KR20220050617A - Method and device for detecting object location based on weak supervised learning - Google Patents

Abstract

The present invention relates to a method and device for detecting an object location based on weak supervised learning. According to the present invention, the method comprises the following steps by an object location detection device: extracting a feature map for an input image by using a first convolutional layer of a deep learning model; generating a first attention map by using the feature map; generating a dropped foreground mask from the first attention map, wherein the dropped foreground mask is generated on the basis of a contrastive attention loss; and estimating an object location on the basis of an importance map generated from the dropped foreground mask or the first attention map. Accordingly, the method provides an effect of preventing performance degradation of object location estimation.

Description

Weak supervised learning-based object location detection method and device {METHOD AND DEVICE FOR DETECTING OBJECT LOCATION BASED ON WEAK SUPERVISED LEARNING}

The disclosed technology relates to a weak supervised learning-based object position detection method and apparatus.

Supervised learning requires labeling close to the perfect answer in order to identify objects included in the image input to the convolutional neural network. Labels are used by neural networks to accurately classify objects. As such, supervised learning has a disadvantage in that the cost required for labeling is increased. Recently, unsupervised learning or weak supervised learning-based object detection technology that only provides image-level annotations is being used.

On the other hand, in the conventional object position detection technology using a convolutional neural network, an object is identified based on a region most strongly identified in the object. Accordingly, rather than searching for the whole object, it showed the discrimination power focused on a part.

To improve this problem, learning methods have been developed in the state that the important area is covered so that the entire object can be found. This leads the learning model to learn from the image to the relatively less important background, so a box that is much larger than the size of the real object is used. As a result, there was a problem in that the accuracy in identifying the location of the object was lowered.

Korean Patent Publication No. 10-2020-0074940

The disclosed technology is to provide a method and apparatus for detecting an object position using weak supervised learning through sample collation in an image.

A first aspect of the disclosed technology to achieve the above technical problem is the step of extracting a feature map for the input image using the first convolutional layer of the deep learning model, by the object position detection apparatus, the object position generating, by a detection device, a first attention map by using the feature map, the object position detection device generating a dropped foreground mask from the first attention map, wherein the dropping The foreground mask is generated based on a contrastive attention loss, and the object position detection device is an important map generated from the dropped foreground mask or the first attention map (Importance Map) An object of the present invention is to provide a method for detecting an object location, including estimating the location of the object as a reference.

A second aspect of the disclosed technology to achieve the above technical task is to use an input device for receiving an image including an object, a storage device for storing a deep learning model, and a first convolutional layer of the deep learning model to the image. A first attention map is generated by extracting a feature map, an attention block is set in the feature map, and a dropped foreground mask from the first attention map ), but the dropped foreground mask is generated based on a contrastive attention loss, and an important map generated from the dropped foreground mask or the first attention map It is to provide an object position detection apparatus including an arithmetic unit for estimating the object position as a reference.

Embodiments of the disclosed technology may have effects including the following advantages. However, since it does not mean that the embodiments of the disclosed technology should include all of them, it should not be understood that the scope of the disclosed technology is limited thereby.

According to an embodiment of the disclosed technology, a weak supervised learning-based method and apparatus for detecting a position of an object generates a box indicating an area of an object, but has an effect of preventing the box from becoming excessively large.

In addition, there is an effect of preventing the degradation of object position estimation performance by using the contrasting attention loss and the foreground consistency loss for training.

In addition, by estimating the location of the object based on the attention map generated by applying the non-local attention block, there is an effect of covering the entire area without being biased toward a specific area of the object.

1 is a diagram illustrating estimating an object position using a weak supervised learning-based object position detection system according to an embodiment of the disclosed technology.
2 is a diagram illustrating an overall flow of estimating an object position according to an embodiment of the disclosed technology.
3 is a diagram illustrating calculating a contrasting attention loss according to an embodiment of the disclosed technology.
4 is a diagram showing the configuration of an object position detection apparatus according to an embodiment of the disclosed technology.
5 is a flowchart of a method for detecting an object position according to an embodiment of the disclosed technology.
6 is a diagram comparing results of object position estimation.
7 is a diagram illustrating application of the disclosed technology to a conventional image data set.

Since the present invention can have various changes and can have various embodiments, specific embodiments are illustrated in the drawings and described in detail in the detailed description. However, this is not intended to limit the present invention to specific embodiments, and it should be understood to include all modifications, equivalents and substitutes included in the spirit and scope of the present invention.

Terms such as first, second, A, and B may be used to describe various components, but the components are not limited by the above terms, and only for the purpose of distinguishing one component from other components. is used only as For example, without departing from the scope of the present invention, a first component may be referred to as a second component, and similarly, a second component may also be referred to as a first component. and/or includes a combination of a plurality of related listed items or any of a plurality of related listed items.

In terms of terms used herein, the singular expression is to be understood as including the plural expression unless the context clearly dictates otherwise. And terms such as "comprising" mean that the specified feature, number, step, operation, component, part, or a combination thereof exists, but one or more other features or number, step operation component, part It should be understood as not excluding the possibility of the presence or addition of or combinations thereof.

Prior to a detailed description of the drawings, it is intended to clarify that the classification of the constituent parts in the present specification is merely a division according to the main function each constituent unit is responsible for. That is, two or more components to be described below may be combined into one component, or one component may be divided into two or more for each more subdivided function.

In addition, each of the constituent units to be described below may additionally perform some or all of the functions of other constituent units in addition to the main function it is responsible for. Of course, it can also be performed by being dedicated to it. Accordingly, the existence or non-existence of each component described through the present specification should be interpreted functionally.

1 is a diagram illustrating estimating an object position using a weak supervised learning-based object position detection system according to an embodiment of the disclosed technology. Referring to FIG. 1 , the object position detection system 100 may receive an image input by the user using the computing device 110 such as the user's PC or smart phone. The computing device may use other types of devices capable of performing calculations according to algorithms. The computing device includes a device for receiving a user's input, a device for storing an algorithm for calculation, a program, etc., a device for performing a calculation using the user's input and the stored algorithm or program, and a device for outputting the calculation result to the user can do. An image input by a user and a comment on the image may be input to the input device. In one embodiment, the input device may include a mouse for uploading an image file and a keyboard for inputting comments on the image file. In order to receive such an input, the computing device may store an application or program that outputs a predetermined interface for a user.

Meanwhile, the user may input data into the computing device 110 . The data input by the user is basically an image of an object. For the image, a user's own shot may be used, or a specific image provided from a separate dataset may be used. The input data includes annotations on the objects along with these images. Here, only image-level annotations may be input, not entire labeling data. That is, the user may input an image including the object and a comment on the object as data to the computing device 110 . In addition, the computing device may receive data and estimate the location of the object by performing weak supervised learning according to the stored algorithm. Estimating the position of the object according to the algorithm may be performed by a device in charge of calculation in the computing device, and the calculation result may be provided to a user through an output device.

Meanwhile, in the disclosed technology, a convolutional layer is trained using two functions, a contrastive attention loss and a foreground consistency loss, in order to more accurately estimate an object position. First, in order to calculate the contrasting attention loss, an attention map is generated based on a feature map extracted from an image, and a threshold value is applied to the attention map to generate a plurality of masks. The attention map is generated by applying the attention block to the feature map, and the attention block may be set somewhat differently according to an algorithm stored in the computing device. However, the attention block referred to in the disclosed technology is intended to replace channel pooling and has a non-local characteristic. That is, an improved attention map is generated from the feature map in consideration of spatial similarity. Given a feature map, the attention block embeds them into Channel, Height, and Weight, respectively, and outputs the spatial sum of the third embedding weighted by the similarity between the first two embeddings. An enhanced attention map may be defined from the feature map by the channel pooling result as described above.

Meanwhile, by using the attention map, three masks are generated: a foreground mask, a background mask, and a dropped foreground mask in the form of obscuring a specific area from the foreground mask. Here, the specific region means the most differentiated region among the background and foreground of the image. In the conventional neural network-based object location detection technology, since the bounding box for the feature map is generated based on this specific area, it may not cover the entire area of the object or generate a bounding box much larger than the size of the actual object. . Therefore, the disclosed technology uses a dropped foreground mask created by covering the most differentiated area in order to solve this problem. The computing device may embed a feature of the image by multiplying each mask by the input feature.

Meanwhile, the dropped foreground mask is generated based on the contrasting attention loss. For example, in embedding the first attention map into the feature space, the dropped foreground mask may be generated so as to be as similar to the embedding result of the foreground mask as possible and as far away from the embedding result of the background mask as possible. The dropped foreground mask can be created by applying a threshold to the first attention map to cover the most differentiated area between the background and the foreground, and the feature values embedded through the generated mask can be located in a space close to the foreground mask. and may be located in a space somewhat distant from the feature value of the background mask.

Meanwhile, the computing device may sequentially perform the contrasting attention loss from the initial layer to the upper layer by inputting the feature map to each of the plurality of convolutional layers. First, in the initial layer that calculates the contrasting attention loss, the feature map extracted from the image is used as an input value as it is, but in the later upper layer, the dropped foreground mask and the important map generated in the previous layer are randomly selected. Multiplying the selected one by the feature map is used as an input value. Here, the dropped foreground mask is generated by applying a threshold to the attention map as described above, and the important map is generated by applying sigmoid activation to the attention map.

That is, in the subsequent layer, a feature map to which the dropped foreground mask or important map is applied as a weight is input, and the contrast-attention loss is performed in the same way as in the previous layer. Here the dropped foreground mask or critical map becomes pixel-wise multiplication. Pixel-wise multiplication refers to element-wise multiplication that is broadcast at the channel level. In this case, since weights are applied to input values in subsequent layers, an attention map different from the attention map generated in the previous layer is generated. Accordingly, the foreground consistency loss is applied to learn to reduce the difference between the two attention maps. Although only two layers have been mentioned as an example for explanation, since there are actually a plurality of layers in the convolutional neural network, the foreground consistency loss can be calculated so that the attention maps between the first and last layers are similar to each other.

2 is a diagram illustrating an overall flow of estimating an object position according to an embodiment of the disclosed technology. Referring to FIG. 2 , each of the contrasting attention loss 220 is performed on a plurality of convolutional layers included in the deep learning model 200, and the attention map 210 of the first layer 201 and the last layer 202 Foreground consistency loss 230 is performed so that the attention map of n has a similar value.

First, a feature map for an image input from the first layer 201 may be extracted. The extraction of the feature map is processed in a manner similar to that performed in a conventional convolutional neural network. For example, an attention map may be generated by applying an attention block having a non-local characteristic to the feature map. Then, a dropped foreground mask is generated based on the generated attention map, and an operation for estimating an object position is performed using the attention map and the dropped foreground mask. A result of estimating an object's position may be adjusted through a contrasting attention loss. For example, the deep learning model may be trained so that the contrasting attention loss value becomes small in order to estimate the position of the object.

Meanwhile, a dropped foreground mask generated based on the contrasting attention loss may be calculated according to Equation 1 below.

는 사전에 정의된 전경(Foreground) 영역이고 ,

는 사전에 정의된 배경(Background) 영역을 의미한다. 즉, 드롭된 포어그라운드 마스크는 어텐션 맵에서 가장 차별화된 영역을 제거하여 생성되되, 임베딩 된 특징이 포어그라운드 마스크의 임베딩 결과와는 유사해지고, 백그라운드 마스크의 임베딩 결과는 서로 멀어지도록 생성될 수 있다.Here, A means attention map,

is the predefined foreground area,

denotes a predefined background area. That is, the dropped foreground mask may be generated by removing the most differentiated region from the attention map, the embedded features may be similar to the embedding result of the foreground mask, and the embedding result of the background mask may be generated to move away from each other.

Meanwhile, the contrasting attention loss may be calculated with reference to Equation 2 below.

는 드롭된 포어그라운드 마스크를 통해 임베딩된 특징이고,

는 포어그라운드 마스크를 통해 임베딩된 특징이고,

는 백그라운드 마스크를 통해 임배딩된 특징을 의미한다. 그리고

는 제 1 마스크 및 제 2 마스크의 인스턴스의 배열이고,

는 제 1 마스크 및 제 3 마스크의 인스턴스의 배열이고,

은 마진이다. 즉, 제 1 마스크를 통해 출력된 특징과 제 2 마스크를 통해 출력된 특징의 인스턴스는 서로 유사하고, 제 1 마스크의 특징과 제 3 마스크의 특징의 인스턴스는 서로 상이하게 출력된다. From here

is the feature embedded through the dropped foreground mask,

is a feature embedded through the foreground mask,

denotes a feature embedded through a background mask. And

is the array of instances of the first mask and the second mask,

is the array of instances of the first mask and the third mask,

is the margin That is, instances of features output through the first mask and features output through the second mask are similar to each other, and instances of features of the first mask and features of the third mask are output differently.

A contrasting attention loss function is a function that yields a small value when a query is similar to an equivalent instance and a query is not similar between different instances. That is, the dropped foreground mask refers to a mask that makes the calculation result of the contrast attention loss smaller. Contrastive attention loss can guide the attention map until it reaches the boundary, because including the background in the attention map causes a penalty due to dissimilarity. Therefore, it is not necessary to extract samples with values such as triplet loss, and negative samples with large values are not managed, and positive samples and negative samples are considered based on the features output through the first to third masks. can do.

Meanwhile, the first layer applies a weight to the feature map to generate the input value of the next layer. Here, a dropped foreground mask may be applied as a weight for the feature map or an important map may be applied. By multiplying the feature map by any one randomly determined in the deep learning model, the input value of the next layer can be calculated.

Meanwhile, the second layer performs an operation in the same manner as the first layer using the weighted feature map as an input value. In addition, the third layer performs the same operation by using the weighted feature map calculated in the second layer as an input value. According to this method, the operation may be sequentially performed up to the last layer 202 . When the operation to the last layer is completed, the attention map of the first layer and the attention map of the last layer can be compared. Depending on the number of layers, the difference between the attention maps generated in the first layer and the last layer may increase. The object position detection apparatus may learn the deep learning model so that the two attention maps are similar to each other. For example, the foreground consistency loss that reduces the difference between the two attention maps may be calculated by reducing the background activation of the first layer based on the last layer.

Meanwhile, the foreground consistency loss may be calculated with reference to Equation 3 below.

는 i번째 어탠션 맵이고,

는 j번째 어탠션 맵을 의미한다. 즉, i번째 어탠션 맵이 초기 레이어에서 출력되는 어탠션 맵이고 j번째 어탠션 맵이 이후의 레이어를 통해 출력되는 어탠션 맵이다. 어탠션 맵은 대략적으로 모든 위치(Location)의 활성화를 나타낸 맵이며, 초기 레이어는 객체의 전체 범위 대신 가장자리나 모서리와 같은 국부적인 영역을 구별하는 것에서 더 활성화될 수 있다. 이러한 문제점을 해소하기 위해서 초기 레이어의 어탠션 맵과 이후 레이어의 어탠션 맵이 일관성을 유지할 수 있도록 포어그라운드 컨시스턴시 로스를 계산함으로써 두 어탠션 맵이 유사해지도록 학습할 수 있다.From here

is the i-th attention map,

denotes the j-th attention map. That is, the i-th attention map is an attention map output from the initial layer, and the j-th attention map is an attention map output through subsequent layers. The attention map is a map showing the activation of approximately all locations, and the initial layer may be further activated in distinguishing local areas such as edges or corners instead of the entire range of an object. In order to solve this problem, the two attention maps can be learned to be similar by calculating the foreground consistency loss so that the attention map of the initial layer and the attention map of the subsequent layer are consistent.

Meanwhile, the object position detection apparatus may calculate the total loss of the deep learning model according to Equation 4 below.

는 종래의 클래스 로스를 의미하고,

는 컨트라스티브 어텐션 로스를 의미하고,

는 포어그라운드 컨시스턴시 로스를 의미한다. 클래스 로스는 네트워크의 마지막에 GAP(Global average pooling) 레이어를 구성함으로써 소프트맥스 출력값을 생성하고 one-hot ground truth label이 주어지면 계산할 수 있으며, 앞서 언급한 수학식 1을 통해 컨트라스티브 어텐션 로스를 계산하고 수학식 2를 통해 포어그라운드 컨시스턴시 로스를 계산할 수 있다. From here

means the conventional class loss,

stands for contrasting attention loss,

is the foreground consistency loss. Class loss can be calculated when a softmax output is generated by configuring a global average pooling (GAP) layer at the end of the network and a one-hot ground truth label is given. and the foreground consistency loss can be calculated through Equation (2).

The object position detection apparatus can calculate the total loss of the deep learning model by calculating the contrasting attention loss and the foreground consistency loss as described above, and adding it to the classification loss for class classification in the image. The object position detecting apparatus may generate a dropped foreground mask so that the contrasting attention loss value is small, and train the deep learning model so that the foreground consistency loss is small so that the attention maps for each layer become similar. That is, by repeating the learning process so that the overall loss is small, a bounding box covering the entire object can be created, but the size of the box can be prevented from becoming excessively large. Accordingly, unlike the performance degradation in the conventional hierarchical model, it is possible to increase the object location estimation performance by forming a box similar to the area corresponding to the entire object.

3 is a diagram illustrating calculating a contrasting attention loss according to an embodiment of the disclosed technology. Referring to FIG. 3 , in each layer, three masks may be generated using an attention map. As in the prior art, a foreground mask may be generated by removing a region corresponding to the background of an image, or a background mask may be generated by removing a region corresponding to the foreground of the image. In addition, the dropped foreground mask 310 may be generated in the manner described above with reference to FIGS. 1 and 2 . The dropped foreground mask 310 is generated by applying a threshold value 310a to the first attention map. Here, the threshold 310a may be an area in which the most differentiated feature appears among the background and foreground of the image. For example, when the object is an animal, the face of the animal may be regarded as a differentiated region.

The dropped foreground mask 310 and the foreground mask may be induced to learn from a differentiated area to a less important part within the object area as learning is progressed. In addition, the background mask may be induced to generate a differentiated region toward the background. According to this process, the dropped foreground mask and the embedding result of the foreground mask may be learned to be similar to each other, and the dropped foreground mask and the embedding result of the background mask may be learned to move away from each other. That is, learning can be performed using the triplet loss form.

The first attention map input to each of the three masks is embedded into the feature space 310b. In the feature space, the characteristics of the dropped foreground mask and the characteristics of the foreground mask are expressed as similar to each other as possible, and the characteristics of the dropped foreground mask and the characteristics of the background mask are expressed as similarly as possible. For example, the characteristic of the dropped foreground mask and the characteristic of the foreground mask may be expressed at adjacent positions in the feature space, and the characteristic of the dropped foreground mask and the characteristic of the background mask may be expressed as far apart as possible. .

4 is a diagram illustrating a configuration of an apparatus for detecting an object position according to an embodiment of the disclosed technology. Referring to FIG. 4 , the object position detecting device 400 includes an input device 410 , a storage device 420 , and an arithmetic device 430 . The object position detection device 400 may be a computing device that has a program or algorithm mounted therein to perform a specific calculation, and outputs data of a specific value by performing calculations based on input data. For example, a device such as a PC or a smartphone may be used.

The input device 410 receives an image including an object. The input device may be mounted on or connected to the object position detection device, and may receive an image input by a user as data. For example, a keyboard or mouse may be used as an input device.

The storage device 420 stores a deep learning model trained to detect an object position with respect to an image. The storage device corresponds to the memory of the object position detection device, and the stored deep learning model may be called under the control of the computing device.

The calculator 430 extracts a feature map for the image by using the first convolutional layer of the deep learning model. Then, a first attention map is generated by setting an attention block in the feature map. Then, a dropped foreground mask is generated from the first attention map. As described above, the dropped foreground mask is generated based on a contrastive attention loss. The computing unit 430 estimates the object location based on the dropped foreground mask or an Importance Map generated from the first attention map. The arithmetic unit may be a CPU or processor capable of calculating a contrasting attention loss and a foreground consistency loss. The computing device may generate output data by performing an operation for object position detection using the input data transmitted from the input device and the deep learning model stored in the storage device. Here, the input data may be an image and an annotation including an object, and the output data may be a position or coordinates of an object included in the image.

Meanwhile, if necessary, an output device capable of outputting the result processed by the arithmetic unit 430 may be further provided. The output device may be a display for outputting the object position detection result transmitted from the computing device. The output device may be a display connected to the object position detection device as one, or a monitor connected separately.

5 is a flowchart of a method for detecting an object position according to an embodiment of the disclosed technology. Referring to FIG. 5 , the apparatus for detecting an object location may perform an object location detecting method according to the following procedure.

First, according to step 510, the object position detecting apparatus extracts a feature map for the input image by using the first convolutional layer. The first convolutional layer refers to an initial layer among the plurality of convolutional layers. The object position detecting apparatus may receive an image input by a user and input the image to the first convolutional layer. In addition, the first convolutional layer may extract a feature map for an object included in the image.

Second, in step 520 , the object position detecting apparatus generates a first attention map by using the feature map. The object position detecting apparatus generates a first attention map for the first convolutional layer by setting an attention block in the feature map.

Third, in step 530 , the object position detecting apparatus generates a dropped foreground mask based on the first attention map. The dropped foreground mask means that it serves as a weight to make the result value for the contrastive attention loss small. The object position detecting apparatus may generate three masks based on the first attention map generated from the input image. As an example, a foreground mask generated by covering the background, a background mask generated by covering the foreground, and a dropped foreground mask generated by covering a region most differentiated between the foreground and the background may be generated. In addition, while the dropped foreground mask is located in a space close to the embedding result of the foreground mask, the dropped foreground mask may be generated so as to be far from the embedding result of the background mask.

Next, in step 540, the location of the object is estimated based on the dropped foreground mask or the important map. The position of the object may be calculated based on the position of the bounding box created around the object. However, the object position detecting apparatus may use the dropped foreground mask so that it is not excessively large while not being biased toward a specific area of the object and including the entire object.

Meanwhile, when the operation in the first convolutional layer is finished as described above, the object position detecting apparatus inputs the weighted object image to the second convolutional layer in step 550 . The weighted object image is generated by applying a weight to a randomly selected one of a dropped foreground mask and an importance map to the first attention map.

In operation 560, the object position detecting apparatus generates a second attention map by using the second convolutional layer. A second attention map may be generated by inputting a weighted image to the second convolutional layer and applying an attention block to the image in the same manner as in the first convolutional layer.

In operation 570, the object location detection apparatus learns the deep learning model so that the first attention map and the second attention map are similar to each other. As an embodiment, by calculating the foreground consistency loss, it is possible to learn such that the first attention map output from the first convolutional layer and the second attention map output from the second convolutional layer have similar values.

6 is a diagram for comparing results of object position estimation. Referring to FIG. 6 , a green box indicates a predicted value, and a red box indicates a value for an area of an actual object. As a result of comparing the conventional Autonomos Deep Learning (ADL) and Self Producted Guidance (SPG) with the object position detection technique according to an embodiment of the disclosed technology, the size of the green box due to the attempt to activate more in the less differentiated part in ADL It can be seen that is excessively expanded. And although SPG tries to suppress the background, it evaluates the object area to an excessive size, so you can see that the size of some green boxes is still set large. On the other hand, in the disclosed object position detection technology, it can be seen that the green box is formed in a shape similar to the red box so as to cover the entire range of the object without being excessively extended to the background area.

7 is a diagram illustrating application of the disclosed technology to a conventional image data set. Referring to FIG. 7 , a region marked in red indicates a high importance region, and a region indicated in green indicates a region of low importance. The high-importance area means the area containing the object, and the low-importance area means the background. As a result of testing through the conventional ImageNet and CUB datasets, it can be confirmed that the green box and the red box generated by the model to which the disclosed technology is applied are similarly formed. Accordingly, by estimating the location of the object based on the non-regional attention map, there is an effect of covering the entire area without being biased toward a specific area.

The weak supervised learning-based object position detection method and apparatus according to an embodiment of the disclosed technology has been described with reference to the embodiment shown in the drawings to help understanding, but this is only an example, and those of ordinary skill in the art It will be understood that various modifications and equivalent other embodiments are possible therefrom. Accordingly, the true technical protection scope of the disclosed technology should be defined by the appended claims.

Claims (19)

extracting, by the object position detection apparatus, a feature map for the input image using the first convolutional layer of the deep learning model;
generating, by the object position detection apparatus, a first attention map using the feature map;
generating, by the object position detecting apparatus, a dropped foreground mask from the first attention map, wherein the dropped foreground mask is generated based on a contrastive attention loss; and
and estimating, by the object location detection device, the location of the object based on the dropped foreground mask or an Importance Map generated from the first attention map. The method of claim 1, wherein generating the first attention map comprises:
An object position detection method in which the object position detection apparatus generates the first attention map by setting an attention block in the feature map. The method of claim 1, wherein generating the dropped foreground mask comprises:
generating, by the object position detection apparatus, a foreground mask for the first attention map;
generating, by the object position detection apparatus, a background mask for the first attention map; and
Embedding the first attention map into a feature space using the dropped foreground mask, the foreground mask, and the background mask,
generating the dropped foreground mask so that the dropped foreground mask and the embedding results of the foreground mask are as similar to each other as possible, and the dropped foreground mask and the embedding results of the background mask are as dissimilar to each other as possible. An object position detection method comprising a. The method of claim 1,
The dropped foreground mask is a mask generated by applying a first threshold value to the first attention map to cover the most discriminative region among the background and the foreground. According to claim 1, wherein the object position detection method,
generating, by the object position detection apparatus, a second attention map using a second convolutional layer; and
and estimating, by the object position detection apparatus, the object position based on a second attention map. 6. The method of claim 5,
The second attention map is an object location detection method generated by applying a weight to the first attention map, the dropped foreground mask or a randomly selected one of the important map. The method of claim 1,
The dropped foreground mask is generated by applying a threshold value to the first attention map, and the important map is generated by applying a sigmoid function to the first attention map. The method of claim 5, wherein the object position detection method comprises:
and calculating, by the object position detection apparatus, a foreground consistency loss such that the first attention map and the second attention map are similar to each other. The method of claim 8, wherein the foreground consistency loss is:
An object position detection method for reducing a background activation of a first convolutional layer that has generated the first attention map based on a background activation of the second convolutional layer. The method of claim 8, wherein the object position detection method comprises:
A final loss of the deep learning model is calculated by summing the contrasting attention loss, the foreground consistency loss, and a classification loss for the first attention map, wherein the contrastive attention loss and the The method of detecting an object position further comprising the step of learning the deep learning model to minimize a foreground consistency loss. an input device for receiving an image including an object;
a storage device for storing the deep learning model; and
Extracting a feature map for the image using the first convolutional layer of the deep learning model, and setting an attention block in the feature map to generate a first attention map and generate a dropped foreground mask from the first attention map, wherein the dropped foreground mask is generated based on a contrastive attention loss,
and an arithmetic unit estimating an object position based on the dropped foreground mask or an importance map generated from the first attention map. 12. The method of claim 11,
The computing device generates a foreground mask and a background mask for the first attention map, and uses the dropped foreground mask, the foreground mask, and the background mask to generate the second 1 Embed the attention map into the feature space,
An object for generating the dropped foreground mask so that the dropped foreground mask and the embedding result of the foreground mask are as similar to each other as possible, and the dropped foreground mask and the embedding result of the background mask are as dissimilar as possible to each other position detection device. 12. The method of claim 11,
The dropped foreground mask is a mask generated by applying a first threshold value to the first attention map to cover the most discriminative region among the background and the foreground. 12. The method of claim 11,
The computing device generates a second attention map by using a second convolutional layer, and estimates an object location based on the second attention map. 15. The method of claim 14,
The second attention map is an object location detecting apparatus generated by applying a weight to the first attention map, the dropped foreground mask or the randomly selected one of the important map. 12. The method of claim 11,
The dropped foreground mask is generated by applying a threshold value to the first attention map, and the important map is generated by applying a sigmoid function to the first attention map. 15. The method of claim 14,
The computing device calculates a foreground consistency loss so that the first attention map and the second attention map are similar to each other. 18. The method of claim 17, wherein the foreground consistency loss is:
An apparatus for detecting an object position for reducing a background activation of a first convolutional layer that has generated the first attention map based on a background activation of the second convolutional layer. 18. The method of claim 17,
The calculating unit calculates the final loss of the deep learning model by summing the contrasting attention loss, the foreground consistency loss, and the classification loss for the first attention map, the contrastive attention Object position detection apparatus for learning the deep learning model to minimize the loss and the foreground consistency loss.

Images