Unsupervised salient object detection method, system, device and medium
Through the pseudo-label generation and contrast clustering methods, combined with attention mechanism and classification network, the problems of training data set bias and pseudo-label noise are solved, and high-precision unsupervised significance object detection is achieved.
Patent Information
- Application Number
- CN202111592536.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-23
- Publication Date
- 2025-08-12
- Estimated Expiration
- 2041-12-23
AI Technical Summary
In the existing significance object detection methods, the training data set has its own bias and pseudo-label noise problems, resulting in a decrease in detection accuracy.
The pseudo-label generation model is used to generate pseudo-labels, the target feature vectors are extracted through the backbone network, and contrast feature screening and clustering learning are performed. Cartesian product calculation and attention mechanism are used, and the classification network is trained, and the loss function is designed to reduce pseudo-label noise.
Under unsupervised conditions, the accuracy and accuracy of significance target detection are improved, and the dependence on manual annotation and the impact of pseudo-label noise is reduced.
Smart Images
Figure CN114266927B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computer vision technology, and in particular to an unsupervised salient object detection method, system, device and medium. Background Art
[0002] When processing visual information, the human eye automatically filters out uninteresting background and leaves the most eye-catching target object. Similarly, the salient object detection algorithm simulates this mechanism with the aim of quickly detecting the most eye-catching target in an image and separating the target from the background. Salient object detection can be used as an important step in image preprocessing and is widely used by various downstream computer vision tasks. Among the existing salient object detection methods, methods based on deep neural networks have excellent performance in detection accuracy. These methods usually require pre-labeled training data, that is, these methods require that every pixel in the image in the training data must be labeled as a salient target or a non-salient target. However, generally speaking, there is a difficulty in obtaining these data, that is, the labeling of image data in the training set requires a lot of manpower and time costs.
[0003] In recent years, weakly supervised learning methods have been widely used in the field of salient object detection. This method reduces the cost of manual labeling to a certain extent. Its training data does not require pixel-level labeling, but only requires that the training data contain image-level labels or image description text. This method is relatively simple to obtain data, and the cost of manual labeling is also lower than the above methods. Taking the learning method of image-level labels as an example, image-level labels can be used to train a convolutional neural network classifier. This classifier, based on the CAM method, outputs the object represented by the label at a certain location in the image, and uses these locations as the output results of salient object detection. In addition, unsupervised learning methods do not require manual labeling. They only need to use traditional rule-based algorithms to generate pseudo-labels. These pseudo-labels are then learned using a fully supervised model method. The model output further refines the pseudo-labels, thus forming a self-supervised, coarse-to-fine learning process.
[0004] However, although the above methods perform well, most existing training data sets have their own biases. The above methods do not take into account the difference between the training data sets and the input data in real applications. During training, the model may learn the bias of these data sets, resulting in a decrease in accuracy during testing. In addition, in unsupervised learning methods, the pseudo-labels used are usually noisy, which usually affects the accuracy of the algorithm to a certain extent. Summary of the Invention
[0005] The present invention provides an unsupervised salient object detection method, system, device and medium, which solve the technical problem that most of the training data sets used in existing salient object detection methods have their own biases and the pseudo labels usually contain noise.
[0006] To solve the above technical problems, the present invention provides an unsupervised salient object detection method, system, device and medium.
[0007] In a first aspect, the present invention provides an unsupervised salient object detection method, the method comprising the following steps:
[0008] Generate pseudo labels for the input training images using the pseudo label generation model;
[0009] Inputting the training image into the backbone network to obtain a target feature vector;
[0010] Obtaining a contrast feature according to the target feature vector and the pseudo label, and filtering the contrast feature to obtain a bottom-level contrast feature and a high-level contrast feature;
[0011] Performing cluster learning on the bottom-level contrast features and the high-level contrast features respectively to obtain corresponding bottom-level cluster features and high-level cluster features;
[0012] Performing Cartesian product calculation on the bottom layer clustering features and the high layer clustering features to obtain Cartesian product features;
[0013] Obtaining an attention vector feature using the target feature vector and the Cartesian product feature;
[0014] The target feature vector and the attention vector features are input into a classification network to obtain a network prediction graph.
[0015] In a further embodiment, the step of obtaining a contrast feature based on the target feature vector and the pseudo label comprises:
[0016] Perform average pooling on all target feature vectors to obtain global features;
[0017] Obtaining a target region according to the pseudo label and the backbone network;
[0018] Performing average pooling on the target area to obtain local features;
[0019] A contrast feature is obtained according to the global feature and the local feature.
[0020] In a further embodiment, the step of obtaining an attention vector feature using the target feature vector and the Cartesian product feature comprises:
[0021] Count the proportion of each cluster in the Cartesian product features to obtain cluster proportion information;
[0022] Attention calculation is performed on the target feature vector, Cartesian product feature and cluster proportion information to obtain the attention vector feature.
[0023] In a further embodiment, the calculation formula of the attention vector feature is:
[0024]
[0025] Where ε represents the attention vector feature, X i represents the target feature vector corresponding to the i-th training image, Q represents the cluster ratio information, C grid,k represents the Cartesian product feature of the k-th cluster, q k represents the cluster ratio of the kth cluster, FC is the fully connected layer, and G represents the projection length.
[0026] In a further embodiment, the loss function used by the classification network is:
[0027]
[0028] in,
[0029]
[0030]
[0031] Where, represents the loss function of the t-th round of iterative training, Mean represents the average function, y represents the network prediction graph, and y i Represents the network prediction graph of the i-th training image, y j Represents the network prediction graph of the jth training image, y b represents the average output of all training images, N represents the total number of training images, prec represents the accuracy, rec represents the recall, W represents the height of the training image, H represents the length of the training image, l represents the pseudo label, Represents the partial derivative of the training image in the x direction, Represents the partial derivative of the training image in the y direction, represents the gradient in the x or y direction, β represents the indicator parameter, w2 and w3 represent the weights, and θ represents the constant.
[0032] In a further embodiment, an updated pseudo-label is obtained based on the network prediction graph output by the classification network and the pseudo-label of the previous round of iterative training, and the updated pseudo-label is used as the pseudo-label for the next round of iterative training.
[0033] In a further embodiment, the backbone network is a ResNet-50 neural network.
[0034] In a second aspect, the present invention provides an unsupervised salient object detection system, the system comprising:
[0035] A pseudo label generation module is used to generate pseudo labels for input training images using a pseudo label generation model;
[0036] a contrast generation module, configured to input the training image into a backbone network to obtain a target feature vector, obtain a contrast feature based on the target feature vector and the pseudo label, and filter the contrast feature to obtain a bottom-level contrast feature and a high-level contrast feature;
[0037] a contrast clustering module, configured to perform cluster learning on the underlying contrast features and the high-level contrast features respectively to obtain corresponding underlying cluster features and high-level cluster features, and perform Cartesian product calculation on the underlying cluster features and the high-level cluster features to obtain Cartesian product features;
[0038] The network training module is used to obtain an attention vector feature by using the target feature vector and the Cartesian product feature, and input the target feature vector and the attention vector feature into a classification network to obtain a network prediction graph.
[0039] In a third aspect, the present invention also provides a computer device comprising a processor and a memory, wherein the processor is connected to the memory, the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory so that the computer device performs the steps of implementing the above method.
[0040] In a fourth aspect, the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above method are implemented.
[0041] The present invention provides an unsupervised salient object detection method, system, device, and medium. This method not only extracts pseudo-labels from a pseudo-label generation model, but also utilizes a contrast clustering method and model training method to fuse low-level contrast features with high-level contrast features in a backbone network. Furthermore, the loss function designed in the present invention reduces the noise level of pseudo-labels in unsupervised learning methods, thereby improving the accuracy of salient object detection. Compared with existing technologies, this method enables unsupervised detection of targets with higher detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 11 is a flow chart of an unsupervised salient object detection method provided by an embodiment of the present invention;
[0043] Figure 2 Schematic diagram of the unsupervised salient object detection process provided by an embodiment of the present invention;
[0044] Figure 3 This is a schematic diagram of the test results provided by an embodiment of the present invention;
[0045] Figure 4 is a block diagram of an unsupervised salient object detection system provided by an embodiment of the present invention;
[0046] Figure 5 It is a structural diagram of a computer device provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0047] The following describes the embodiments of the present invention in detail with reference to the accompanying drawings. The embodiments are provided for illustrative purposes only and are not to be construed as limiting the present invention. The accompanying drawings are provided for reference and illustration only and do not constitute a limitation on the scope of protection of the present invention. Many changes may be made to the present invention without departing from the spirit and scope of the present invention.
[0048] refer to Figure 1 , the embodiment of the present invention provides an unsupervised salient object detection method, such as Figure 1 As shown, the method includes:
[0049] S1. Generate pseudo labels for the input training images using the pseudo label generation model.
[0050] In one embodiment, when performing the first round of iterative training, the present embodiment generates an initial pseudo label I using the pseudo label generation model for the acquired training image. 0 In this embodiment, the pseudo-label generation model preferably selects a dense and sparse reconstruction algorithm. In this embodiment, the dense and sparse reconstruction algorithm is used to perform saliency calculation on the training image to generate rough but accurate initial pseudo-labels. It should be noted that those skilled in the art can set the algorithm used by the pseudo-label generation model according to specific implementation conditions, and the present invention is not limited thereto.
[0051] S2. Input the training image into the backbone network to obtain the target feature vector.
[0052] This embodiment defines the input training image set D I ={I1,I2,…,I i}, for the i-th training image I iIn this embodiment, the training image I is processed by a backbone network. i Extract and obtain the corresponding target feature vector X i =f(I i ), where X i It is composed of B vector features extracted by B network layers of the backbone network, where f represents the backbone network. Therefore, after all training images are processed by the backbone network, this embodiment defines the target feature vector D X ={X1,X2,…,X i}.
[0053] In this embodiment, the backbone network preferably selects the ResNet-50 neural network. Those skilled in the art can adjust the backbone network according to actual conditions. This embodiment uses the ResNet-50 neural network to detect targets, which can analyze deeper target features and better explore the relationship between features, thereby obtaining a more comprehensive and discriminative feature representation.
[0054] S3. Obtain a contrast feature based on the target feature vector and the pseudo label, and filter the contrast feature to obtain a bottom-level contrast feature and a high-level contrast feature.
[0055] In one embodiment, the step of obtaining a contrast feature based on the target feature vector and the pseudo label includes:
[0056] Perform average pooling on all target feature vectors to obtain global features;
[0057] Obtaining a target region according to the pseudo label and the backbone network;
[0058] Performing average pooling on the target area to obtain local features;
[0059] A contrast feature is obtained according to the global feature and the local feature.
[0060] In this embodiment, the target feature vector set of the backbone network layer d is first Perform average pooling operation to obtain the global features of the dth layer of the backbone network Then, in the dth layer of the backbone network, this embodiment obtains the salient target area in the pseudo label according to the pseudo label in this layer, and inputs the salient area in the pseudo label into the backbone network to obtain the salient area detection result, and performs an average pooling operation on the salient area detection result to obtain the local feature of the dth layer of the backbone network. Finally, this embodiment calculates the contrast feature of the dth layer of the backbone network based on the global features and local features of the dth layer of the backbone network, wherein the calculation formula of the contrast feature is:
[0061]
[0062] Where, Represents the contrast feature of the dth layer of the backbone network.
[0063] After obtaining the contrast features of each layer of the backbone network, this embodiment screens the contrast features according to the distances between each layer of the neural network and the input layer and the output layer to obtain the bottom layer contrast features. and high-level contrast features It should be noted that, in a neural network structure, this embodiment defines the neural network layer features that are relatively close to the network input layer as bottom-level features, and its corresponding contrast features are bottom-level contrast features. At the same time, this embodiment defines the neural network layer features that are relatively close to the network output layer as high-level features, and its corresponding contrast features are high-level contrast features. Those skilled in the art can select the neural network layers as the bottom and high levels according to the specific implementation situation, and the embodiments of the present invention do not limit this.
[0064] S4. Perform cluster learning on the bottom-level contrast features and the high-level contrast features respectively to obtain corresponding bottom-level cluster features and high-level cluster features.
[0065] In one embodiment, this embodiment uses the K-Means algorithm to perform cluster learning on the bottom contrast feature and the high-level contrast feature to obtain the corresponding bottom cluster feature. and high-level clustering features
[0066] For the sake of brief description, the underlying clustering features are and high-level clustering features The calculation process of is described uniformly, namely:
[0067] This embodiment defines the output of the K-Means algorithm as CT = {C1, C2, ..., C k}, where k represents the number of given clusters. At the same time, for each cluster vector matrix C k , which is defined in this embodiment as Then, this embodiment obtains the cluster prototype according to the cluster vector matrix Among them, |C k | represents the total number of clusters, thus, the clustering feature E is obtained CST ={A1,A2,…,A k}.
[0068] S5. Perform Cartesian product calculation on the bottom-level clustering features and the high-level clustering features to obtain Cartesian product features.
[0069] This embodiment sets the underlying clustering features and high-level clustering features The scales are γ low and γ high In this embodiment, a Cartesian product feature is obtained by performing Cartesian product calculation on the underlying clustering feature and the high-level clustering feature, wherein the calculation formula of the Cartesian product feature is:
[0070]
[0071] in,
[0072]
[0073] Where C grid Represents the Cartesian product feature, cat represents the vector concatenation operation, represents the u-th underlying clustering feature, Represents the vth high-level clustering feature.
[0074] In this embodiment, the above steps S2 to S5 are all performed as follows Figure 2 The contrast clusterer implementation shown.
[0075] S6. Use the target feature vector and the Cartesian product feature to obtain the attention vector feature.
[0076] In one embodiment, Figure 2 As shown, this embodiment inputs the Cartesian product feature generated by the contrast clusterer into the model trainer, and uses the model trainer to process the target feature vector and the Cartesian product feature to obtain the attention vector feature, which specifically includes:
[0077] Count the proportion of each cluster in the Cartesian product features to obtain cluster proportion information;
[0078] Attention calculation is performed on the target feature vector, Cartesian product feature and cluster proportion information to obtain the attention vector feature.
[0079] In this embodiment, the calculation formula of the cluster proportion information Q is:
[0080] Q={q1,q2,…,q k}
[0081] in,
[0082]
[0083] Where C grid,k represents the Cartesian product feature of the k-th cluster, |C k | represents the number of vectors in the clustering vector matrix, q k Indicates the cluster ratio of the k-th cluster.
[0084] This embodiment achieves the effect of suppressing unimportant features and enhancing important features by performing attention mechanism calculation, thereby improving the prediction performance of the classification network model.
[0085] In this embodiment, the calculation formula of the attention vector feature is:
[0086]
[0087] Where ε represents the attention vector feature, X i represents the target feature vector corresponding to the i-th training image. In this embodiment, FC is a fully connected layer, which projects a vector of a certain length to a vector of length G.
[0088] S7. Input the target feature vector and the attention vector features into the classification network to obtain a network prediction graph.
[0089] In this embodiment, the target feature vector and the attention vector feature are merged together and input into the classification network to obtain a network prediction graph. It should be noted that those skilled in the art can select a classification network according to the specific implementation situation. For example, a multilayer perceptron is selected as the classification network in this embodiment, but it is necessary to ensure that the selected classification network can realize gradient backpropagation and can use the backpropagation algorithm.
[0090] During the t-th round of iterative training, this embodiment defines the loss function of the classification network as:
[0091]
[0092] in,
[0093]
[0094]
[0095] Where, represents the loss function of the t-th round of iterative training, Mean represents the average function, y represents the network prediction graph, and y i Represents the network prediction graph of the i-th training image, y j Represents the network prediction graph of the jth training image, y brepresents the average output of all training images, N represents the total number of training images, prec represents the accuracy, rec represents the recall, W represents the height of the training image, H represents the length of the training image, l represents the pseudo label, Represents the partial derivative of the training image in the x direction, Represents the partial derivative of the training image in the y direction, represents the gradient in the x or y direction, β represents the index parameter, w2 and w3 represent weights, θ represents a constant, and in this embodiment, θ=1.5 is preferably set.
[0096] The input of the Mean function used in this embodiment is a two-dimensional matrix. The Mean function represents the calculation of the average value of the matrix, and this average value can effectively alleviate the spatial distribution bias.
[0097] In one embodiment, the unsupervised salient object detection method provided by this embodiment further includes the following steps:
[0098] The pseudo-label generation model obtains updated pseudo-labels according to the network prediction graph output by the classification network and the pseudo-labels of the previous round of iterative training, and uses the updated pseudo-labels as the pseudo-labels for the next round of iterative training.
[0099] This embodiment needs to be updated before each round of iteration, that is, when the t-th round of iterative training is performed, this embodiment uses the network prediction graph obtained by the t-th round of iterative training to perform the t-th pseudo-label update, and obtains the updated pseudo-label I t , and the pseudo label I t As the pseudo label used in the next round of iterative training, the t-th pseudo label update formula is:
[0100]
[0101] in,
[0102]
[0103] Where, represents the sliding average of the t-th round of iterative training, CRF represents the Conditional Random Field function, represents the sliding average of the (t-1)th round of iterative training, y t represents the network prediction graph obtained by the t-th round of iterative training, α represents the hyperparameter, θ t It represents the average value of all network prediction graphs output by the (t-1)th round of iterative training. In this embodiment, t>1, 0<α<1.
[0104] This embodiment uses the calculated sliding average as the pseudo label after this update, which can effectively reduce the impact of noise and make the prediction model more accurate. The pseudo label generation model in this embodiment accepts the output of the model trainer each time and updates the pseudo label according to the output of the model trainer.
[0105] Figure 3 This is the test result of this embodiment when using MSRA-B and DUTS datasets. Figure 3 It can be seen that the E-measure, mean comprehensive evaluation index (Mean F-measure), and MAE are all good when using the MSRA-B dataset.
[0106] This embodiment provides an unsupervised salient object detection method. Contrast features are obtained using pseudo-labels and target feature vectors. The contrast features are then filtered, and the filtered low-level and high-level contrast features are clustered and processed. Finally, a prediction map corresponding to each training image is obtained using attention calculation and a classification network. Compared with the existing technology, this embodiment can achieve excellent performance on salient object detection datasets without relying on manual labels. At the same time, the loss function used in this embodiment has a certain positive effect on reducing pseudo-label noise, thereby improving the salient object prediction effect.
[0107] It should be noted that the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of this application.
[0108] In one embodiment, Figure 4 As shown, this embodiment provides an unsupervised salient object detection system, the system comprising:
[0109] A pseudo label generation module 101 is configured to generate pseudo labels from input training images using a pseudo label generation model;
[0110] A contrast generation module 102 is configured to input the training image into a backbone network to obtain a target feature vector, obtain a contrast feature based on the target feature vector and the pseudo label, and filter the contrast feature to obtain a bottom-level contrast feature and a high-level contrast feature;
[0111] a contrast clustering module 103 for performing cluster learning on the underlying contrast features and the high-level contrast features respectively to obtain corresponding underlying cluster features and high-level cluster features, and performing Cartesian product calculation on the underlying cluster features and the high-level cluster features to obtain Cartesian product features;
[0112] The network training module 104 is used to use the target feature vector and the Cartesian product feature to obtain the attention vector feature, and input the target feature vector and the attention vector feature into the classification network to obtain a network prediction graph.
[0113] For the specific definition of an unsupervised salient target detection system, please refer to the above-mentioned definition of an unsupervised salient target detection method, which will not be repeated here. Those of ordinary skill in the art will appreciate that the various modules and steps described in conjunction with the embodiments disclosed in this application can be implemented in hardware, software, or a combination of both. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.
[0114] This embodiment provides an unsupervised salient object detection system. This system addresses the inherent bias in the training dataset and takes into account possible differences in contrast distribution and centralization distribution between the training dataset and the test input data. This system generates pseudo-labels corresponding to the training images through a pseudo-label generation module. Furthermore, it implements unsupervised salient object detection through a contrast generation module, a contrast clustering module, and a network training module. Compared to supervised or weakly supervised object detection, this embodiment not only avoids the reliance on manual labeling in salient object detection methods, significantly reducing costs, but also reduces the noise level of pseudo-labels in unsupervised learning methods through a loss function, thereby improving detection accuracy to a certain extent.
[0115] Figure 5 A computer device provided in an embodiment of the present invention includes a memory, a processor and a transceiver, which are connected via a bus; the memory is used to store a set of computer program instructions and data, and can transmit the stored data to the processor, and the processor can execute the program instructions stored in the memory to perform the steps of the above method.
[0116] The memory may include volatile memory or non-volatile memory, or may include both volatile and non-volatile memory; the processor may be a central processing unit, a microprocessor, an application-specific integrated circuit, a programmable logic device, or a combination thereof. By way of example and not limitation, the programmable logic device may be a complex programmable logic device, a field programmable gate array, a general purpose array logic, or any combination thereof.
[0117] Additionally, the memory may be a physically separate unit or integrated with the processor.
[0118] It can be understood by those skilled in the art that Figure 5The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have the same component arrangement.
[0119] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above method are implemented.
[0120] Embodiments of the present invention provide an unsupervised salient object detection method, system, device, and medium. This unsupervised salient object detection method utilizes generated pseudo-labels for iterative training, clusters contrast features using a K-Means clustering algorithm, and optimizes training using an attention mechanism and a classification network to achieve unsupervised yet reliable salient object detection. This method achieves excellent performance on salient object detection datasets without relying on manual labeling. It also uses a loss function to reduce the noise level of pseudo-labels in unsupervised learning methods, resulting in superior performance compared to existing salient object detection methods.
[0121] In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line, or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., an SSD).
[0122] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing related hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods.
[0123] The above-described embodiments merely represent several preferred implementations of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art could make several improvements and substitutions without departing from the technical principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be based on the scope of protection of the claims.
Claims
1. An unsupervised salient object detection method, characterized in that: The following steps are involved: Generate pseudo labels for the input training images using the pseudo label generation model; Inputting the training image into the backbone network to obtain a target feature vector; Obtaining contrast features based on the target feature vector and the pseudo-label, and screening the contrast features based on the distances between each layer of the backbone network and the input layer and the output layer to obtain bottom-level contrast features and high-level contrast features; wherein the bottom-level features are defined as the features of the neural network layer that are relatively close to the network input layer, and the contrast features corresponding thereto are the bottom-level contrast features; and the high-level features are defined as the features of the neural network layer that are relatively close to the network output layer, and the contrast features corresponding thereto are the high-level contrast features; Performing cluster learning on the bottom-level contrast features and the high-level contrast features respectively to obtain corresponding bottom-level cluster features and high-level cluster features; Performing Cartesian product calculation on the bottom layer clustering features and the high layer clustering features to obtain Cartesian product features; Obtaining an attention vector feature using the target feature vector and the Cartesian product feature; Inputting the target feature vector and the attention vector features into a classification network to obtain a network prediction graph; The step of obtaining a contrast feature according to the target feature vector and the pseudo label includes: All target feature vectors of each layer of the backbone network are average-pooled to obtain the global features of each layer of the backbone network; According to the pseudo labels of each layer of the backbone network, the salient target regions in the pseudo labels are obtained, and the salient target regions in the pseudo labels are input into the backbone network to obtain the salient region detection results; Perform average pooling on the salient region detection results to obtain local features of each layer of the backbone network; According to the global features and local features of each layer of the backbone network, the contrast features of each layer of the backbone network are obtained; the calculation formula of the contrast features is: Where, represents the contrast feature of the dth layer of the backbone network; Represents the global features of the dth layer of the backbone network; Represents the local features of the dth layer of the backbone network; The step of obtaining the attention vector feature by using the target feature vector and the Cartesian product feature includes: counting the proportion of each cluster in the Cartesian product feature to obtain cluster proportion information; performing attention calculation on the target feature vector, the Cartesian product feature and the cluster proportion information to obtain the attention vector feature.
2. The unsupervised salient object detection method according to claim 1, wherein: The calculation formula of the attention vector feature is: Where, represents the attention vector feature, represents the target feature vector corresponding to the i-th training image, Q represents the cluster ratio information, represents the Cartesian product feature of the k-th cluster, represents the cluster ratio of the kth cluster, FC represents the fully connected layer, Indicates the projection length.
3. The unsupervised salient object detection method according to claim 1, wherein: The loss function used by the classification network is: in, Where, represents the loss function of the t-th round of iterative training, represents the average function, y represents the network prediction graph, represents the network prediction graph for the i-th training image, represents the network prediction graph for the j-th training image, represents the average output of all training images, N represents the total number of training images, Indicates the accuracy, represents the recall rate, W represents the height of the training image, H represents the length of the training image, represents a pseudo label, represents the partial derivative of the training image in the x direction, Represents the partial derivative of the training image in the y direction, Represents the gradient in the x or y direction, Represents the indicator parameters, 、 represents the weight, Represents a constant.
4. The unsupervised salient object detection method according to claim 1, wherein: Also includes: According to the network prediction graph output by the classification network and the pseudo labels of the previous round of iterative training, updated pseudo labels are obtained, and the updated pseudo labels are used as pseudo labels for the next round of iterative training.
5. The unsupervised salient object detection method according to claim 1, wherein: The backbone network is the ResNet-50 neural network.
6. An unsupervised salient object detection system, characterized in that The system comprises: A pseudo-label generation module is used to generate pseudo-labels from input training images using a pseudo-label generation model; A contrast generation module is configured to input the training image into the backbone network to obtain a target feature vector, obtain a contrast feature based on the target feature vector and the pseudo label, and screen the contrast feature based on the distance between each layer of the backbone network and the input layer and the output layer to obtain a bottom-level contrast feature and a high-level contrast feature; wherein the bottom-level feature is defined as a neural network layer feature that is relatively close to the network input layer, and its corresponding contrast feature is the bottom-level contrast feature; the high-level feature is defined as a neural network layer feature that is relatively close to the network output layer, and its corresponding contrast feature is the high-level contrast feature; a contrast clustering module, configured to perform cluster learning on the underlying contrast features and the high-level contrast features respectively to obtain corresponding underlying cluster features and high-level cluster features, and perform Cartesian product calculation on the underlying cluster features and the high-level cluster features to obtain Cartesian product features; A network training module is used to obtain an attention vector feature using a target feature vector and the Cartesian product feature, and input the target feature vector and the attention vector feature into a classification network to obtain a network prediction graph; The step of obtaining a contrast feature based on the target feature vector and the pseudo label specifically includes: All target feature vectors of each layer of the backbone network are average-pooled to obtain the global features of each layer of the backbone network; According to the pseudo labels of each layer of the backbone network, the salient target regions in the pseudo labels are obtained, and the salient target regions in the pseudo labels are input into the backbone network to obtain the salient region detection results; Perform average pooling on the salient region detection results to obtain local features of each layer of the backbone network; According to the global features and local features of each layer of the backbone network, the contrast features of each layer of the backbone network are obtained; the calculation formula of the contrast features is: Where, represents the contrast feature of the dth layer of the backbone network; Represents the global features of the dth layer of the backbone network; Represents the local features of the dth layer of the backbone network; The method of using the target feature vector and the Cartesian product feature to obtain the attention vector feature specifically includes: counting the proportion of each cluster in the Cartesian product feature to obtain cluster proportion information; performing attention calculation on the target feature vector, the Cartesian product feature and the cluster proportion information to obtain the attention vector feature.
7. A computer device, characterized in that: The computer device comprises a processor and a memory, wherein the processor is connected to the memory, the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the computer device performs the method according to any one of claims 1 to 5.
8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a computer, the method according to any one of claims 1 to 5 is implemented.