CN114332489A - Image salient target detection method and system based on uncertainty perception - Google Patents
Image salient target detection method and system based on uncertainty perception Download PDFInfo
- Publication number
- CN114332489A CN114332489A CN202210249631.1A CN202210249631A CN114332489A CN 114332489 A CN114332489 A CN 114332489A CN 202210249631 A CN202210249631 A CN 202210249631A CN 114332489 A CN114332489 A CN 114332489A
- Authority
- CN
- China
- Prior art keywords
- uncertainty
- features
- contour
- image
- characteristic
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention provides an image salient target detection method and system based on uncertainty perception, and the method comprises the following steps: calculating an inner contour uncertainty image and an outer contour uncertainty image through a saliency map in the image; extracting and obtaining an inner contour uncertainty characteristic of the detail, an outer contour uncertainty characteristic of the detail and a significance characteristic through a characteristic encoder; fusing the salient features and the contour uncertainty features of the details to generate complementary features, calculating through a dynamic weighting branch to obtain dynamic weights, distributing, and fusing the salient features through addition operation to obtain refined salient features; inputting the refined significant features and the detailed contour uncertainty features into a decoding branch, and circularly outputting corresponding features to finally obtain a contour uncertainty prediction graph and a significant prediction graph. The method has good effectiveness and superiority in the aspect of detecting the salient object of the image, and has better performance in processing a complex scene.
Description
Technical Field
The invention relates to the technical field of computer image processing, in particular to a method and a system for detecting an image salient target based on uncertainty perception.
Background
Today, with the rapid development of informatization, people are exposed to a large amount of image information every day, and people often only care about a small amount of information which is useful for themselves in images. Therefore, it is extremely important and meaningful to extract a portion attracting attention of people from an image efficiently and quickly. The salient object detection is one of the most basic and important research contents in the image processing field, can be used for quickly positioning a region concerned by people and extracting salient information from mass data, and can also be used as a preprocessing step of computer vision tasks such as semantic segmentation, object tracking, image description and the like. Meanwhile, the saliency target detection is widely applied to the artificial intelligence fields of human eye detection, intelligent video monitoring, automobile auxiliary driving systems and the like.
Under the influence of deep learning techniques, Significant Object Detection (SOD) research has made significant progress in the field of computer vision. However, the "uncertainty" problem remains an obstacle to SOD development. In particular, most SOD models are prone to false predictions when facing pixels around (inside or outside) the contours of a salient object, assigning similar saliency probabilities to the pixels of that region. To solve this problem, some existing research methods introduce contour or detail maps as supervisory signals to train the SOD model to make the predicted salient objects visually clearer.
However, the prior art methods only consider the target contour or salient pixels near the target contour and ignore the non-salient pixels near the target contour, which makes them susceptible to the non-salient pixels near the contour. Furthermore, the underutilization of contour information and saliency information also makes prediction of contours difficult.
Disclosure of Invention
In view of the above situation, the main objective of the present invention is to provide a method and a system for detecting an image salient object based on uncertainty perception, so as to solve the above technical problems.
The embodiment of the invention provides an image salient target detection method based on uncertainty perception, wherein the method comprises the following steps:
step one, calculating an inner contour uncertainty image and an outer contour uncertainty image through a saliency image in an image, wherein the saliency image represents a salient target in the image, the inner contour uncertainty image represents salient pixel points near a target contour, the outer contour uncertainty image represents non-salient pixel points near the target contour, and the inner contour uncertainty image and the outer contour uncertainty image are used as binarization labels during iterative calculation;
inputting the image into a feature encoder, and extracting an internal contour uncertainty feature of the detail, an external contour uncertainty feature of the detail and a significant feature through the feature encoder;
fusing the salient features and the detail inner contour uncertainty features through an attention mask to generate first complementary features, fusing the salient features and the detail outer contour uncertainty features through a reverse attention mask to generate second complementary features, respectively carrying out weight distribution on the first complementary features and the second complementary features after dynamic weights are obtained through dynamic weighting branch calculation, and fusing the first complementary features and the second complementary features through addition operation to obtain refined salient features;
inputting the refined significant features and the detailed internal contour uncertainty features into an internal contour decoding branch to output and obtain internal contour uncertainty features; inputting the refined significant features and the outer contour uncertainty features of the details into an outer contour decoding branch to output the outer contour uncertainty features; inputting the refined salient features into a salient decoding branch, and obtaining the salient features of the next stage through two convolution operations;
and step five, circulating the extraction output process corresponding to the internal contour uncertainty characteristic, the external contour uncertainty characteristic and the next-stage saliency characteristic in the step four or more for four times to gradually refine the internal contour uncertainty characteristic, the external contour uncertainty characteristic and the next-stage saliency characteristic respectively to finally obtain an internal contour uncertainty characteristic prediction graph, an external contour uncertainty characteristic prediction graph and a saliency prediction graph.
According to the image salient target detection method based on uncertainty perception, an inner contour uncertainty image and an outer contour uncertainty image are obtained through calculation of a salient image in an image; extracting and obtaining an inner contour uncertainty characteristic of the detail, an outer contour uncertainty characteristic of the detail and a significance characteristic through a characteristic encoder; fusing the salient features and the contour uncertainty features of the details to generate complementary features, calculating through a dynamic weighting branch to obtain dynamic weights, distributing, and fusing the salient features through addition operation to obtain refined salient features; inputting the refined significant features and the detailed contour uncertainty features into a decoding branch, and circularly outputting corresponding features to finally obtain a contour uncertainty prediction graph and a significant prediction graph. The method has good effectiveness and superiority in the aspect of detecting the salient object of the image, and has better performance in processing a complex scene.
In the image salient object detection method based on uncertainty perception, in the first step, the calculation formula of the inner contour uncertainty image and the outer contour uncertainty image obtained through calculation of the saliency map in the image is represented as follows:
wherein the content of the first and second substances,a saliency map representing the image is shown,representing the internal contour uncertainty map in a manner that,representing the outer contour uncertainty map,the corrosion function is expressed as a function of corrosion,the function of the expansion is represented by,which represents the size of the kernel for the etching operation,representing the kernel size of the dilation operation.
The image salient object detection method based on uncertainty perception comprises the following two specific steps:
inputting the image into a ResNet-50 network to generate features of five scales;
using two 64 channels for the five scale featuresThe convolution operation and a spatial attention module to obtain an inner contour uncertainty characteristic of the detail and an outer contour uncertainty characteristic of the detail respectively;
based on the characteristics of the five scales, the significant characteristics are obtained by extracting two convolution layers。
The image salient object detection method based on uncertainty perception is characterized in that the characteristics of the five scales are represented as follows:;
the spatial attention module is calculated in the following way:
wherein the content of the first and second substances,the spatial attention module is represented as a function of the spatial attention module,a feature representing the input is presented to the user,to representThe amount of the above-mentioned convolution layer is,a sigmoid function is represented as a function,representing a multiplication operation at the element level,an internal contour uncertainty characteristic representing the detail,showing detailsAn outer contour uncertainty feature.
The image salient object detection method based on uncertainty perception comprises the following steps of fusing a salient feature and an internal contour uncertainty feature of a detail through an attention mask to generate a first complementary feature:
connecting the salient features with the internal contour uncertainty features of the detail to obtain first connected features, usingThe convolution operation of (a) compresses the first connected feature to obtain a first compressed feature;
Applying the first compression characteristicInputting into two convolution layers, and generating attention mask by sigmoid function;
Enhancing complementary information between the salient features and the internal contour uncertainty features of the detail by using multiplication to obtain the first complementary features;
the corresponding calculation formula is expressed as:
wherein the content of the first and second substances,the uncertainty characteristics of the internal contour are represented,a first compression characteristic is represented by a first compression characteristic,to representThe convolution operation of (a) is performed,to representThe convolution operation of (a) is performed,it is shown that the connection operation is performed,an attention mask is shown in which the position of the eye,the salient features are represented by a representation of,representing the first complementary feature.
The image salient object detection method based on uncertainty perception comprises the following steps of fusing salient features and external contour uncertainty features of details through a reverse attention mask to generate second complementary features:
connecting the salient features with the outer contour uncertainty features of the detail to obtain second connected features, usingThe convolution operation of (a) compresses the second connected feature to obtain a second compressed feature;
Subtracting each pixel value in the generated attention mask with 1 to obtain a reverse attention mask;
Enhancing complementary information between the salient features and the outer contour uncertainty features of the detail using multiplication operations to obtain the second complementary features;
the corresponding calculation formula is expressed as:
wherein the content of the first and second substances,the uncertainty characteristics of the outer contour are represented,the second compression characteristic is represented by a second set of values,the reverse attention mask is shown with the intent of,represents the aboveA second complementary feature.
The image salient object detection method based on uncertainty perception is characterized in that a calculation formula of dynamic weight obtained through dynamic weighting branch calculation is represented as follows:
wherein the content of the first and second substances,the dynamic weight is represented by a weight value representing the dynamic weight,which represents the average pooling layer, is,a fully-connected layer is shown,it is indicated that the softmax function is,indicating the ReLU activation function.
The image salient object detection method based on uncertainty perception comprises the following steps of:
first output characteristic of previous layer through up-sampling layerUpsampling to obtain internal profile uncertainty features associated with the detailSame size sampled first output characteristic;
Pairing the salient features by an upsampling layerUpsampling to obtain internal profile uncertainty features associated with the detailSame size post-sampling first salient feature;
To the first output characteristic after samplingAnd the first significant feature after samplingPerforming multiplication and addition to obtain a decoded first intermediate feature;
According to the decoded first intermediate featureInternal profile uncertainty characterization from said detailPerforming a join and convolution operation to obtain the internal contour uncertainty feature prediction graph;
The corresponding calculation formula is expressed as:
wherein the content of the first and second substances,it is indicated that the up-sampling operation,representing an element-level addition operation.
The image salient object detection method based on uncertainty perception comprises the following steps of:
second output characteristic of previous layer by up-sampling layerUpsampling to obtain an outer contour uncertainty characteristic associated with the detailSame size of sampled second output characteristic;
Pairing the salient features by an upsampling layerUpsampling to obtain an outer contour uncertainty characteristic associated with the detailSame size post-sampling second salient feature;
To the second output characteristic after samplingAnd second significant feature after samplingPerforming multiplication and addition to obtain a decoded second intermediate feature;
According to the decoded second intermediate characteristicExternal profile uncertainty characterization of said detailPerforming a join and convolution operation to obtain the outer contour uncertainty feature prediction map;
The corresponding calculation formula is expressed as:
the invention also provides an image salient object detection system based on uncertainty perception, wherein the system comprises:
the image processing device comprises a first calculation module, a second calculation module and a third calculation module, wherein the first calculation module is used for calculating an inner contour uncertainty image and an outer contour uncertainty image through a saliency image in an image, the saliency image represents a salient object in the image, the inner contour uncertainty image represents salient pixel points near a contour of the object, the outer contour uncertainty image represents non-salient pixel points near the contour of the object, and the inner contour uncertainty image and the outer contour uncertainty image are used as binarization labels during iterative calculation;
the characteristic extraction module is used for inputting the image into a characteristic encoder, and extracting and obtaining an inner contour uncertainty characteristic of the detail, an outer contour uncertainty characteristic of the detail and a significance characteristic through the characteristic encoder;
the second calculation module is used for fusing the salient features and the internal contour uncertainty features of the details through an attention mask to generate first complementary features, fusing the salient features and the external contour uncertainty features of the details through a reverse attention mask to generate second complementary features, respectively carrying out weight distribution on the first complementary features and the second complementary features after dynamic weights are obtained through dynamic weighting branch calculation, and fusing the first complementary features and the second complementary features through addition operation to obtain refined salient features;
a feature output module, configured to input the refined significant features and the internal contour uncertainty features of the details into an internal contour decoding branch to output an internal contour uncertainty feature; inputting the refined significant features and the outer contour uncertainty features of the details into an outer contour decoding branch to output the outer contour uncertainty features; inputting the refined salient features into a salient decoding branch, and obtaining the salient features of the next stage through two convolution operations;
and the prediction graph determining module is used for circulating the extraction output process corresponding to the internal contour uncertainty characteristic, the external contour uncertainty characteristic and the next-stage saliency characteristic for more than four times so as to gradually refine the internal contour uncertainty characteristic, the external contour uncertainty characteristic and the next-stage saliency characteristic respectively and finally obtain an internal contour uncertainty characteristic prediction graph, an external contour uncertainty characteristic prediction graph and a saliency prediction graph.
Additional aspects and advantages of the invention will be set forth in part in the description which follows and, in part, will be obvious from the description, or may be learned by practice of the invention.
Drawings
FIG. 1 is a flowchart of a method for detecting a salient object in an image based on uncertainty sensing according to the present invention;
fig. 2 is a schematic structural diagram of an image salient object detection system based on uncertainty perception according to the present invention.
Detailed Description
Reference will now be made in detail to embodiments of the present invention, examples of which are illustrated in the accompanying drawings, wherein like or similar reference numerals refer to the same or similar elements or elements having the same or similar function throughout. The embodiments described below with reference to the accompanying drawings are illustrative only for the purpose of explaining the present invention, and are not to be construed as limiting the present invention.
These and other aspects of embodiments of the invention will be apparent with reference to the following description and attached drawings. In the description and drawings, particular embodiments of the invention have been disclosed in detail as being indicative of some of the ways in which the principles of the embodiments of the invention may be practiced, but it is understood that the scope of the embodiments of the invention is not limited correspondingly. On the contrary, the embodiments of the invention include all changes, modifications and equivalents coming within the spirit and terms of the claims appended hereto.
Referring to fig. 1, the present invention provides a method for detecting an image salient object based on uncertainty perception, wherein the method includes the following steps:
and S101, calculating an inner contour uncertainty image and an outer contour uncertainty image through a saliency map in the image.
The salient map represents a salient target in an image, the inner contour uncertainty map represents salient pixel points near a target contour, the outer contour uncertainty map represents non-salient pixel points near the target contour, and the inner contour uncertainty map and the outer contour uncertainty map are used as binarization labels in iterative computation.
In the invention, in order to relieve the uncertainty problem of pixel significance probability prediction around a significant target contour, three supervision signals of an internal contour uncertainty image, a significant image of an image and an external contour uncertainty image are introduced for joint supervision so as to guide model learning. Wherein the saliency map focuses on salient objects in the image, and the inner contour uncertainty map and the outer contour uncertainty map focus on pixels around the contours of the salient objects.
Specifically, the calculation formula of the inner contour uncertainty map and the outer contour uncertainty map obtained by calculating the saliency map in the image is expressed as follows:
wherein the content of the first and second substances,a saliency map representing the image is shown,representing the internal contour uncertainty map in a manner that,representing the outer contour uncertainty map,indicating corrosion contentThe number of the first and second groups is,the function of the expansion is represented by,which represents the size of the kernel for the etching operation,representing the kernel size of the dilation operation.
S102, inputting the image into a feature encoder, and extracting the internal contour uncertainty feature, the external contour uncertainty feature and the saliency feature of the detail through the feature encoder.
Specifically, step S102 includes the following substeps:
and S1021, inputting the image into a ResNet-50 network to generate characteristics of five scales.
S1022, for the characteristics of the five scales, the method utilizes two 64 channelsAnd a spatial attention module to obtain an inner contour uncertainty characteristic of the detail and an outer contour uncertainty characteristic of the detail, respectively.
S1023, extracting the salient features through two convolution layers based on the features of the five scales。
the spatial attention module is calculated in the following way:
wherein the content of the first and second substances,the spatial attention module is represented as a function of the spatial attention module,a feature representing the input is presented to the user,to representThe amount of the above-mentioned convolution layer is,a sigmoid function is represented as a function,representing a multiplication operation at the element level,an internal contour uncertainty characteristic representing the detail,an outer contour uncertainty characteristic representing the detail.
S103, fusing the salient features and the internal contour uncertainty features of the details through an attention mask to generate first complementary features, fusing the salient features and the external contour uncertainty features of the details through a reverse attention mask to generate second complementary features, respectively carrying out weight distribution on the first complementary features and the second complementary features after dynamic weights are obtained through dynamic weighting branch calculation, and fusing the first complementary features and the second complementary features through addition operation to obtain refined salient features.
Wherein the method of fusing the salient features and the internal contour uncertainty features of the detail through the attention mask to generate first complementary features comprises the steps of:
s1031a, connecting the significant features and the internal contour uncertainty features of the details to obtain first connected features, and using the first connected featuresThe convolution operation of (a) compresses the first connected feature to obtain a first compressed feature。
S1031b, compressing the first compression characteristicInputting into two convolution layers, and generating attention mask by sigmoid function。
And S1031c, enhancing complementary information between the significant features and the internal contour uncertainty features of the details by utilizing multiplication operation to obtain the first complementary features.
The corresponding calculation formula is expressed as:
wherein the content of the first and second substances,the uncertainty characteristics of the internal contour are represented,a first compression characteristic is represented by a first compression characteristic,to representThe convolution operation of (a) is performed,to representThe convolution operation of (a) is performed,it is shown that the connection operation is performed,an attention mask is shown in which the position of the eye,the salient features are represented by a representation of,representing the first complementary feature.
Further, the method of fusing the salient features and the outer contour uncertainty features of the detail through the inverse attention mask to generate second complementary features comprises the steps of:
s1032a, distinguishing the significanceConnecting the feature with the outer contour uncertainty feature of the detail to obtain a second connected feature, usingThe convolution operation of (a) compresses the second connected feature to obtain a second compressed feature。
S1032b, subtracting each pixel value in the generated attention mask by 1 to obtain a reverse attention mask。
S1032c, enhancing complementary information between the salient feature and the outer contour uncertainty feature of the detail by utilizing multiplication operation to obtain the second complementary feature.
The corresponding calculation formula is expressed as:
wherein the content of the first and second substances,the uncertainty characteristics of the outer contour are represented,the second compression characteristic is represented by a second set of values,the reverse attention mask is shown with the intent of,representing the second complementary feature.
In this embodiment, the calculation formula of the dynamic weight obtained by the dynamic weighting branch calculation is represented as:
wherein the content of the first and second substances,the dynamic weight is represented by a weight value representing the dynamic weight,which represents the average pooling layer, is,a fully-connected layer is shown,it is indicated that the softmax function is,indicating the ReLU activation function.
As supplementary explanation, after the dynamic weight is obtained by calculation, weights are assigned to the salient features, the inner contour uncertainty features and the outer contour uncertainty features to extract complementary features between the contour uncertainty features and the salient features, and the salient features are fused by addition operation to refine and enhance the salient features.
S104, inputting the refined significant features and the detailed internal contour uncertainty features into an internal contour decoding branch to output and obtain internal contour uncertainty features; inputting the refined significant features and the outer contour uncertainty features of the details into an outer contour decoding branch to output the outer contour uncertainty features; inputting the refined salient features into a salient decoding branch, and obtaining the salient features of the next stage through two convolution operations.
In this step, the method for generating the prediction graph of uncertainty characteristics of the internal contour includes the following steps:
s1041a, first output characteristic of previous layer through up-sampling layerUpsampling to obtain internal profile uncertainty features associated with the detailSame size sampled first output characteristic;
S1041b pairing the salient features by an upsampling layerUpsampling to obtain internal profile uncertainty features associated with the detailSame size post-sampling first salient feature;
S1041c, sampling the first output characteristicAnd the first significant feature after samplingPerforming multiplication and addition to obtain a decoded first intermediate feature;
S1041d, according to the decoded first intermediate characteristicInternal profile uncertainty characterization from said detailPerforming a join and convolution operation to obtain the internal contour uncertainty feature prediction graph;
The corresponding calculation formula is expressed as:
wherein the content of the first and second substances,it is indicated that the up-sampling operation,representing an element-level addition operation.
Further, the method for generating the external contour uncertainty characteristic prediction map comprises the following steps:
s1042a, throughSecond output characteristic of sampling layer to previous layerUpsampling to obtain an outer contour uncertainty characteristic associated with the detailSame size of sampled second output characteristic;
S1042b, comparing the significant features through an upsampling layerUpsampling to obtain an outer contour uncertainty characteristic associated with the detailSame size post-sampling second salient feature;
S1042c, second output characteristic after samplingAnd second significant feature after samplingPerforming multiplication and addition to obtain a decoded second intermediate feature;
S1042d, according to the decoded second intermediate characteristicExternal profile uncertainty characterization of said detailPerforming a join and convolution operation to obtain the outer contour uncertainty feature prediction map;
The corresponding calculation formula is expressed as:
and S105, circulating the extraction output process corresponding to the internal contour uncertainty characteristic, the external contour uncertainty characteristic and the next-stage saliency characteristic for more than four times to respectively refine the internal contour uncertainty characteristic, the external contour uncertainty characteristic and the next-stage saliency characteristic step by step so as to finally obtain an internal contour uncertainty characteristic prediction graph, an external contour uncertainty characteristic prediction graph and a saliency prediction graph.
In this embodiment, for the final output inner contour uncertainty prediction map, the significant prediction map and the outer contour uncertainty prediction map, a loss function corresponding to each is used, and an overall training loss function is combined to optimize the prediction. The significance prediction graph is calculated by using loss binarization cross entropy and an intersection ratio, and a loss function corresponding to the internal contour uncertainty characteristic prediction graph and a loss function corresponding to the external contour uncertainty prediction graph are calculated by using the intersection ratio.
wherein the content of the first and second substances,a normalized prediction graph is represented that represents the predicted graph,a label representing the binary value is attached to the label,andrespectively represent the width and height corresponding to the binary label,pixel coordinate values representing the binarized label,pixel coordinate values representing the prediction map,indicating the coordinate position.
In addition, the overall training loss function is a weighted sum of losses corresponding to the internal contour uncertainty prediction graph, the significant prediction graph, and the external contour uncertainty prediction graph, and is specifically expressed as:
wherein the content of the first and second substances,the training loss function as a whole is represented,a function representing the significant loss is represented by,the internal profile loss function is represented as,the outer contour loss function is represented as,andare all weights used to balance the losses.
According to the image salient target detection method based on uncertainty perception, an inner contour uncertainty image and an outer contour uncertainty image are obtained through calculation of a salient image in an image; extracting and obtaining an inner contour uncertainty characteristic of the detail, an outer contour uncertainty characteristic of the detail and a significance characteristic through a characteristic encoder; fusing the salient features and the contour uncertainty features of the details to generate complementary features, calculating through a dynamic weighting branch to obtain dynamic weights, distributing, and fusing the salient features through addition operation to obtain refined salient features; inputting the refined significant features and the detailed contour uncertainty features into a decoding branch, and circularly outputting corresponding features to finally obtain a contour uncertainty prediction graph and a significant prediction graph. The method has good effectiveness and superiority in the aspect of detecting the salient object of the image, and has better performance in processing a complex scene.
The invention has the advantages that:
(1) guiding the proposed model to pay attention to the detection of the significant target by jointly supervising three labels of the internal contour uncertainty image, the significant image and the external contour uncertainty image, and transferring partial attention to 'uncertainty' pixels near the contour of the significant target;
(2) a new feature interaction module is provided to enhance complementary feature interaction between the inner contour uncertainty feature, the saliency feature and the outer contour uncertainty feature;
(3) the method for detecting the salient object of the research image is not only beneficial to people to quickly find out useful information from the image, but also beneficial to the development of other computer vision tasks taking salient object detection as a preprocessing step. Therefore, an effective and accurate prediction image salient object detection algorithm can promote the development of image processing.
Referring to fig. 2, the present invention further provides a system for detecting an image salient object based on uncertainty perception, wherein the system includes:
the image processing device comprises a first calculation module, a second calculation module and a third calculation module, wherein the first calculation module is used for calculating an inner contour uncertainty image and an outer contour uncertainty image through a saliency image in an image, the saliency image represents a salient object in the image, the inner contour uncertainty image represents salient pixel points near a contour of the object, the outer contour uncertainty image represents non-salient pixel points near the contour of the object, and the inner contour uncertainty image and the outer contour uncertainty image are used as binarization labels during iterative calculation;
the characteristic extraction module is used for inputting the image into a characteristic encoder, and extracting and obtaining an inner contour uncertainty characteristic of the detail, an outer contour uncertainty characteristic of the detail and a significance characteristic through the characteristic encoder;
the second calculation module is used for fusing the salient features and the internal contour uncertainty features of the details through an attention mask to generate first complementary features, fusing the salient features and the external contour uncertainty features of the details through a reverse attention mask to generate second complementary features, respectively carrying out weight distribution on the first complementary features and the second complementary features after dynamic weights are obtained through dynamic weighting branch calculation, and fusing the first complementary features and the second complementary features through addition operation to obtain refined salient features;
a feature output module, configured to input the refined significant features and the internal contour uncertainty features of the details into an internal contour decoding branch to output an internal contour uncertainty feature; inputting the refined significant features and the outer contour uncertainty features of the details into an outer contour decoding branch to output the outer contour uncertainty features; inputting the refined salient features into a salient decoding branch, and obtaining the salient features of the next stage through two convolution operations;
and the prediction graph determining module is used for circulating the extraction output process corresponding to the internal contour uncertainty characteristic, the external contour uncertainty characteristic and the next-stage saliency characteristic for more than four times so as to gradually refine the internal contour uncertainty characteristic, the external contour uncertainty characteristic and the next-stage saliency characteristic respectively and finally obtain an internal contour uncertainty characteristic prediction graph, an external contour uncertainty characteristic prediction graph and a saliency prediction graph.
It should be understood that portions of the present invention may be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, the various steps or methods may be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any one or combination of the following techniques, which are known in the art, may be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application specific integrated circuit having an appropriate combinational logic gate circuit, a Programmable Gate Array (PGA), a Field Programmable Gate Array (FPGA), or the like.
In the description herein, references to the description of the term "one embodiment," "some embodiments," "an example," "a specific example," or "some examples," etc., mean that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above do not necessarily refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.
The above-mentioned embodiments only express several embodiments of the present invention, and the description thereof is more specific and detailed, but not construed as limiting the scope of the present invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the inventive concept, which falls within the scope of the present invention. Therefore, the protection scope of the present patent shall be subject to the appended claims.
Claims (10)
1. An image salient object detection method based on uncertainty perception is characterized by comprising the following steps: step one, calculating an inner contour uncertainty image and an outer contour uncertainty image through a saliency image in an image, wherein the saliency image represents a salient target in the image, the inner contour uncertainty image represents salient pixel points near a target contour, the outer contour uncertainty image represents non-salient pixel points near the target contour, and the inner contour uncertainty image and the outer contour uncertainty image are used as binarization labels during iterative calculation;
inputting the image into a feature encoder, and extracting an internal contour uncertainty feature of the detail, an external contour uncertainty feature of the detail and a significant feature through the feature encoder;
fusing the salient features and the detail inner contour uncertainty features through an attention mask to generate first complementary features, fusing the salient features and the detail outer contour uncertainty features through a reverse attention mask to generate second complementary features, respectively carrying out weight distribution on the first complementary features and the second complementary features after dynamic weights are obtained through dynamic weighting branch calculation, and fusing the first complementary features and the second complementary features through addition operation to obtain refined salient features;
inputting the refined significant features and the detailed internal contour uncertainty features into an internal contour decoding branch to output and obtain internal contour uncertainty features; inputting the refined significant features and the outer contour uncertainty features of the details into an outer contour decoding branch to output the outer contour uncertainty features; inputting the refined salient features into a salient decoding branch, and obtaining the salient features of the next stage through two convolution operations;
and step five, circulating the extraction output process corresponding to the internal contour uncertainty characteristic, the external contour uncertainty characteristic and the next-stage saliency characteristic in the step four or more for four times to gradually refine the internal contour uncertainty characteristic, the external contour uncertainty characteristic and the next-stage saliency characteristic respectively to finally obtain an internal contour uncertainty characteristic prediction graph, an external contour uncertainty characteristic prediction graph and a saliency prediction graph.
2. The method for detecting the image salient objects based on the uncertainty perception according to claim 1, wherein in the step one, the calculation formula of the inner contour uncertainty map and the outer contour uncertainty map obtained by the calculation of the saliency map in the image is represented as follows:
wherein the content of the first and second substances,a saliency map representing the image is shown,representing the internal contour uncertainty map in a manner that,representing the outer contour uncertainty map,the corrosion function is expressed as a function of corrosion,the function of the expansion is represented by,which represents the size of the kernel for the etching operation,representing the kernel size of the dilation operation.
3. The image salient object detection method based on uncertainty perception according to claim 2, wherein the second step specifically comprises:
inputting the image into a ResNet-50 network to generate features of five scales;
using two 64 channels for the five scale featuresThe convolution operation and a spatial attention module to obtain an inner contour uncertainty characteristic of the detail and an outer contour uncertainty characteristic of the detail respectively;
4. The image salient object detection method based on uncertainty perception according to claim 3, wherein the five scales of features are represented as:;
the spatial attention module is calculated in the following way:
wherein the content of the first and second substances,the spatial attention module is represented as a function of the spatial attention module,a feature representing the input is presented to the user,to representThe amount of the above-mentioned convolution layer is,a sigmoid function is represented as a function,representing a multiplication operation at the element level,an internal contour uncertainty characteristic representing the detail,an outer contour uncertainty characteristic representing the detail.
5. The image salient object detection method based on uncertainty perception according to claim 4, wherein the method for fusing the salient features and the internal contour uncertainty features of the details through the attention mask to generate the first complementary features comprises the following steps:
connecting the salient features with the internal contour uncertainty features of the detail to obtain first connected features, usingThe convolution operation of (a) compresses the first connected feature to obtain a first compressed feature;
Applying the first compression characteristicInputting into two convolution layers, and generating attention mask by sigmoid function;
Enhancing complementary information between the salient features and the internal contour uncertainty features of the detail by using multiplication to obtain the first complementary features;
the corresponding calculation formula is expressed as:
wherein the content of the first and second substances,the uncertainty characteristics of the internal contour are represented,a first compression characteristic is represented by a first compression characteristic,to representThe convolution operation of (a) is performed,to representThe convolution operation of (a) is performed,it is shown that the connection operation is performed,an attention mask is shown in which the position of the eye,the salient features are represented by a representation of,representing the first complementary feature.
6. The image salient object detection method based on uncertainty perception according to claim 5, wherein the method for fusing the salient features and the external contour uncertainty features of the details through the inverse attention mask to generate second complementary features comprises the following steps:
connecting the salient features with the outer contour uncertainty features of the detail to obtain second connected features, usingThe convolution operation of (a) compresses the second connected feature to obtain a second compressed feature;
Subtracting each pixel value in the generated attention mask with 1 to obtain a reverse attention mask;
Enhancing complementary information between the salient features and the outer contour uncertainty features of the detail using multiplication operations to obtain the second complementary features;
the corresponding calculation formula is expressed as:
wherein the content of the first and second substances,the uncertainty characteristics of the outer contour are represented,the second compression characteristic is represented by a second set of values,the reverse attention mask is shown with the intent of,representing the second complementary feature.
7. The image salient object detection method based on uncertainty perception according to claim 6, wherein the calculation formula for obtaining dynamic weight through dynamic weighting branch calculation is represented as:
wherein the content of the first and second substances,the dynamic weight is represented by a weight value representing the dynamic weight,which represents the average pooling layer, is,a fully-connected layer is shown,it is indicated that the softmax function is,indicating the ReLU activation function.
8. The image salient object detection method based on uncertainty perception according to claim 7, wherein the generation method of the internal contour uncertainty feature prediction map comprises the following steps:
first output characteristic of previous layer through up-sampling layerUpsampling to obtain internal profile uncertainty features associated with the detailSame size sampled first output characteristic;
Pairing the salient features by an upsampling layerUpsampling to obtain internal profile uncertainty features associated with the detailSame size post-sampling first salient feature;
To the first output characteristic after samplingAnd the first significant feature after samplingPerforming multiplication and addition to obtain a decoded first intermediate feature;
According to the decoded first intermediate featureInternal profile uncertainty characterization from said detailPerforming a join and convolution operation to obtain the internal contour uncertainty feature prediction graph;
The corresponding calculation formula is expressed as:
9. The image salient object detection method based on uncertainty perception according to claim 7, wherein the generation method of the outer contour uncertainty feature prediction map comprises the following steps:
second output characteristic of previous layer by up-sampling layerUpsampling to obtain an outer contour uncertainty characteristic associated with the detailSame size of sampled second output characteristic;
Pairing the salient features by an upsampling layerUpsampling to obtain an outer contour uncertainty characteristic associated with the detailSame size post-sampling second salient feature;
To the second output characteristic after samplingAnd second significant feature after samplingPerforming multiplication and addition to obtain a decoded second intermediate feature;
According to the decoded second intermediate characteristicExternal profile uncertainty characterization of said detailPerforming a join and convolution operation to obtain the outer contour uncertainty feature prediction map;
The corresponding calculation formula is expressed as:
10. an image salient object detection system based on uncertainty perception, characterized in that the system comprises:
the image processing device comprises a first calculation module, a second calculation module and a third calculation module, wherein the first calculation module is used for calculating an inner contour uncertainty image and an outer contour uncertainty image through a saliency image in an image, the saliency image represents a salient object in the image, the inner contour uncertainty image represents salient pixel points near a contour of the object, the outer contour uncertainty image represents non-salient pixel points near the contour of the object, and the inner contour uncertainty image and the outer contour uncertainty image are used as binarization labels during iterative calculation;
the characteristic extraction module is used for inputting the image into a characteristic encoder, and extracting and obtaining an inner contour uncertainty characteristic of the detail, an outer contour uncertainty characteristic of the detail and a significance characteristic through the characteristic encoder;
the second calculation module is used for fusing the salient features and the internal contour uncertainty features of the details through an attention mask to generate first complementary features, fusing the salient features and the external contour uncertainty features of the details through a reverse attention mask to generate second complementary features, respectively carrying out weight distribution on the first complementary features and the second complementary features after dynamic weights are obtained through dynamic weighting branch calculation, and fusing the first complementary features and the second complementary features through addition operation to obtain refined salient features;
a feature output module, configured to input the refined significant features and the internal contour uncertainty features of the details into an internal contour decoding branch to output an internal contour uncertainty feature; inputting the refined significant features and the outer contour uncertainty features of the details into an outer contour decoding branch to output the outer contour uncertainty features; inputting the refined salient features into a salient decoding branch, and obtaining the salient features of the next stage through two convolution operations;
and the prediction graph determining module is used for circulating the extraction output process corresponding to the internal contour uncertainty characteristic, the external contour uncertainty characteristic and the next-stage saliency characteristic for more than four times so as to gradually refine the internal contour uncertainty characteristic, the external contour uncertainty characteristic and the next-stage saliency characteristic respectively and finally obtain an internal contour uncertainty characteristic prediction graph, an external contour uncertainty characteristic prediction graph and a saliency prediction graph.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210249631.1A CN114332489B (en) | 2022-03-15 | 2022-03-15 | Image salient target detection method and system based on uncertainty perception |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210249631.1A CN114332489B (en) | 2022-03-15 | 2022-03-15 | Image salient target detection method and system based on uncertainty perception |
Publications (2)
Publication Number | Publication Date |
---|---|
CN114332489A true CN114332489A (en) | 2022-04-12 |
CN114332489B CN114332489B (en) | 2022-06-24 |
Family
ID=81033136
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210249631.1A Active CN114332489B (en) | 2022-03-15 | 2022-03-15 | Image salient target detection method and system based on uncertainty perception |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114332489B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115359258A (en) * | 2022-08-26 | 2022-11-18 | 中国科学院国家空间科学中心 | Weak and small target detection method and system for component uncertainty measurement |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2015017941A1 (en) * | 2013-08-09 | 2015-02-12 | Sweep3D Corporation | Systems and methods for generating data indicative of a three-dimensional representation of a scene |
US20160140424A1 (en) * | 2014-11-13 | 2016-05-19 | Nec Laboratories America, Inc. | Object-centric Fine-grained Image Classification |
CN110443257A (en) * | 2019-07-08 | 2019-11-12 | 大连理工大学 | A kind of conspicuousness detection method based on Active Learning |
US20210146531A1 (en) * | 2019-11-20 | 2021-05-20 | Nvidia Corporation | Guided uncertainty-aware policy optimization: combining model-free and model-based strategies for sample-efficient learning |
CN113326886A (en) * | 2021-06-16 | 2021-08-31 | 中山大学 | Salient object detection method and system based on unsupervised learning |
CN114067101A (en) * | 2021-11-05 | 2022-02-18 | 浙江工业大学 | Image significance detection method of double-stream decoder based on information complementation |
CN114078138A (en) * | 2021-11-23 | 2022-02-22 | 中国银行股份有限公司 | Image significance detection method and device |
-
2022
- 2022-03-15 CN CN202210249631.1A patent/CN114332489B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2015017941A1 (en) * | 2013-08-09 | 2015-02-12 | Sweep3D Corporation | Systems and methods for generating data indicative of a three-dimensional representation of a scene |
US20160140424A1 (en) * | 2014-11-13 | 2016-05-19 | Nec Laboratories America, Inc. | Object-centric Fine-grained Image Classification |
CN110443257A (en) * | 2019-07-08 | 2019-11-12 | 大连理工大学 | A kind of conspicuousness detection method based on Active Learning |
US20210146531A1 (en) * | 2019-11-20 | 2021-05-20 | Nvidia Corporation | Guided uncertainty-aware policy optimization: combining model-free and model-based strategies for sample-efficient learning |
CN113326886A (en) * | 2021-06-16 | 2021-08-31 | 中山大学 | Salient object detection method and system based on unsupervised learning |
CN114067101A (en) * | 2021-11-05 | 2022-02-18 | 浙江工业大学 | Image significance detection method of double-stream decoder based on information complementation |
CN114078138A (en) * | 2021-11-23 | 2022-02-22 | 中国银行股份有限公司 | Image significance detection method and device |
Non-Patent Citations (3)
Title |
---|
A. LI等: ""Uncertainty-aware Joint Salient Object and Camouflaged Object Detection"", 《2021 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR)》 * |
YUMING FANG等: ""Visual attention prediction for Autism Spectrum Disorder with hierarchical semantic fusion"", 《SIGNAL PROCESSING: IMAGE COMMUNICATION》 * |
包晓安等: ""基于背景感知的显著性目标检测算法"", 《计算机系统应用》 * |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115359258A (en) * | 2022-08-26 | 2022-11-18 | 中国科学院国家空间科学中心 | Weak and small target detection method and system for component uncertainty measurement |
Also Published As
Publication number | Publication date |
---|---|
CN114332489B (en) | 2022-06-24 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN112990211B (en) | Training method, image processing method and device for neural network | |
KR20220002067A (en) | method, apparatus, computer equipment, computer readable storage medium and computer program for visual qustion answering | |
CN113592060A (en) | Neural network optimization method and device | |
CN115512103A (en) | Multi-scale fusion remote sensing image semantic segmentation method and system | |
CN113901900A (en) | Unsupervised change detection method and system for homologous or heterologous remote sensing image | |
CN116416416A (en) | Training method of virtual fitting model, virtual fitting method and electronic equipment | |
CN114049356B (en) | Method, device and system for detecting structure apparent crack | |
CN113012177A (en) | Three-dimensional point cloud segmentation method based on geometric feature extraction and edge perception coding | |
CN114359289A (en) | Image processing method and related device | |
CN114332489B (en) | Image salient target detection method and system based on uncertainty perception | |
CN117078930A (en) | Medical image segmentation method based on boundary sensing and attention mechanism | |
CN114821096A (en) | Image processing method, neural network training method and related equipment | |
CN112668608A (en) | Image identification method and device, electronic equipment and storage medium | |
CN111652181A (en) | Target tracking method and device and electronic equipment | |
Zuo et al. | A remote sensing image semantic segmentation method by combining deformable convolution with conditional random fields | |
CN113158970B (en) | Action identification method and system based on fast and slow dual-flow graph convolutional neural network | |
CN114037678A (en) | Urine visible component detection method and device based on deep learning | |
CN112884702A (en) | Polyp identification system and method based on endoscope image | |
CN114565092A (en) | Neural network structure determining method and device | |
CN116824291A (en) | Remote sensing image learning method, device and equipment | |
CN116309050A (en) | Image super-resolution method, program product, storage medium and electronic device | |
CN115984093A (en) | Depth estimation method based on infrared image, electronic device and storage medium | |
CN114399648A (en) | Behavior recognition method and apparatus, storage medium, and electronic device | |
CN113256556A (en) | Image selection method and device | |
CN116824308B (en) | Image segmentation model training method and related method, device, medium and equipment |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |