stored thereon computer program instructions, which, when executed by a processor, realizes the method of any one of claims 1 to 10. A computer program comprising computer readable code, wherein when the computer readable code operates in an electronic device, a processor of the electronic device executes instructions for realizing the method of any one of claims 1 to 10.

Images