Ship detection method and device based on trusted quantum evidence neural network

Abstract

The application provides a ship detection method and device based on a trusted quantum evidence neural network, the method comprising: acquiring a synthetic aperture radar image and inputting the image into a plurality of full complex neural networks based on complex value contrast learning to extract complex value feature vectors respectively; constructing a quantum recognition framework containing known category propositions and an empty set proposition; each network generates a corresponding generalized quantum basic probability amplitude function according to the feature vector thereof, and fuses the function according to a generalized quantum evidence combination rule to obtain a generalized quantum basic probability distribution; and based on the distribution and a preset decision rule, outputting a detection result of a target being a known or unknown ship category. The method synchronously optimizes the feature amplitude and phase discriminability through full complex networks and joint loss, introduces quantum evidence theory to model and fuse multi-source uncertain information, can simultaneously realize accurate classification of known categories and effective rejection of unknown categories, and improves the reliability and robustness of ship detection in an open scene.

Description

Technical Field

[0001] This application relates to the field of image or video recognition in next-generation information technology, and in particular to a ship detection method and device based on a trusted quantum evidence neural network. Background Technology

[0002] Currently, deep learning-based image target detection and recognition technology is the primary means of interpreting Synthetic Aperture Radar (SAR) images from ships. Furthermore, current research on open set identification and uncertainty estimation aims to enable models to identify and reject unknown samples, thereby improving decision-making reliability.

[0003] However, several fundamental bottlenecks remain in practical applications in open sea areas. First, mainstream models are built on the real number domain, making it difficult to fully utilize the inherent complex numerical amplitude and phase information of SAR images, thus limiting the discriminative power of features. Second, traditional detection methods lack a unified formal framework to characterize the model's understanding of "known" and "unknown" states, and also struggle to effectively fuse multi-source uncertain evidence to resolve conflicts. These fundamental problems limit the system's performance in achieving reliable and trustworthy decision-making in complex open environments. Summary of the Invention

[0004] This application proposes a ship detection method and device based on a credible quantum evidence neural network. The aim is to improve the open set recognition performance of ship targets in synthetic aperture radar images by deeply integrating a fully complex valued neural network with quantum evidence theory, simultaneously achieving high-precision classification of known categories and effective rejection of unknown targets, thereby enhancing the practicality and decision credibility of the system in open sea scenarios.

[0005] In a first aspect, embodiments of this application provide a ship detection method based on a trusted quantum evidence neural network, including: Acquire synthetic aperture radar images as the images to be detected; The image to be detected is input into multiple fully complex value neural networks that have been pre-trained with a minimized joint magnitude-phase loss function to extract the corresponding complex value feature vectors. Construct a quantum recognition framework. The proposition set of the quantum recognition framework includes reference propositions for multiple known ship categories, as well as empty set propositions representing unknown ship categories. By using multiple fully complex-valued neural networks, the generalized quantum fundamental probability amplitudes corresponding to all propositions in the quantum recognition framework are determined based on their respective complex-valued feature vectors. According to the generalized quantum evidence combination rule, multiple sets of generalized quantum fundamental probability amplitudes are fused to obtain the fused generalized quantum fundamental probability distribution. The fused generalized quantum fundamental probability distribution is the set of support degrees corresponding to each proposition in the quantum recognition framework. Based on the generalized quantum fundamental probability distribution and the preset classification decision rules, the detection result is determined. The detection result is used to characterize whether the ship target in the image to be detected belongs to a known ship category or an unknown ship category.

[0006] Secondly, embodiments of this application provide a ship detection device based on a trusted quantum evidence neural network, comprising: An image acquisition unit is used to acquire synthetic aperture radar images as images to be detected. The vector extraction unit is used to input the image to be detected into multiple fully complex value neural networks that have been pre-trained with minimizing the joint magnitude-phase loss function, so as to extract the corresponding complex value feature vectors respectively; The framework building unit is used to construct the quantum recognition framework. The proposition set of the quantum recognition framework includes reference propositions for multiple known ship categories, as well as empty set propositions representing unknown ship categories. The distribution determination unit is used to determine the generalized quantum fundamental probability amplitude corresponding to all propositions in the quantum recognition framework by using multiple fully complex-valued neural networks according to their respective complex-valued feature vectors; and to fuse multiple sets of generalized quantum fundamental probability amplitudes according to the generalized quantum evidence combination rule to obtain the fused generalized quantum fundamental probability distribution, which is the set of support corresponding to each proposition in the quantum recognition framework. The category detection unit is used to determine the detection result based on the generalized quantum fundamental probability distribution and the preset classification decision rules. The detection result is used to characterize whether the ship target in the image to be detected belongs to a known ship category or an unknown ship category.

[0007] Thirdly, embodiments of this application provide an electronic device including a processor, a memory, and one or more programs, the one or more programs being stored in the memory and configured to be executed by the processor, the programs including instructions for performing the steps as described in the first aspect of embodiments of this application.

[0008] Fourthly, embodiments of this application provide a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps in the first aspect of embodiments of this application.

[0009] Fifthly, embodiments of this application provide a computer program product, including a computer program / instructions, which, when executed by a processor, implement some or all of the steps described in the first aspect of embodiments of this application.

[0010] As can be seen, in this embodiment, joint amplitude-phase loss is first used for training, fully utilizing all the information of complex-valued SAR data to extract more discriminative features. Then, a quantum recognition framework incorporating the empty set proposition is constructed to formalize the unknown category, achieving explicit quantitative modeling of uncertainty. Finally, based on the generalized quantum evidence combination rule, multi-source complex-valued evidence is fused, and quantum interference effects are used to resolve conflicts, resulting in a robust decision distribution. This method can simultaneously achieve high-precision classification of known categories and high-reliability rejection of unknown targets, improving the system's practicality and decision credibility in open sea scenarios. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart illustrating a ship detection method based on a trusted quantum evidence neural network provided in an embodiment of this application. Figure 2 This is a schematic diagram of a typical general architecture and mainstream method for identifying SAR images based on real-valued neural networks, provided in the embodiments of this application. Figure 3a This is a schematic diagram of another ship detection process based on a trusted quantum evidence neural network provided in an embodiment of this application; Figure 3b This is a schematic diagram of the structure of a complex value feature extraction module provided in an embodiment of this application; Figure 3c This is a schematic diagram of the structure of a generalized quantum evidence fundamental probability amplitude function generation module provided in an embodiment of this application; Figure 3d This is a schematic diagram of the structure of a feature fusion and decision module provided in an embodiment of this application; Figure 4 This is a functional unit block diagram of a ship detection device based on a trusted quantum evidence neural network provided in an embodiment of this application; Figure 5 This is a functional unit block diagram of another ship detection device based on a trusted quantum evidence neural network provided in this application embodiment; Figure 6 This is a structural block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0013] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0014] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0015] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0016] This application provides a ship detection method based on a trusted quantum evidence neural network. The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0017] Please see Figure 1 , Figure 1 This is a flowchart illustrating a ship detection method based on a trusted quantum evidence neural network, as provided in an embodiment of this application. The method includes: Step S101: Acquire a synthetic aperture radar image as the image to be detected.

[0018] Synthetic Aperture Radar (SAR) images are complex numerical images formed by a spaceborne or airborne SAR system actively emitting microwave pulses and receiving backscattered echoes from ground objects. Unlike optical images, SAR images inherently contain information in two dimensions: amplitude (reflecting the intensity of target backscattering) and phase (carrying information about the target's geometry, attitude, and subtle motion). In shipboard inspection missions, the images to be inspected typically originate from continuous scanning of vast sea areas. These images cover a large area, contain targets of varying sizes, and are easily affected by sea clutter, islands, and man-made debris at sea.

[0019] Furthermore, due to the diverse types of vessels in open seas, in addition to the vessel categories known during the training phase (such as cargo ships, tankers, and fishing boats), the detection system also needs to deal with a large number of targets that never appeared during the training phase, such as new military vessels, floating platforms, or various non-ship objects. Therefore, acquiring SAR images as the images to be detected is a prerequisite for performing open-set vessel detection. The aim is to simultaneously achieve the identification of two types of targets from a complex marine background: providing accurate classification results for known vessel categories, while reliably rejecting all unknown targets (judging them as "unknown vessel category"), in order to avoid serious decision-making errors caused by forced classification, thereby improving the operational usability and security of the autonomous monitoring system.

[0020] Step S102: The image to be detected is input into multiple fully complex neural networks that have been pre-trained with a minimized joint magnitude-phase loss function to extract the corresponding complex feature vectors.

[0021] In this application, the "Trustworthy Quantum Evidence Neural Network" (Trustworthy QENN) refers to a ship detection architecture built upon complex-valued contrastive learning and quantum evidence theory (QET). This architecture comprises a complex-valued feature extraction component and a quantum evidence processing component. The complex-valued feature extraction component employs a Complex Valued Neural Network (CVNN) trained with a Joint Amplitude-Phase Loss (JAP-Loss) function to extract highly discriminative complex-valued features from synthetic aperture radar images. The quantum evidence processing component performs quantum recognition framework construction, generalized quantum fundamental probability amplitude generation, multi-source evidence fusion, and decision-making.

[0022] For example, please refer to Figure 2 , Figure 2 This is a schematic diagram illustrating a typical general architecture and mainstream method for recognizing SAR images based on real-valued neural networks, as provided in this application embodiment. The main flow of this architecture is as follows: input a ship SAR image, extract features through a real-valued neural network, and then feed it into a classifier to complete open set recognition and classification. In-Distribution (ID) targets (i.e., ships of known categories) refer to known ship categories seen during the model training phase, requiring explicit classification results; Out-Of-Distribution (OOD) targets (i.e., samples of unknown categories) refer to new types of ships, marine debris, or noise not seen by the model, which need to be classified as "unknown" to avoid forced misclassification. Figure 2 In this context, real-valued neural networks serve as the backbone for feature extraction. However, their core limitation lies in their ability to process only single-channel amplitude data, failing to leverage the inherent amplitude and phase coupling information within ship SAR images. Based on this, traditional methods are categorized into probabilistic models (such as OpenMax), discriminative models (such as ConOSR), and evidence models (such as Evidential Deep Learning (EDL)).

[0023] Furthermore, probabilistic models rely on probability thresholds or extreme value theory derived from the output of Softmax (a normalized exponential function, a commonly used activation function in machine learning and deep learning, especially in multi-class classification tasks to convert model output into a probability distribution) to distinguish out-of-distribution (OOD) samples. However, fixed thresholds lack scenario adaptability, requiring manual readjustment when migrating to new data, resulting in poor model generalization ability and difficulty in practical deployment. Discriminative models construct decision boundaries through feature distance or contrast constraints, but they only focus on whether the feature distance falls within the decision boundary, failing to quantify the model's confidence in the judgment result (i.e., cognitive uncertainty). This leads to the erroneous rejection of difficult samples with ambiguous features that should be classified as known. Evidence models, while incorporating the Dempster-Shafer (DS) evidence theory framework, have basic probability assignments defined only in the real domain, unable to handle complex-valued features, and difficult to integrate with end-to-end deep learning. The aforementioned methods generally suffer from insufficient real-valued feature representation capabilities and a lack of systematic uncertainty modeling theory, leading to a reduction in the classification accuracy of known categories to improve the detection rate of unknown targets.

[0024] To address the aforementioned technical challenges, this application proposes a Trustworthy Quantum Evidence Neural Network (QENN) based on quantum evidence theory. This network employs a fully complex-valued architecture, treating the SAR image to be detected as a complex-valued whole for end-to-end processing. To overcome the inherent bottlenecks of real-valued models, the network is pre-trained by minimizing the Joint Amplitude-Phase Loss (JAP-Loss) function. This loss function simultaneously optimizes the amplitude distribution and phase correlation of the complex-valued features, ensuring that the extracted complex-valued feature vectors retain complete phase semantic information while possessing high intra-class compactness and high inter-class separability.

[0025] The specific training process and internal composition of the fully complex value neural network will be described in detail below with reference to specific embodiments.

[0026] In one possible embodiment, the fully complex-valued neural network consists of a complex-valued encoding network. Complex projection networks and complex value classification network The method comprises the following components: before inputting the image to be detected into multiple pre-trained fully complex-valued neural networks trained with a minimized joint magnitude-phase loss function to extract corresponding complex-valued feature vectors, the method further includes: inputting pre-stored training samples into a complex-valued encoding network to obtain first complex-valued features; inputting the first complex-valued features into a complex-valued projection network to obtain complex-valued projection vectors; calculating a joint magnitude-phase loss function based on the complex-valued projection vectors; performing backpropagation based on the joint magnitude-phase loss function to update the parameters of the complex-valued encoding network and the complex-valued projection network; inputting training samples into the updated complex-valued encoding network to obtain second complex-valued features; inputting the second complex-valued features into a complex-valued classification network to obtain classification features; and updating the parameters of the complex-valued classification network based on a cross-entropy loss function determined by the magnitude of the classification features.

[0027] Please refer to Figure 3a , Figure 3a This is a schematic diagram of another ship detection process based on a trusted quantum evidence neural network provided in this application embodiment. Figure 3a This document fully illustrates the entire training and testing process of the aforementioned credible quantum evidence neural network, covering training samples. Test samples The system comprises five core modules: complex-valued feature extraction, generalized quantum evidence fundamental probability amplitude function generation, feature fusion, and decision-making. During the training phase, pre-labeled images within and outside the distribution are used as training samples. To complete network parameter optimization, the testing phase uses test samples. As input, the complex value feature extraction module outputs complex value features. ,Will The generalized quantum evidence fundamental probability amplitude is obtained by inputting into the generalized quantum evidence fundamental probability amplitude generation module. Then The input feature fusion and decision-making modules perform fusion and decision discrimination, and finally output the final detection result. The detection results This is used to characterize ship targets as either known ship categories within the distribution or unknown ship categories outside the distribution.

[0028] Please see Figure 3b , Figure 3b This is a schematic diagram of the structure of a complex value feature extraction module provided in an embodiment of this application. For example... Figure 3b As shown, this embodiment is Figure 3a The training process performed by the complex value feature extraction module is divided into two stages: Step 1 and Step 2. The first stage is used to optimize the complex value encoding network. and complex projection network The second stage is used for freezing. And only optimize the complex value classification network .

[0029] Specifically, before inputting the image to be detected into multiple fully complex value neural networks pre-trained with minimizing the joint magnitude-phase loss function to extract the corresponding complex value feature vectors, the method also includes the following training steps: First, store the pre-stored training samples Input complex value encoding network The first complex-valued feature is obtained; the first complex-valued feature is then input into the complex-valued projection network. The complex-valued projection vector is obtained; based on the complex-valued projection vector, the joint magnitude-phase loss function (JAP-Loss) is calculated; backpropagation is performed based on the joint magnitude-phase loss function to update the complex-valued coding network. and complex projection network The parameters.

[0030] Then, the training samples Input the updated complex value encoding network (This stage of complex value encoding network) The parameters are completely frozen, such as Figure 3b (As indicated by the "parameter freeze" label in the image), the second complex-valued feature is obtained; the second complex-valued feature is then input into the complex-valued classification network. The classification features are obtained; the complex value classification network is updated based on the cross-entropy loss function determined by the magnitude of the classification features. The parameters. In the second stage, the complex-valued projection network. It will no longer participate in the calculation.

[0031] in, Figure 3b The diagram shows a complex value coding network. The internal structure consists of a learnable Gabor complex-valued convolutional layer as the first layer. Gabor filters are widely used in image processing to extract local features with direction and frequency selectivity. Their kernel function is a Gaussian envelope-modulated sine / cosine wave in the spatial domain. This embodiment employs a learnable Gabor complex-valued convolutional layer, generating the real and imaginary parts of the output features through a pair of orthogonal filters (the real part being an even-symmetric cosine component and the imaginary part an odd-symmetric sine component), thus converting the real-valued input into a complex-valued representation. Unlike traditional Gabor filters with fixed parameters, "learnable" means that the filter's direction and frequency parameters can be optimized end-to-end through backpropagation, enabling the network to adaptively adjust the filtering response based on the geometric characteristics of different ship targets in SAR images. This design can simultaneously extract both direction-selective and frequency-selective features of the image, providing information-rich complex-domain feature maps for subsequent complex-valued convolutional blocks. Multiple complex-valued convolutional blocks (e.g., 8) are then cascaded. Each convolutional block contains a complex-valued convolutional layer, a complex-valued batch normalization layer (CBN), and a complex-valued activation function (CReLU). Finally, after global complex-valued average pooling and flattening operations, a complex-valued feature vector of a preset dimension (e.g., 128 dimensions) is output.

[0032] Complex projection network It consists of two complex-valued fully connected layers: the first layer is a complex-valued linear layer that maps the input to a preset dimension (e.g., 128→128), followed by complex-valued one-dimensional batch normalization and CReLU activation; the second layer is also a complex-valued linear layer that maps to a preset dimension (e.g., 128→128), followed by complex-valued one-dimensional batch normalization, and finally outputs a complex-valued projection vector of a preset dimension (e.g., 128-dimensional), which is used for comparative learning optimization.

[0033] Complex value classification network It also consists of two complex-valued fully connected layers: the first layer is a complex-valued linear layer that maps the input to a preset dimension (e.g., 128→128), followed by complex-valued one-dimensional batch normalization and CReLU activation; the second layer is a complex-valued linear layer that maps the input to the total number of known ship categories. Matching dimensions (e.g., 128 →) (This is followed by complex-valued one-dimensional batch normalization, and the final output is the sum of the total number of known ship categories.) The matched complex-valued classification features are used for subsequent classification and quantum evidence mapping.

[0034] The complex-valued batch normalization and complex-valued activation functions in the above network are specifically designed for the complex domain to ensure that the gradient can propagate stably in the real and imaginary parts. This is the key to achieving end-to-end training of complex-valued networks.

[0035] The joint amplitude-phase loss function is the core of this embodiment for achieving synchronous optimization of amplitude and phase. This loss function is composed of a weighted amplitude loss term and a phase loss term. The amplitude loss term constrains the amplitude features of similar samples to be closer together in cosine space, while keeping dissimilar samples further apart. The phase loss term eliminates the ambiguity of phase periodicity by mapping the original phase to a phase projection vector, ensuring that the phase projection vectors of similar samples are closer together and dissimilar samples are further apart. By jointly optimizing these two losses, the network can learn complex-valued features that simultaneously possess amplitude and phase discriminative properties. The cross-entropy loss function used in the second stage is calculated based on the amplitude of each component of the classification feature, ensuring that the amplitude of the complex-valued feature vector is strictly aligned with the classification confidence, while fully preserving phase information.

[0036] As can be seen, in this example, the credible quantum evidence neural network achieves joint optimization of amplitude and phase during the feature extraction stage, overcoming the inherent defect of existing real-valued models that lose phase information, and obtaining complex-valued features with high intra-class compactness and high inter-class separability. Meanwhile, the strategy of freezing the encoder and fine-tuning the classifier in the second stage ensures the consistency between feature discriminativeness and classification confidence.

[0037] In one possible embodiment, calculating the joint magnitude-phase loss function based on the complex-valued projection vector includes: obtaining a training batch used in a single parameter update iteration; determining the magnitude component and original phase vector corresponding to the complex-valued projection vector of each training sample within the training batch; calculating the magnitude loss term based on the magnitude component using a normalized exponential function; calculating the cosine and sine values ​​for each phase component corresponding to each original phase vector, and arranging all calculation results sequentially to form the corresponding phase projection vector; calculating the phase loss term based on the phase projection vector using a normalized exponential function; and weighting and summing the magnitude loss term and the phase loss term according to a preset balancing weight coefficient to obtain the joint magnitude-phase loss function.

[0038] For example, the training batch contains multiple ships carrying known ship category ID tags. training samples Let the ID training set be... Known category set ,in, Let be the total number of ID categories. And let the set of sample indices within a training batch be . B Batch size is |B| For anchor samples (the current sample within the batch) t Define the set of positive samples This refers to all samples within the batch that belong to the same category as the anchor sample but are not itself. Furthermore, the "normalized exponential function" is the softmax function. The "amplitude similarity" uses cosine similarity. Its range is [-1, 1]. , These are real-valued vectors of the same dimension; a larger value indicates that the directions of the two magnitude vectors are more aligned. The "phase projection vector" is constructed as follows: For anchor point sample complex value projection vector amplitude components, =∣ |, whose original phase vector is Each phase component ,for The argument of the corresponding complex-valued component. Addressing the periodic ambiguity of the original phase ( and (Geometrically adjacent, but numerically the largest distance, leading to anomalies in training gradients). The original phase vectors are mapped to Cartesian coordinates to construct a 2D phase projection vector. The formula is as follows: .

[0039] Furthermore, this transformation makes the cosine similarity of the two phase projection vectors equivalent to the mean of the cosine values ​​of the original phase component differences, i.e. This completely eliminates the problem of training gradient anomalies caused by the periodicity of the original phase values ​​(e.g., π and -π are adjacent on the circumference but are far apart in value).

[0040] Among them, the "balance weight coefficient" It is a fixed proportional parameter between 0 and 1, used to adjust the relative importance of the amplitude loss term and the phase loss term in the total loss. In this embodiment, A value of 0.5 indicates that amplitude and phase are considered equally important information sources. Furthermore, a temperature hyperparameter τ (with a value range of 0.01) is introduced. 1.0 (In this embodiment, τ=0.1) is used to control the degree of attention the contrast loss pays to difficult samples: the lower the temperature, the more the model pays attention to negative samples similar to the current sample, which helps to form more compact feature clusters.

[0041] The design principle of the joint amplitude-phase loss function proposed in this embodiment is as follows: Amplitude loss term Using the idea of ​​supervised contrastive learning, for the current sample t (which can be called the anchor sample) in the batch, its amplitude components are... =∣ | Amplitude components of all positive samples ( It is brought closer in cosine space, while being compared with the magnitude components of all different classes (negative samples). ( (Pushing it further away.) The mathematical expression for the amplitude loss term is: ; This loss causes the amplitude features of similar samples to be highly clustered on the hypersphere, while the amplitude features of dissimilar samples are far apart.

[0042] Phase loss term The design needs to overcome the periodicity ambiguity of the original phase. This is achieved using the aforementioned phase projection vector. Supervised contrastive learning is performed in the phase projection space using the same form as the amplitude loss: maximizing the similarity of phase projection vectors between samples of the same class and minimizing the similarity of phase projection vectors between samples of different classes. Its mathematical expression is: ; Ultimately, the joint magnitude-phase loss function is defined as a weighted sum of magnitude loss and phase loss: ; By minimizing The network can synchronously learn complex-valued features with amplitude and phase discriminative properties.

[0043] As can be seen, in this example, by extending supervised contrastive learning to the complex domain, joint optimization of amplitude and phase is achieved, fully preserving the phase semantic information of SAR images and overcoming the fundamental defect of information loss in real-valued models. The training gradient anomaly caused by phase periodicity ambiguity is solved by phase projection transformation. Complex-valued features simultaneously possess high intra-class compactness and high inter-class separability in both amplitude and phase dimensions, breaking through the upper limit of the feature expression capability of real-valued models and laying a solid foundation for subsequent high-reliability detection.

[0044] In one possible embodiment, the parameters of the complex-valued classification network are updated based on the cross-entropy loss function determined by the magnitude of the classification features, including: [See also...] Figure 3b Step 2 in the process. In this second stage, the complex-valued coding network (This stage of complex value encoding network) The parameters are completely frozen, such as Figure 3b (As indicated by the "parameter freeze" label in the image), complex value projection network No longer involved in calculations. Training samples Frozen After extracting the complex-valued features, they are directly input into the complex-valued classification network. Output 3D complex value classification features , Total number of ID categories. Then, based only on... The cross-entropy loss function is calculated based on the magnitude of each component, and its mathematical expression is as follows: ; in Given the total number of known ship categories, for The One component, | | represents the modulus of that component. ( The indicator function is 1 if the condition is met, and 0 otherwise. This loss function only constrains the magnitude of the classification feature to align with the true label, without imposing any constraint on the phase, thus fully preserving the phase discrimination information learned earlier through the joint magnitude-phase loss. Through backpropagation, only the following is updated: The network parameters are shown in Figure 3. The cross-entropy loss is related to the classifier (…). As shown in the connection, the output of this stage is the trained classification feature, which is used for subsequent generation of generalized quantum fundamental probability amplitude.

[0045] As can be seen in this example, the second stage, by freezing the encoder and optimizing the classifier based solely on the amplitude, fully preserves the phase information while ensuring that the classification confidence is aligned with the amplitude. This provides complex-valued features that are both highly discriminative and consistent with quantum theory inputs for subsequent quantum evidence mapping.

[0046] Step S103: Construct a quantum recognition framework.

[0047] The Quantum Frame of Discernment (QFOD) proposition set includes reference propositions for multiple known ship categories, and an empty set of propositions representing unknown ship categories. Specifically, let the total number of known ship categories (IDs) be... The quantum recognition framework is defined as follows: each For a known category, it is an orthogonal basis vector in Hilbert space. The power set of the quantum recognition framework. Includes all possible combinations of propositions, where the empty set This is explicitly defined as an out-of-distribution (OOD) state, i.e., an unknown ship category. Compared with existing technologies, this application incorporates the unknown category as an independent proposition into a formal framework, achieving explicit modeling of cognitive uncertainty.

[0048] Please refer to Figure 3c , Figure 3cThis is a schematic diagram of a generalized quantum evidence fundamental probability amplitude function generation module provided in an embodiment of this application. The purpose of constructing the quantum recognition framework is to provide a unified proposition space for the subsequent generation of the generalized quantum basic probability amplitude (GQBPA). Existing evidence models can only define the basic probability assignment in the real number domain and do not treat the out-of-determinacy (OOD) state as an independent proposition, resulting in incomplete uncertainty modeling. This application, by introducing the quantum recognition framework and the empty set proposition, for the first time unifies known classification and unknown detection under the same theoretical system.

[0049] Step S104: Using a multi-full complex-valued neural network, determine the generalized quantum fundamental probability amplitude corresponding to all propositions in the quantum recognition framework based on the corresponding complex-valued feature vectors.

[0050] In one possible embodiment, multiple fully complex-valued neural networks are used to determine the generalized quantum fundamental probability amplitude corresponding to all propositions in the quantum recognition framework based on their respective complex-valued feature vectors. This includes: determining the feature prototype of each known ship category based on pre-stored training samples; calculating the distance between the complex-valued feature vector and the feature prototype of each known ship category; calculating the reliability weight of the image to be detected belonging to each known ship category based on the Weibull distribution parameters pre-fitted for each known ship category and the distance; constructing an augmented complex-valued feature vector based on the reliability weight and the complex-valued feature vector, including the dimension corresponding to each known ship category and a dimension corresponding to the empty set proposition; and performing quantum probability amplification and normalization mapping on the augmented complex-valued feature vector to generate the generalized quantum fundamental probability amplitude.

[0051] In one possible embodiment, the augmented complex-valued eigenvector is subjected to quantum probability amplification and normalization mapping to generate a generalized quantum fundamental probability amplitude. This includes: transforming the augmented complex-valued eigenvector so that the magnitude of each complex-valued component in the augmented complex-valued eigenvector becomes an exponential function of the magnitude of the original component, while keeping the phase unchanged; normalizing the transformed augmented complex-valued eigenvector so that the sum of the squares of the magnitudes of all complex-valued components is 1; and using each normalized complex-valued component as the generalized quantum fundamental probability amplitude of the corresponding proposition in the quantum recognition framework.

[0052] The Weibull distribution parameters are obtained by fitting a three-parameter Weibull distribution to the feature distance set of training samples for each known ship category using the maximum likelihood estimation method. Specifically, for each known category... The process involves collecting all correctly classified samples from the training set and calculating the Euclidean distance between their complex-valued categorical features and the prototype features of that category, forming a distance sample set. Then, a three-parameter Weibull distribution is fitted using maximum likelihood estimation. For example, the parameters are obtained by numerically optimizing and maximizing the log-likelihood function, resulting in scale, shape, and location parameters. These parameters are used to calculate the reliability weights of the test samples: the closer a sample is to the prototype category's distance distribution within the main region of the training samples for that category, the closer the reliability weight is to 1; the farther the distance, the closer the weight is to 0.

[0053] Specifically, the feature prototype is determined as follows: for each known category Based on all correctly classified categories in the training set For each sample, calculate the centroid of its complex-valued categorical feature, and use it as the category. Prototype Center The formula is: ; in, Total number of ID categories; For all correctly classified categories in the training set The set of indexes for the samples; For set The number of samples; , for sample Complex value classification features; for A complex-valued vector, and Consistent dimensions, as a category The characteristic prototype.

[0054] Furthermore, for the test samples After the trained network is processed, it obtains Dimensional complex value characteristics Calculate its relationship with each category prototype Euclidean distance: .

[0055] The Weibull distribution parameters are obtained by fitting a three-parameter Weibull distribution to the feature distance set of training samples for each known class using the maximum likelihood estimation method. Specifically, for each class... Collect all samples correctly classified into that category in the training set, and calculate their correlation with the training set. The Euclidean distances are used to form a distance sample set. The probability density function of the Weibull distribution is defined as: ; in, .

[0056] The scaling parameter is solved using maximum likelihood estimation. Shape parameters and position parameters Specifically, this is obtained by taking the partial derivative of the log-likelihood function and numerically solving the system of nonlinear equations. After fitting, the test samples... Corresponding category Reliability weight Defined as: ; The closer the weight value is to 1, the more closely the sample conforms to the category. The closer the ID distribution is to 0, the more likely it is to be an OOD sample.

[0057] Based on this, an augmented complex-valued eigenvector is constructed. Its dimensions are from (Number of ID categories) augmented to +1, the newly added number +1 dimension corresponds to the empty set of OOD. For each ID category The values ​​for the corresponding dimensions in the augmented features are: ; The first corresponding to the empty set of OOD The +1 dimension is the sum of the products of all original feature components of the ID category and "1 - reliability weight", as shown in the following formula: ; This construction makes the magnitude of the OOD dimension dominant when the reliability weights of the sample for all ID categories approach 0, naturally highlighting the OOD characteristics of the sample.

[0058] Specifically, the process of performing quantum probability amplification and normalization mapping on the augmented complex-valued eigenvector to generate the generalized quantum fundamental probability amplitude (GQBPA) includes the following sub-steps: First, define the quantum probability amplification factor. For complex-valued components The magnification factor is: ; in, It is a very small positive numerical stability constant, taking the value of The core characteristic of this amplification factor is: for complex-valued components... After magnification, the result is Its amplitude is (Achieving exponential amplification and improving feature discrimination), its phase is the same as the original. The phases are completely consistent (phase information is fully preserved).

[0059] Among them, the normalization factor is calculated. : ; Then, GQBPA is generated. For the quantum recognition framework power set... any proposition in ; (i.e., the empty set), then: ; Among them, for ID category , The value is the amplified and normalized complex value of the j-th dimension of the augmented feature; for the OOD empty set , The value is the complex value of the augmented feature in dimension K+1 after amplification and normalization.

[0060] For other propositions (such as compound propositions), the value is 0. This GQBPA strictly satisfies the normalization axiom of the Generalized Quantum Evidence Theory (GQET): ; As can be seen, in this example, the design of reliability weights and augmented feature vectors decouples ID confidence from OOD uncertainty; statistical modeling based on the Weibull distribution replaces the traditional heuristic threshold, giving OOD discrimination an interpretable theoretical basis; exponential enhancement of feature amplitude and lossless preservation of phase are achieved through quantum probability amplification factors; and for the first time, end-to-end mapping from complex-valued deep features to generalized quantum fundamental probability amplitudes is realized, opening up a feasible path for the deep integration of quantum evidence theory and deep learning.

[0061] Step S105: According to the generalized quantum evidence combination rule, multiple sets of generalized quantum fundamental probability amplitudes are fused to obtain the fused generalized quantum fundamental probability distribution.

[0062] The fused generalized quantum fundamental probability distribution is the set of support levels for each proposition within the quantum recognition framework. This distribution integrates the outputs of multiple independent fully complex-valued neural networks, leveraging quantum interference effects to enhance consistent evidence and suppress conflicting evidence, thereby improving the robustness and reliability of decision-making.

[0063] In one possible embodiment, according to the generalized quantum evidence combination rule, multiple sets of generalized quantum fundamental probability amplitudes are fused to obtain a fused generalized quantum fundamental probability distribution, including: for each proposition in the quantum recognition framework, performing a fusion quantity calculation to obtain the corresponding fusion quantity; if the currently processed proposition is an empty set proposition, determining the complex value corresponding to the empty set proposition in each set of generalized quantum fundamental probability amplitudes; and calculating the product of all complex values, and determining the square of the modulus of the product as the fusion quantity of the empty set proposition; if the currently processed proposition is... When considering a reference proposition, determine the complex value corresponding to the reference proposition in each set of generalized quantum fundamental probability amplitudes, and determine the proposition combination that makes the intersection of a single set of propositions equal to the reference proposition being processed; and calculate the square of the modulus of the sum of the products of the corresponding complex values ​​under each proposition combination as the fusion amount of the reference proposition being processed; calculate the sum of the fusion amounts of the empty set proposition and each reference proposition to obtain the total fusion amount; divide the fusion amount corresponding to each empty set proposition or reference proposition by the total fusion amount to obtain the support of the corresponding proposition in the generalized quantum fundamental probability distribution after fusion.

[0064] For example, please refer to Figure 3d , Figure 3d This is a schematic diagram of the structure of a feature fusion and decision module provided in an embodiment of this application. Figure 3d As shown, the feature fusion and decision module includes multiple independent credible quantum evidence neural network models, a generalized quantum evidence combination unit, and a decision output unit. The generalized quantum evidence combination unit is used to perform generalized quantum evidence combination rule operations on multiple sets of generalized quantum basic probability amplitudes to obtain the fused generalized quantum basic probability distribution. The decision output unit is used to output the final ship detection result based on the fused probability distribution and the preset classification decision rules.

[0065] Furthermore, an ensemble learning framework is constructed using N (5 in the experiments of this application) independently trained Trustworthy QENN models, with each model initialized with different training parameters to ensure diversity among models; for the same test sample Each model independently outputs a set of GQBPAs, resulting in N independent sets of GQBPAs: . arrive Represents the 1st to Nth independent models in the ensemble learning framework (i.e. Figure 3d In F 1 , F 2 ,..., F N To differentiate the outputs of different models, the first... Sub-model The output GQBPA instantiation is represented as .Right now It is a general mapping function In a specific sub-model The specific calculation results are as follows.

[0066] For the GQBPA generated by the model, its corresponding Generalized Quantum Basic Probability Distribution (GQBPD) is defined as the square of its modulus. The Generalized Quantum Evidence Combination Rule (GQECR) is used to... The independent GQBPAs are fused to obtain the fused GQBPD. The core difference between this fused GQBPD and the existing DS evidence combination rules is that it simultaneously utilizes the amplitude and phase information of the evidence, introducing a quantum interference effect to effectively resolve the conflict problem of multi-source evidence. The fusion formula is as follows: For any proposition in the power set of the quantum recognition framework The merged GQBPD is: ; For the empty set of OOD The merged GQBPD is: ; in, , These are all propositions within the power set of the quantum recognition framework; The intersection of all single propositions represents the target proposition. The combination of these elements is used to extract joint consensus among multi-source information; For N sets of GQBPA propositions The product of the values.

[0067] Wherein, the distribution satisfies =1, and each It is a non-negative real number, representing the overall support for the proposition.

[0068] This fusion rule is the core operation of the Generalized Quantum Evidence Combination Rule (GQECR). It differs from the classic Dempster-Shafer (DS) combination rule in that the DS rule can only handle real-valued evidence and cannot resolve the paradox of highly conflicting evidence. This rule, however, introduces a quantum interference effect by utilizing the phase information of complex-valued evidence—when the phases of evidence from multiple models on the same proposition are close, summation produces coherent enhancement; when the phases are contradictory, they cancel each other out. This mechanism naturally resolves evidence conflicts, making the fusion result more robust.

[0069] The above integration and decision-making processes are all related to Figure 3d The feature fusion and decision-making modules shown work in complete accordance with each other.

[0070] As can be seen, in this example, deep fusion of multi-source complex-valued quantum evidence was achieved through GQECR, which made full use of the quantum interference effect in phase information and effectively solved the problem of anti-intuitive results in the fusion of high-conflict evidence in classical DS theory. The fused probability distribution integrates the decision information of multiple independent models, which significantly improves the robustness and accuracy of reasoning and lays a theoretical foundation for subsequent high-confidence decision-making.

[0071] Step S106: Determine the detection result based on the generalized quantum fundamental probability distribution and the preset classification decision rules.

[0072] The detection results are used to characterize whether the ship targets in the image to be detected belong to a known ship category or an unknown ship category.

[0073] In one possible embodiment, the detection result is determined based on the generalized quantum fundamental probability distribution and a preset classification decision rule, including: determining whether the support corresponding to the empty set proposition in the generalized quantum fundamental probability distribution is the maximum value; if the support corresponding to the empty set proposition is the maximum value, then the ship target is determined to be an unknown ship category; if the support corresponding to the empty set proposition is not the maximum value, then determining whether the maximum support in the known ship category proposition is greater than or equal to a preset confidence threshold; if the maximum support in the known ship category proposition is greater than or equal to the confidence threshold, then the ship target is determined to be a known ship category that obtains the maximum support; if the maximum support in the known ship category proposition is less than the confidence threshold, then the ship target is determined to be an unknown ship category; and generating and outputting the detection result based on the discrimination result.

[0074] The preset confidence threshold δ is adaptively determined using training data. Specifically, the maximum fusion support of all correctly classified samples in the training set is extracted.

[0075] For example, suppose the total number of correctly classified samples in the training set is . For each correctly classified training sample, calculate its maximum fusion support. ,in The generalized quantum fundamental probability distribution after fusion.

[0076] Furthermore, this The support values ​​are sorted in ascending order to obtain an ordered sequence. Confidence threshold Take the 10th percentile of the sequence, its mathematical expression is: ; in, This is a rounding function. This ensures that a predetermined proportion (preferably 90%) of known samples in the training data are correctly identified.

[0077] Furthermore, the final decision-making rules are as follows: ; The decision rule is explained as follows: If the proposition corresponding to the maximum support after fusion is an OOD empty set... If the maximum support of all propositions is lower than the confidence threshold δ, it indicates that the model's prediction confidence for this sample is insufficient, and it is determined to be OOD. In other cases, the target is the ID category corresponding to the maximum support. Finally, based on the discrimination results, the detection results are generated and output.

[0078] As can be seen, in this example, compared with existing technologies, this embodiment achieves more reliable open set decision-making by combining the maximum support judgment and confidence threshold judgment of the empty set proposition: direct rejection when multi-source quantum evidence consistently points to the unknown; rejection is also made when the confidence of the known class is insufficient, thereby effectively controlling the risk of false alarms. Unlike the existing "accuracy-rejection tradeoff" dilemma, this embodiment significantly improves the detection reliability of unknown targets while maintaining high classification accuracy for known classes.

[0079] As can be seen, in the method of this application, by introducing fully complex-valued neural networks and supervised complex-valued comparative learning into the field of open-set recognition for the first time, and by simultaneously optimizing the amplitude and phase discriminativeness of complex-valued features through a joint amplitude-phase loss function, the inherent phase semantic information of SAR images is fully preserved, fundamentally overcoming the inherent defects of existing real-valued models, such as insufficient information utilization and upper limits in feature representation. Furthermore, a quantum recognition framework incorporating the empty set proposition is constructed, and through reliability weights, quantum probability amplification, and normalization mapping, an end-to-end mapping from complex-valued deep features to generalized quantum fundamental probability amplitudes is achieved, formalizing unknown states as quantum empty sets. This is the first time that explicit and interpretable quantitative modeling of out-of-distribution uncertainty has been realized, replacing traditional heuristic thresholding methods. Moreover, based on the generalized quantum evidence combination rule, complex-valued quantum evidence from multiple independent network outputs is fused, utilizing quantum interference effects to enhance consistent evidence and suppress conflicting evidence. This effectively solves the counterintuitive problem of classical DS evidence theory in highly conflicting scenarios, significantly improving the robustness of multi-source information fusion.

[0080] The following are embodiments of the apparatus of this application. These embodiments of the apparatus and the embodiments of the method of this application belong to the same concept and are used to execute the methods described in the embodiments of this application. For ease of explanation, only the parts related to the apparatus embodiments of this application are shown in the embodiments of this application. For specific technical details not disclosed, please refer to the description of the embodiments of the method of this application, which will not be repeated here.

[0081] This application provides a ship detection device based on a trusted quantum evidence neural network. Specifically, the ship detection device based on the trusted quantum evidence neural network is used to execute the steps performed by the controller in the above-described ship detection method based on the trusted quantum evidence neural network. The ship detection device based on the trusted quantum evidence neural network in this application may include modules corresponding to the respective steps.

[0082] This application embodiment can divide the ship detection device based on the trusted quantum evidence neural network into functional modules according to the above method example. For example, each function can be divided into a separate functional module, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware or as a software functional module. The module division in this application embodiment is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0083] When dividing each function into modules according to its corresponding function. Figure 4 This is a functional block diagram of a ship detection device based on a trusted quantum evidence neural network, provided in an embodiment of this application. The ship detection device includes: an image acquisition unit 401 for acquiring synthetic aperture radar images as images to be detected; a vector extraction unit 402 for inputting the image to be detected into multiple fully complex value neural networks pre-trained with a minimized joint amplitude-phase loss function; a framework construction unit 403 for constructing a quantum recognition framework, wherein the proposition set of the quantum recognition framework includes multiple reference propositions of known ship categories and an empty set proposition representing unknown ship categories; and a distribution determination unit. 404 is used to determine the generalized quantum fundamental probability amplitude corresponding to all propositions in the quantum recognition framework based on the corresponding complex-valued feature vectors through multiple fully complex-valued neural networks; and to fuse multiple sets of generalized quantum fundamental probability amplitudes according to the generalized quantum evidence combination rule to obtain the fused generalized quantum fundamental probability distribution, which is the set of support corresponding to each proposition in the quantum recognition framework; the category detection unit 405 is used to determine the detection result according to the generalized quantum fundamental probability distribution and the preset classification decision rule, and the detection result is used to characterize whether the ship target in the image to be detected belongs to a known ship category or an unknown ship category.

[0084] In one possible embodiment, the fully complex valued neural network consists of a complex value encoding network, a complex value projection network, and a complex value classification network. Before inputting the image to be detected into multiple fully complex valued neural networks pre-trained with a minimized joint magnitude-phase loss function to extract corresponding complex value feature vectors, the vector extraction unit 402 is further configured to: input pre-stored training samples into the complex value encoding network to obtain a first complex value feature; input the first complex value feature into the complex value projection network to obtain a complex value projection vector; calculate the joint magnitude-phase loss function based on the complex value projection vector; perform backpropagation based on the joint magnitude-phase loss function to update the parameters of the complex value encoding network and the complex value projection network; input the training samples into the updated complex value encoding network to obtain a second complex value feature; input the second complex value feature into the complex value classification network to obtain classification features; and update the parameters of the complex value classification network based on the cross-entropy loss function determined by the magnitude of the classification features.

[0085] In one possible embodiment, in calculating the joint magnitude-phase loss function based on the complex-valued projection vector, the vector extraction unit 402 is specifically configured to: acquire a training batch used in a single parameter update iteration, the training batch containing multiple training samples carrying category labels of known ship categories; determine the magnitude component and original phase vector corresponding to the complex-valued projection vector of each training sample within the training batch; calculate the magnitude loss term based on the magnitude component using a normalized exponential function, wherein the calculation process is configured to maximize the magnitude similarity between training samples carrying the same category label and minimize the magnitude similarity between training samples carrying different category labels; and, For each phase component corresponding to each original phase vector, cosine and sine values ​​are calculated, and all calculation results are arranged sequentially to form the corresponding phase projection vector. Based on the phase projection vector, a phase loss term is calculated using a normalized exponential function. The calculation process is configured to maximize the similarity of phase projection vectors between training samples carrying the same category label and minimize the similarity of phase projection vectors between training samples carrying different category labels. The amplitude loss term and the phase loss term are weighted and summed according to a preset balance weight coefficient to obtain a joint amplitude-phase loss function. The balance weight coefficient is a fixed ratio parameter that adjusts the proportion of the amplitude loss term and the phase loss term.

[0086] In one possible embodiment, in determining the generalized quantum fundamental probability amplitude corresponding to all propositions in the quantum recognition framework through multiple fully complex-valued neural networks based on their respective complex-valued feature vectors, the distribution determination unit 404 is specifically used for: determining the feature prototype of each known ship category based on pre-stored training samples; calculating the distance between the complex-valued feature vector and the feature prototype of each known ship category; calculating the reliability weight of the image to be detected belonging to each known ship category based on the Weibull distribution parameters pre-fitted for each known ship category and the distance, wherein the Weibull distribution parameters are obtained by fitting a three-parameter Weibull distribution to the feature distance set of the training samples of each known ship category using the maximum likelihood estimation method; constructing an augmented complex-valued feature vector including the dimension corresponding to each known ship category and a dimension corresponding to the empty set proposition based on the reliability weight and the complex-valued feature vector; and performing quantum probability amplification and normalization mapping on the augmented complex-valued feature vector to generate the generalized quantum fundamental probability amplitude.

[0087] In one possible embodiment, in generating a generalized quantum fundamental probability amplitude by performing quantum probability amplification and normalization mapping on the augmented complex-valued eigenvector, the distribution determination unit 404: transforms the augmented complex-valued eigenvector so that the magnitude of each complex-valued component in the augmented complex-valued eigenvector becomes an exponential function of the magnitude of the original component, while keeping the phase unchanged; normalizes the transformed augmented complex-valued eigenvector so that the sum of the squares of the magnitudes of all complex-valued components is 1; and uses each normalized complex-valued component as the generalized quantum fundamental probability amplitude of the corresponding proposition in the quantum recognition framework.

[0088] In one possible embodiment, regarding the fusion of multiple sets of generalized quantum fundamental probability amplitudes according to the generalized quantum evidence combination rule to obtain the fused generalized quantum fundamental probability distribution, the distribution determination unit 404 is specifically used for: performing fusion quantity calculation for each proposition in the quantum recognition framework to obtain the corresponding fusion quantity; if the currently processed proposition is an empty set proposition, determining the complex value corresponding to the empty set proposition in each set of generalized quantum fundamental probability amplitudes; and calculating the product of all complex values, and determining the square of the modulus of the product as the fusion quantity of the empty set proposition; if when When the preprocessed proposition is a reference proposition, the complex value corresponding to the reference proposition in each group of generalized quantum fundamental probability amplitudes is determined, and the proposition combination that makes the intersection of a single group of propositions equal to the reference proposition being processed is determined; and the square of the modulus of the sum of the products of the corresponding complex values ​​under each proposition combination is calculated as the fusion amount of the reference proposition being processed; the sum of the fusion amounts of the empty set proposition and each reference proposition is calculated to obtain the total fusion amount; the fusion amount corresponding to each empty set proposition or reference proposition is divided by the total fusion amount to obtain the support of the corresponding proposition in the generalized quantum fundamental probability distribution after fusion.

[0089] In one possible embodiment, in determining the detection result based on the generalized quantum fundamental probability distribution and a preset classification decision rule, the category detection unit 405 is specifically used to: determine whether the support corresponding to the empty set proposition in the generalized quantum fundamental probability distribution is the maximum value; if the support corresponding to the empty set proposition is the maximum value, then determine that the ship target is an unknown ship category; if the support corresponding to the empty set proposition is not the maximum value, then determine whether the maximum support in the known ship category proposition is greater than or equal to a preset confidence threshold; if the maximum support in the known ship category proposition is greater than or equal to the confidence threshold, then determine that the ship target is a known ship category that obtains the maximum support; if the maximum support in the known ship category proposition is less than the confidence threshold, then determine that the ship target is an unknown ship category; and generate and output the detection result based on the discrimination result.

[0090] When using integrated units, such as Figure 5 As shown, Figure 5 This is a functional unit block diagram of another ship detection device based on a trusted quantum evidence neural network provided in this application embodiment. Figure 5 The ship detection device 40 based on a trusted quantum evidence neural network includes a processing module 502 and a communication module 501. The processing module 502 controls and manages the actions of the ship detection device 40, such as the steps of the image acquisition unit 401, vector extraction unit 402, frame construction unit 403, distribution determination unit 404, and category detection unit 405, and / or other processes for executing the techniques described herein. The communication module 501 supports interaction between the ship detection device based on the trusted quantum evidence neural network and other devices. Figure 5 As shown, the ship detection device based on the trusted quantum evidence neural network may include a storage module 503, which is used to store the program code and data of the ship detection device based on the trusted quantum evidence neural network.

[0091] The processing module 502 can be a processor or processing module, such as a central processing unit (CPU), a general-purpose processor, a digital signal processor (DSP), an ASIC, an FPGA, or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. The processor can also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc. The communication module 501 can be a transceiver, RF circuitry, or a communication interface, etc. The storage module 503 can be a memory.

[0092] All relevant content in each scenario involved in the above method embodiments can be referenced from the functional descriptions of the corresponding functional modules, and will not be repeated here. The ship detection device 40 based on the trusted quantum evidence neural network described above can perform the above-mentioned... Figure 1 The ship detection method shown is based on a trusted quantum evidence neural network.

[0093] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions according to the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0094] Figure 6 This is a structural block diagram of an electronic device provided in an embodiment of this application. For example... Figure 6 As shown, the electronic device 60 may include one or more of the following components: a processor 601 and a memory 602 coupled to the processor 601, wherein the memory 602 may store one or more computer programs 603, and the one or more computer programs 603 may be configured to implement the methods described in the above embodiments when executed by one or more processors 601.

[0095] Processor 601 may include one or more processing cores. Processor 601 connects to various parts within the electronic device 60 using various interfaces and lines, and performs various functions and processes data of the electronic device 60 by running or executing instructions, programs, code sets, or instruction sets stored in memory 602, and by calling data stored in memory 602. Optionally, processor 601 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). Processor 601 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the displayed content; and the modem handles wireless communication. It is understood that the modem may also not be integrated into processor 601 and may be implemented separately using a communication chip.

[0096] The memory 602 may include random access memory (RAM) or read-only memory (ROM). The memory 602 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 602 may include a program storage area and a data storage area. The program storage area may store instructions for implementing an operating system, instructions for implementing at least one function (such as touch functionality, sound playback functionality, image playback functionality, etc.), and instructions for implementing the various method embodiments described above. The data storage area may also store data created by the electronic device 60 during use.

[0097] It is understood that the electronic device 60 may include more or fewer structural elements than those shown in the above block diagram, without limitation herein.

[0098] This application also provides a computer storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements some or all of the steps of any of the methods described in the above method embodiments.

[0099] This application also provides a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program operable to cause a computer to perform some or all of the steps of any of the methods described in the above method embodiments.

[0100] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply 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 embodiments of this application.

[0101] In the several embodiments provided in this application, it should be understood that the disclosed methods, apparatuses, and systems can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for example, the division of units is merely a logical functional division, and there may be other division methods in actual implementation; for example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0102] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0103] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can be physically comprised separately, or two or more units can be integrated into one unit. The integrated unit described above can be implemented in hardware or in the form of hardware plus software functional units.

[0104] The integrated units implemented as software functional units described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute partial steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes: a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, volatile memory, or non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM), etc., which are various media capable of storing program code.

[0105] While the present invention has been disclosed above, it is not limited thereto. Any person skilled in the art can easily conceive of variations or substitutions without departing from the spirit and scope of the present invention, and various modifications and alterations can be made, including combinations of the different functions and implementation steps described above, as well as software and hardware implementation methods, all of which are within the protection scope of the present invention.

Claims

1. A ship detection method based on a trusted quantum evidence neural network, characterized in that, include: Acquire synthetic aperture radar images as the images to be detected; The image to be detected is input into multiple fully complex value neural networks pre-trained with a joint magnitude-phase loss function to extract corresponding complex value feature vectors. The joint magnitude-phase loss function is: ； in, For amplitude loss, For phase loss term, The weighting coefficient between the amplitude loss term and the phase loss term is: ； The phase loss term is: ； in, For the set of sample indices within the training batch, For samples within the batch and anchor points A collection of sample indexes that are of the same category but not themselves. To traverse the set of positive samples index variable, To traverse the set of all samples except the anchor point sample index variable, The cosine similarity function is used. For the anchor point sample The magnitude components of the complex-valued projection vector. Positive sample amplitude components, negative samples The magnitude vector, For the anchor point sample The phase projection vector, The positive sample The phase projection vector, For the negative sample The phase projection vector, This refers to temperature hyperparameters. A quantum recognition framework is constructed, wherein the proposition set of the quantum recognition framework includes reference propositions of multiple known ship categories, and empty set propositions representing unknown ship categories; The generalized quantum fundamental probability amplitude corresponding to all propositions in the quantum recognition framework is determined by the multiple fully complex-valued neural networks based on the corresponding complex-valued feature vectors, including: determining the feature prototype of each known ship category based on pre-stored training samples; Calculate the distance between the complex-valued feature vector and the feature prototype of each known ship category; Based on the Weibull distribution parameters prefitted for each of the known ship categories and the distance, the reliability weight of the image to be detected belonging to each of the known ship categories is calculated. The Weibull distribution parameters are obtained by fitting a three-parameter Weibull distribution to the feature distance set of the training samples of each of the known ship categories using the maximum likelihood estimation method. Based on the reliability weights and the complex-valued feature vectors, an augmented complex-valued feature vector is constructed, which includes the dimension corresponding to each of the known ship categories and a dimension corresponding to the empty set proposition. The augmented complex-valued eigenvector is transformed so that the magnitude of each complex-valued component in the augmented complex-valued eigenvector becomes an exponential function of the magnitude of the original component, while the phase remains unchanged; The transformed augmented complex-valued eigenvector is normalized so that the sum of the squares of the magnitudes of all the complex-valued components is 1. Each of the normalized complex numerical components is used as the generalized quantum fundamental probability amplitude of the corresponding proposition in the quantum recognition framework, wherein a numerical stability constant is introduced in the normalization. According to the generalized quantum evidence combination rule, multiple sets of the generalized quantum fundamental probability amplitudes are fused to obtain the fused generalized quantum fundamental probability distribution, including: for each proposition in the quantum recognition framework, performing a fusion quantity calculation to obtain the corresponding fusion quantity: If the proposition being processed is the empty set proposition, determine the complex value corresponding to the empty set proposition in each group of generalized quantum fundamental probability amplitudes; and calculate the product of all the complex values, and determine the square of the modulus of the product as the fusion amount of the empty set proposition. If the proposition being processed is the reference proposition, determine the complex value corresponding to the reference proposition in each group of generalized quantum fundamental probability amplitudes, and determine the proposition combination that makes the intersection of a single group of propositions equal to the reference proposition being processed; and calculate the square of the modulus of the sum of the product of the corresponding complex values ​​under each proposition combination as the fusion amount of the reference proposition being processed. The sum of the fusion amounts of the empty set proposition and each of the reference propositions is calculated to obtain the total fusion amount, wherein the fused generalized quantum fundamental probability distribution is obtained by fusing multi-source evidence from multiple fully complex valued neural networks; Divide the fusion amount corresponding to each empty set proposition or the reference proposition by the sum of the fusion amounts to obtain the support of the corresponding proposition in the generalized quantum fundamental probability distribution after fusion. Based on the generalized quantum fundamental probability distribution and the preset classification decision rule, the detection result is determined. The detection result is used to characterize whether the ship target in the image to be detected belongs to the known ship category or the unknown ship category. The confidence threshold in the preset classification decision rule is adaptively determined by extracting the maximum fusion support of all correctly classified samples in the training set, and taking the 10th percentile after sorting them in ascending order as the confidence threshold.

2. The method according to claim 1, characterized in that, The fully complex value neural network consists of a complex value encoding network, a complex value projection network, and a complex value classification network; before inputting the image to be detected into multiple fully complex value neural networks pre-trained with a minimized joint magnitude-phase loss function to extract the corresponding complex value feature vectors, the method further includes: The pre-stored training samples are input into the complex value encoding network to obtain the first complex value feature; The first complex-valued feature is input into the complex-valued projection network to obtain a complex-valued projection vector; Calculate the joint magnitude-phase loss function based on the complex-valued projection vector; Backpropagation is performed based on the joint magnitude-phase loss function to update the parameters of the complex-valued coding network and the complex-valued projection network; The training samples are input into the updated complex value encoding network to obtain the second complex value feature; The second complex-valued feature is input into the complex-valued classification network to obtain classification features; The parameters of the complex-valued classification network are updated based on the cross-entropy loss function determined by the magnitude of the classification features.

3. The method according to claim 2, characterized in that, The step of calculating the joint magnitude-phase loss function based on the complex-valued projection vector includes: Obtain a training batch used in a single parameter update iteration, wherein the training batch contains multiple training samples carrying category labels of the known ship categories; Determine the magnitude component and original phase vector corresponding to the complex-valued projection vector of each training sample in the training batch; Based on the amplitude components, an amplitude loss term is calculated using a normalized exponential function, wherein the calculation process is configured to maximize the amplitude similarity between training samples carrying the same class label and minimize the amplitude similarity between training samples carrying different class labels; and, For each phase component corresponding to each original phase vector, calculate the cosine and sine values ​​respectively, and arrange all the calculation results in order to form the corresponding phase projection vector; Based on the phase projection vector, the phase loss term is calculated through a normalized exponential function, wherein the calculation process is configured to maximize the phase projection vector similarity between training samples carrying the same category label and minimize the phase projection vector similarity between training samples carrying different category labels. The amplitude loss term and the phase loss term are weighted and summed according to a preset balance weight coefficient to obtain the joint amplitude-phase loss function. The balance weight coefficient is a fixed ratio parameter that adjusts the proportion of the amplitude loss term and the phase loss term.

4. The method according to claim 1, characterized in that, The step of determining the detection result based on the generalized quantum fundamental probability distribution and the preset classification decision rules includes: Determine whether the support degree corresponding to the empty set proposition in the generalized quantum fundamental probability distribution is the maximum value; If the support corresponding to the empty set proposition is the maximum value, then the ship target is determined to be the unknown ship category; If the support corresponding to the empty set proposition is not the maximum value, then determine whether the maximum support in the known ship category proposition is greater than or equal to the preset confidence threshold. If the maximum support in the known ship category proposition is greater than or equal to the confidence threshold, then the ship target is determined to be the known ship category that obtains the maximum support; If the maximum support of the known ship category proposition is less than the confidence threshold, then the ship target is determined to be the unknown ship category; Based on the discrimination result, the detection result is generated and output.

5. A ship detection device based on a trusted quantum evidence neural network, characterized in that, include: An image acquisition unit is used to acquire synthetic aperture radar images as images to be detected. The vector extraction unit is used to input the image to be detected into multiple fully complex value neural networks pre-trained with a joint magnitude-phase loss function to extract corresponding complex value feature vectors. The joint magnitude-phase loss function is: ； in, For amplitude loss, For phase loss term, The weighting coefficient between the amplitude loss term and the phase loss term is: ； The phase loss term is: ； in, For the set of sample indices within the training batch, For samples within the batch and anchor points A collection of sample indexes that are of the same category but not themselves. To traverse the set of positive samples index variable, To traverse the set of all samples except the anchor point sample index variable, The cosine similarity function is used. For the anchor point sample The magnitude components of the complex-valued projection vector. Positive sample amplitude components, negative samples The magnitude vector, For the anchor point sample The phase projection vector, The positive sample The phase projection vector, For the negative sample The phase projection vector, This refers to temperature hyperparameters. A framework building unit is used to construct a quantum recognition framework. The proposition set of the quantum recognition framework includes reference propositions for multiple known ship categories, as well as empty set propositions representing unknown ship categories. The distribution determination unit is used to determine the generalized quantum fundamental probability amplitude corresponding to all propositions in the quantum recognition framework based on the corresponding complex-valued feature vectors through the multiple fully complex-valued neural networks. This includes: determining the feature prototype of each known ship category based on pre-stored training samples; calculating the distance between the complex-valued feature vector and the feature prototype of each known ship category; and calculating the reliability weight of the image to be detected belonging to each known ship category based on the Weibull distribution parameters pre-fitted for each known ship category and the distances. The Weibull distribution parameters are obtained by using the maximum likelihood estimation method to perform a three-parameter Weibull distribution on the feature distance set of the training samples for each known ship category. The distribution is fitted to obtain the result; based on the reliability weights and the complex-valued feature vector, an augmented complex-valued feature vector is constructed, including the dimension corresponding to each of the known ship categories and a dimension corresponding to the empty set proposition; the augmented complex-valued feature vector is transformed so that the magnitude of each complex-valued component in the augmented complex-valued feature vector becomes an exponential function of the magnitude of the original component, while the phase remains unchanged; the transformed augmented complex-valued feature vector is normalized so that the sum of the squares of the magnitudes of all the complex-valued components is 1; each of the normalized complex-valued components is used as the generalized quantum fundamental probability amplitude of the corresponding proposition in the quantum recognition framework, wherein a numerical stability constant is introduced in the normalization; and, According to the generalized quantum evidence combination rule, multiple sets of generalized quantum fundamental probability amplitudes are fused to obtain a fused generalized quantum fundamental probability distribution, including: for each proposition in the quantum recognition framework, performing a fusion quantity calculation to obtain the corresponding fusion quantity; if the currently processed proposition is the empty set proposition, determining the complex value corresponding to the empty set proposition in each set of generalized quantum fundamental probability amplitudes; and calculating the product of all the complex values, and determining the square of the modulus of the product as the fusion quantity of the empty set proposition; if the currently processed proposition is the reference proposition, determining the complex value corresponding to the reference proposition in each set of generalized quantum fundamental probability amplitudes. The method involves considering the complex values ​​corresponding to the propositions, determining the proposition combinations that make the intersection of a single set of propositions equal to the reference proposition being processed, and calculating the square of the modulus of the sum of the products of the corresponding complex values ​​under each proposition combination as the fusion amount of the reference proposition being processed; calculating the sum of the fusion amounts of the empty set proposition and each of the reference propositions to obtain the total fusion amount, wherein the fused generalized quantum fundamental probability distribution is obtained by fusing multi-source evidence from multiple fully complex valued neural networks; dividing the fusion amount corresponding to each empty set proposition or the reference proposition by the total fusion amount to obtain the support of the corresponding proposition in the fused generalized quantum fundamental probability distribution; A category detection unit is used to determine the detection result based on the generalized quantum fundamental probability distribution and a preset classification decision rule. The detection result is used to characterize whether the ship target in the image to be detected belongs to the known ship category or the unknown ship category. The confidence threshold in the preset classification decision rule is adaptively determined by extracting the maximum fusion support of all correctly classified samples in the training set, and taking the 10th percentile after sorting them in ascending order as the confidence threshold.

6. An electronic device, characterized in that, The method includes a processor, a memory, a communication interface, and one or more programs, said programs being stored in the memory and configured to be executed by the processor, said programs including instructions for performing the steps of the method as described in any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, A computer program for storing electronic data interchange is provided, wherein the computer program causes a computer to perform the method as described in any one of claims 1-4.

Images