A picture target recognition method, device, equipment and medium

Abstract

The application discloses a picture target recognition method, device, equipment and medium, first, the input picture is converted into a high-dimensional visual feature vector, dimension reduction is carried out through a learnable sparse projection matrix to obtain a low-dimensional embedding representation, then the low-dimensional embedding representation is mapped to the embedding space of a large language model with frozen parameters through a lightweight adaptive network to extract semantic enhancement features, and the semantic enhancement features are fused with visual features output by a visual feature extraction network; then the fused features are encoded into quantum states, evolution and measurement are carried out through a parameterized quantum circuit, the probability distribution of the target category is obtained and the prediction uncertainty is calculated, finally, the sparse projection matrix, the adaptive network and the quantum circuit parameters are cooperatively optimized based on a joint loss function, the target recognition with efficient calculation, semantic enhancement and uncertainty estimation capability is realized, and the energy consumption is significantly reduced while the recognition accuracy is ensured.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more specifically, to a method, apparatus, device, and medium for image target recognition. Background Technology

[0002] With the tremendous success of deep learning, especially convolutional neural networks and their variants, in object recognition, the complexity and computational demands of these models have increased exponentially. Traditional CNNs rely on stacked convolutional layers and large-scale fully connected layers for feature extraction and classification, resulting in a massive number of model parameters and extremely high computational energy consumption. This computationally intensive nature makes it difficult to deploy existing high-performance models on resource-constrained edge devices such as mobile phones, IoT devices, and drones. Although techniques such as model pruning and quantization have been used to alleviate this problem, these methods often come at the cost of sacrificing model accuracy and robustness, and inherently still involve a large amount of redundant computation, failing to fundamentally meet the stringent requirements of low power consumption and high real-time performance on edge devices.

[0003] Secondly, existing visual models suffer from significant bottlenecks in semantic understanding and generalization capabilities in complex scenes. Mainstream methods focus on learning local features from pixel-level data, lacking the ability to deeply model contextual relationships between objects, scene logic, and higher-order semantics. For example, model performance drops sharply when faced with occlusion, abrupt changes in lighting, unknown categories, or long-tailed data distributions. This stems from the limitations of their visual modalities, failing to effectively incorporate the rich semantic information contained in human knowledge or large-scale text corpora. Consequently, the model's understanding of images remains at the surface feature level rather than the conceptual level, limiting its practical application in open environments.

[0004] Finally, existing target recognition systems face severe challenges in terms of reliability, interpretability, and system architecture. On the one hand, the models' predictions are mostly deterministic category outputs, lacking reliable estimates of their own prediction confidence and uncertainty quantification, making their application in safety-critical fields extremely risky and turning them into untrustworthy black boxes. On the other hand, in terms of computing architecture, existing edge-cloud collaborative solutions typically only perform simple task division, failing to fully utilize the powerful computing power of the cloud and the real-time advantages of the edge. The lack of a dynamic, adaptive, and efficient collaborative mechanism among the three results in the overall system failing to achieve an optimal balance in energy efficiency, speed, and performance. Summary of the Invention

[0005] The purpose of this application is to overcome the shortcomings of existing technologies and provide an image target recognition method, device, equipment and medium. It realizes a target recognition method with high efficiency, low energy consumption and interpretable uncertainty estimation. Experimental results show that it improves the accuracy of traditional CNN by 4.7% and reduces inference energy consumption by 62% on the COCO dataset. It is suitable for edge, cloud and hybrid computing scenarios.

[0006] The objective of this application is achieved through the following technical solution:

[0007] Firstly, this application proposes an image target recognition method, the method comprising:

[0008] The input image is converted into a high-dimensional visual feature vector, and the high-dimensional visual feature vector is reduced in dimensionality using a learnable sparse projection matrix to obtain a low-dimensional embedding representation.

[0009] A lightweight adapting network maps low-dimensional embedding representations to the embedding space of a large language model with frozen parameters, extracts semantic enhancement features, and fuses the semantic enhancement features with the visual features output by the visual feature extraction network to generate fused features.

[0010] The fused features are encoded into quantum states, and the quantum states are evolved and measured through parameterized quantum circuits to obtain the probability distribution of the target category. The prediction uncertainty is then calculated based on the probability distribution.

[0011] Based on the constructed joint loss function, the parameters of the sparse projection matrix, lightweight adaptation network, and parameterized quantum circuit are optimized in a coordinated manner to minimize the target recognition error and prediction uncertainty.

[0012] In one possible embodiment, the steps of converting the input image into a high-dimensional visual feature vector and then reducing the dimensionality of the high-dimensional visual feature vector using a learnable sparse projection matrix to obtain a low-dimensional embedding representation include:

[0013] Perform superpixel segmentation or block processing on the input image to form an image tensor. , For the set of real numbers, Image height, Image width, This refers to the number of image channels;

[0014] Expand the image tensor into a one-dimensional vector. , For tensor flattening operations;

[0015] Through a learnable sparse random projection matrix For a one-dimensional vector Compression is performed to obtain low-dimensional embeddings. , It is the target dimension after dimensionality reduction.

[0016] In one possible embodiment, after obtaining the low-dimensional embedding, the method further includes:

[0017] For low-dimensional embedding representation Apply constraints: ,in For group sparse regularization terms, The sparsity threshold is adjustable.

[0018] In one possible embodiment, the steps of mapping low-dimensional embedding representations to the embedding space of a large language model with frozen parameters using a lightweight adapting network, extracting semantic enhancement features, and fusing the semantic enhancement features with the visual features output by the visual feature extraction network to generate fused features include:

[0019] A lightweight adapting network is used to map low-dimensional embedding representations to the word embedding space of a large language model with frozen parameters, generating semantically enhanced features.

[0020] Extracting the first category logical value from the decoding end of the large language model Simultaneously, the second category logical value is obtained from the visual feature extraction network. ;

[0021] First category logical value With second category logical value Perform dynamic gating fusion to generate fused features;

[0022] The dynamic gating fusion method uses a weighted sum of the first-class and second-class logical values ​​through gating weights. The gating weights are determined by the first-class logical value... Second category logical value The results were obtained through joint calculations.

[0023] In one possible embodiment, the steps of encoding the fused features into quantum states, evolving and measuring the quantum states using parameterized quantum circuits to obtain the probability distribution of the target category, and calculating the prediction uncertainty based on the probability distribution include:

[0024] The fused features are encoded into quantum states, and the quantum states are evolved through parameterized quantum circuits to obtain the evolved quantum states.

[0025] The evolved quantum state is repeatedly sampled under a preset measurement basis to obtain the probability distribution of the target category;

[0026] The quantum Shannon entropy is calculated based on the probability distribution and used as the model's prediction uncertainty for the sample.

[0027] If the prediction uncertainty exceeds the preset uncertainty threshold, an active learning strategy is triggered, and the sample is added to the next training set.

[0028] In one possible embodiment, the joint loss function is: , For classifying losses, To control the weights of sparse regularization, To control quantum entropy regularity, To control the distribution divergence, For real labels, The merged category logical value, For group sparse regularization terms, For Shannon entropy, Let KL divergence be the KL divergence. This represents the probability distribution of the quantum circuit output. This represents the probability distribution of the output of a traditional CNN path.

[0029] In one possible embodiment, the step of co-optimizing the parameters of the sparse projection matrix, the lightweight adaptation network, and the parameterized quantum circuit based on the constructed joint loss function to minimize the target recognition error and prediction uncertainty includes:

[0030] An alternating training strategy is adopted between GPU and quantum coprocessor. Quantum natural gradient descent and zero-order gradient estimation methods are used to jointly optimize the sparse projection matrix parameters, lightweight adaptation network parameters and parameterized quantum circuit parameters, so as to simultaneously minimize the target recognition error and prediction uncertainty.

[0031] Secondly, this application proposes an image target recognition device, the device comprising:

[0032] The conversion module is used to convert the input image into a high-dimensional visual feature vector, and to reduce the dimensionality of the high-dimensional visual feature vector by using a learnable sparse projection matrix to obtain a low-dimensional embedding representation.

[0033] The fusion module is used to map low-dimensional embedding representations to the embedding space of a large language model with frozen parameters through a lightweight adaptation network, extract semantic enhancement features, and fuse the semantic enhancement features with the visual features output by the visual feature extraction network to generate fused features.

[0034] The encoding module is used to encode the fused features into quantum states, evolve and measure the quantum states through parameterized quantum circuits, obtain the probability distribution of the target category, and calculate the prediction uncertainty based on the probability distribution.

[0035] The collaborative optimization module is used to collaboratively optimize the parameters of the sparse projection matrix, lightweight adaptation network, and parameterized quantum circuit based on the constructed joint loss function, thereby minimizing target recognition error and prediction uncertainty.

[0036] Thirdly, this application also proposes a computer device comprising a processor and a memory, wherein the memory stores a computer program, which is loaded and executed by the processor to implement the image target recognition method as described in any of the first aspects.

[0037] Fourthly, this application also proposes a computer-readable storage medium storing a computer program that is loaded and executed by a processor to implement the image target recognition method as described in any of the first aspects.

[0038] The main solution and its various further alternatives described above can be freely combined to form multiple solutions, all of which are solutions that can be adopted and are claimed in this application; furthermore, the (non-conflicting alternatives) can also be freely combined with each other and with other alternatives. Those skilled in the art, after understanding the solution of this application, will realize from the prior art and common general knowledge that there are many combinations, all of which are technical solutions to be protected in this application, and will not be exhaustively listed here.

[0039] This application discloses an image target recognition method, apparatus, device, and medium. First, the input image is converted into a high-dimensional visual feature vector. Then, a low-dimensional embedding representation is obtained by dimensionality reduction using a learnable sparse projection matrix. Next, a lightweight adaptation network maps the low-dimensional embedding representation to the embedding space of a large language model with frozen parameters to extract semantically enhanced features, which are then fused with the visual features output by the visual feature extraction network. The fused features are then encoded into quantum states, evolved and measured using a parameterized quantum circuit to obtain the probability distribution of the target category and calculate the prediction uncertainty. Finally, the parameters of the sparse projection matrix, adaptation network, and quantum circuit are synergistically optimized based on a joint loss function. This achieves computationally efficient, semantically enhanced target recognition with uncertainty estimation capabilities, significantly reducing energy consumption while maintaining recognition accuracy. Attached Figure Description

[0040] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 A flowchart illustrating an image target recognition method proposed in an embodiment of this application is shown. Detailed Implementation

[0042] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0043] Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0044] In existing technologies, convolutional neural networks generally suffer from the following problems: First, high energy consumption and redundant computation. Traditional CNNs rely on large-scale convolution and fully connected operations, resulting in significant computational redundancy and making them unsuitable for low-power edge devices. Second, insufficient semantic representation. Visual models have limited generalization ability in complex scenes and struggle to effectively model high-order semantic relationships. Third, lack of uncertainty estimation. Existing recognition models are mostly "black box" predictions, lacking reasonable estimates of prediction confidence and uncertainty. Fourth, bottleneck in edge-cloud collaboration. Efficient collaboration between edge, cloud, and quantum coprocessors is difficult to achieve during model inference.

[0045] Therefore, in order to solve the above-mentioned technical problems, this application proposes an image target recognition method, device, equipment and medium, which combines the high efficiency of matrix compression, the semantic enhancement capability of large language models and the interpretability of quantum uncertainty estimation to achieve low-energy and robust target recognition.

[0046] Please refer to Figure 1 , Figure 1 This paper illustrates a flowchart of an image target recognition method proposed in an embodiment of this application. The method includes:

[0047] Step S1: Convert the input image into a high-dimensional visual feature vector, and use a learnable sparse projection matrix to reduce the dimensionality of the high-dimensional visual feature vector to obtain a low-dimensional embedding representation.

[0048] First, the input image is converted into a high-dimensional visual feature vector. Then, a learnable sparse projection matrix is ​​used to reduce the dimensionality of the high-dimensional vector, resulting in a low-dimensional embedding representation containing key visual information. This significantly reduces the amount of data required for subsequent computation while retaining the main features needed for recognition.

[0049] Step S1 includes:

[0050] Perform superpixel segmentation or block processing on the input image to form an image tensor. , For the set of real numbers, Image height, Image width, This refers to the number of image channels;

[0051] Expand the image tensor into a one-dimensional vector. , For tensor flattening operations;

[0052] Through a learnable sparse random projection matrix For a one-dimensional vector Compression is performed to obtain low-dimensional embeddings. , It is the target dimension after dimensionality reduction.

[0053] This step first preprocesses the input image by using superpixel segmentation or regular block processing to structurally represent the image as a three-dimensional tensor. The tensor is then flattened to convert it into a one-dimensional vector. This is to adapt to subsequent linear projection operations. The core compression process uses a learnable sparse random projection matrix. Achievement, including the target dimension Much smaller than the original dimension (Right now (Through matrix multiplication) This generates a low-dimensional embedding containing core visual information. The matrix here... It is initialized as a sparse matrix (containing only a small number of non-zero elements) and optimized during training, thereby preserving key features while greatly reducing computation and storage requirements.

[0054] After obtaining the low-dimensional embedding, it also includes:

[0055] For low-dimensional embedding representation Apply constraints: ,in For group sparse regularization terms, The sparsity threshold is adjustable.

[0056] After obtaining the low-dimensional embedding representation This is followed by structured sparsity constraints: by applying a set of sparsity regularization terms. To achieve feature selection, where This represents the L2,1 norm, a constraint that prompts the model to automatically identify and eliminate redundant dimensions in the embedding; The preset sparsity threshold is adjusted by... The value can effectively control the balance between feature compression rate and information fidelity, ensuring that visual features that are crucial to the recognition task are preserved while reducing dimensionality.

[0057] Step S2: Map the low-dimensional embedding representation to the embedding space of the large language model with frozen parameters through a lightweight adaptation network, extract semantic enhancement features, and fuse the semantic enhancement features with the visual features output by the visual feature extraction network to generate fused features.

[0058] A lightweight adapting network maps the compressed low-dimensional embedding representation to the embedding space of a large language model with frozen parameters, thereby extracting context-rich enhanced features. Subsequently, the semantically enhanced features are dynamically gated and fused with the visual features output by a visual feature extraction network (such as CNN) to generate fused features that combine visual details and deep semantic information.

[0059] Step S2 includes:

[0060] A lightweight adapting network is used to map low-dimensional embedding representations to the word embedding space of a large language model with frozen parameters, generating semantically enhanced features.

[0061] Extracting the first category logical value from the decoding end of the large language model Simultaneously, the second category logical value is obtained from the visual feature extraction network. ;

[0062] First category logical value With second category logical value Perform dynamic gating fusion to generate fused features;

[0063] The dynamic gating fusion method uses a weighted sum of the first-class and second-class logical values ​​through gating weights. The gating weights are determined by the first-class logical value... Second category logical value The results were obtained through joint calculations.

[0064] The low-dimensional embedding vector obtained after matrix-domain compression The adapter module uses a lightweight trainable adapter. , embed low-dimensional vector Mapped into the word embedding space of a Large Language Model (LLM) with frozen parameters. Adapter Typically composed of simple linear layers or small neural networks, the purpose is to embed visual data. Convert to a format compatible with LLM word embeddings without changing the parameters of the LLM itself. Generate semantically enhanced feature vectors. ,Right now , This represents the implicit representation of an image after LLM semantic enhancement, containing high-level semantic information that may be missing in traditional visual features.

[0065] semantic vectors The input is fed into the frozen LLM and processed by the decoder part of the LLM. The LLM will then process the semantic vector. Generate output logits (i.e., unnormalized class scores) that are associated with the target class, denoted as the first class logit. A traditional convolutional neural network (CNN) is used to process the original image or intermediate visual features, extract visual features, and output category logits, denoted as the second category logical value. .

[0066] First category logical value Second category logical value Perform dynamic gating fusion: Gating weights From the first category of logical values Second category logical value It was calculated jointly. Specifically, it will be... and After concatenation or summation, pass through a Sigmoid function ( Generate gate weights ,Right now Using gating weights right and Perform dynamic weighted summation to generate the fused category logical value. Dynamic gating fusion allows the model to adaptively adjust the contributions of LLM and CNN. When the semantic predictions of LLM are more reliable, As the value approaches 1, the fusion result becomes more dependent on... When the visual features of CNNs are more reliable, Approaching 0, the fusion result depends more on This mechanism improves the model's generalization ability across diverse scenarios.

[0067] Step S3: Encode the fused features into quantum states, evolve and measure the quantum states through parameterized quantum circuits to obtain the probability distribution of the target category, and calculate the prediction uncertainty based on the probability distribution.

[0068] The fused semantic features are encoded into quantum states and then evolved using parameterized quantum circuits. This process leverages the properties of quantum superposition and entanglement to extract deep information from the features. Subsequently, the evolved quantum states are repeatedly measured under a computational basis, and the probability distribution of the target category is obtained through statistical analysis of the measurement results. Based on the obtained probability distribution, quantum Shannon entropy is calculated as a quantitative indicator of prediction uncertainty. This entropy value effectively reflects the model's confidence in the recognition results: a higher entropy value indicates greater uncertainty in the model's prediction.

[0069] Step S3 includes:

[0070] The fused features are encoded into quantum states, and the quantum states are evolved through parameterized quantum circuits to obtain the evolved quantum states.

[0071] The evolved quantum state is repeatedly sampled under a preset measurement basis to obtain the probability distribution of the target category;

[0072] The quantum Shannon entropy is calculated based on the probability distribution and used as the model's prediction uncertainty for the sample.

[0073] If the prediction uncertainty exceeds the preset uncertainty threshold, an active learning strategy is triggered, and the sample is added to the next training set.

[0074] Step S3 input is the fused feature vector output from the previous stage. It can be transformed into the state of a quantum mechanical system, i.e., a quantum state, by using methods such as amplitude encoding or angle encoding. . For example, let the vector be the amplitude vector of n qubits. Each value, after a linear transformation, is mapped to the superposition state amplitude or rotation angle of n qubits. Parameterized quantum circuit. It is a computational module consisting of a series of pre-defined quantum logic gates, some of which contain trainable parameters. The encoded quantum state The input is fed into a quantum circuit. The quantum gates in the circuit perform a series of superposition, interference, and entanglement operations (evolution) on the quantum states. This transformation can uncover complex relationships between features that are difficult for classical neural networks to learn. Parameters The transformation is optimized through training to make it most advantageous for the target recognition task.

[0075] After the evolution is complete, the final quantum state is repeatedly sampled and measured using the calculated basis vectors. Due to the inherent probabilistic nature of quantum mechanics, each measurement is a random variable. Through hundreds or thousands of repeated measurements, the frequency of each possible output state can be statistically determined. In this application, these output states are designed to be related to the target category. Correlation. Through a pre-defined mapping relationship, these state frequencies are converted into probability distributions of the model for each category. .

[0076] Based on the obtained probability distribution Calculate the Shannon entropy using the following formula: .entropy Here, it is defined as the model's prediction uncertainty. If the probability distribution is highly concentrated, then the entropy value... It will be very low. If the probability distribution is very even, then the entropy value... It will be very high. The system sets a preset uncertainty threshold. When the calculated entropy value When this threshold is exceeded (i.e.) > If the system determines that the current sample is information-rich and difficult for the model to judge, it automatically triggers an active learning strategy and adds the sample to the training set for the next round.

[0077] Step S4: Based on the constructed joint loss function, the parameters of the sparse projection matrix, lightweight adaptation network, and parameterized quantum circuit are optimized in a coordinated manner to minimize the target recognition error and prediction uncertainty.

[0078] First, construct the joint loss function: , For classification loss (such as cross-entropy), it measures the difference between the prediction and the true label. The gap To control the weights of sparse regularization, To control quantum entropy regularity and avoid overconfidence, To control the distribution divergence and align the quantum distribution with the CNN distribution, For real labels, The merged category logical value, For group sparse regularization terms, For Shannon entropy, Let KL divergence be the KL divergence. This represents the probability distribution of the quantum circuit output. This represents the probability distribution of the output of a traditional CNN path.

[0079] Step S4 includes:

[0080] An alternating training strategy is adopted between GPU and quantum coprocessor. Quantum natural gradient descent and zero-order gradient estimation methods are used to jointly optimize the sparse projection matrix parameters, lightweight adaptation network parameters and parameterized quantum circuit parameters, so as to simultaneously minimize the target recognition error and prediction uncertainty.

[0081] The training process employs quantum natural gradient descent to collaboratively update all trainable parameters: These parameters include the sparse matrix parameters of the matrix projection module. Semantic adapter parameters and quantum circuit parameters At the hardware execution level, the system adopts a heterogeneous computing architecture, establishing an alternating training mechanism between the GPU and the quantum coprocessor: the CNN backbone network and adapter perform gradient calculations on the GPU, while the quantum circuit part completes state evolution and gradient estimation on the quantum coprocessor, and parameter synchronization and updates are achieved across hardware platforms through a parameter server.

[0082] To address the technical challenge of non-differentiability in quantum circuits, a zero-order gradient estimator is introduced as a solution. This estimator samples the probability gradient in the quantum state space through parameter translation, effectively avoiding the dependence of traditional backpropagation on quantum channels, thus enabling end-to-end training of classical-quantum hybrid models. This design retains the advantages of quantum computing while solving the algorithm compatibility issues in hybrid model training.

[0083] In terms of system-level acceleration, extreme inference efficiency is achieved through multi-level optimization: computationally intensive tasks such as matrix projection and adapter inference are deployed on-chip SRAM memory, leveraging near-memory computing characteristics to achieve sub-millisecond forward inference; pulse-level compilation optimization is implemented for quantum circuits, compressing the single quantum state evolution time to the order of 100 microseconds; simultaneously, a hybrid precision computing strategy (FP16+FP32 combination) and dynamic batch processing technology are adopted to dynamically adjust the computation precision and batch size according to input characteristics, ultimately achieving a significant improvement in overall energy efficiency of 3-5 times. These optimization measures together ensure the efficient deployment capability of this method in edge devices, cloud platforms, and hybrid quantum-classical computing environments.

[0084] In one possible implementation, the end-to-end training process on the COCO dataset is as follows:

[0085] First, data preparation: The publicly available COCO 2017 dataset was used as the experimental subject, containing approximately 118k training samples and 5k validation samples; superpixel segmentation and normalization were performed on the input images to generate tensors. ,in =256, =256, =3.

[0086] Second, matrix field compression, initializing the sparse random projection matrix. ,in =4096, with non-zero elements accounting for approximately 3%; the image tensor is vectorized and left-multiplied by the matrix. To obtain low-dimensional embeddings During the training process, for Add structured norm regularization to constrain the sparsity and robustness of the embedding dimension.

[0087] Third, semantic enhancement of the large language model: The multimodal large language model with frozen weights uses LLaVA-Mini (approximately 7B parameters) as the semantic engine; a lightweight adapter is designed. Its parameter count does not exceed 5M, and it is used to map the embedding Z to the word vector space of LLM, outputting semantic vectors. Obtain category-related information from the LLM decoder. Compared to CNN (ResNet-50 backbone) Fusion: .

[0088] Fourth, quantum domain uncertainty estimation, which will affect semantic vectors. Normalize and map to n=8 qubit amplitude states Design of a 12-layer parameterized quantum circuit It includes rotation gates, controlled phase gates, and entanglement gates; it is run through a quantum simulator (Qiskit Aer) or an actual superconducting quantum chip, with 2048 samples; the class distribution is obtained. And calculate the quantum Shannon entropy. like If this is triggered, active learning will be initiated, and the sample will be added to the incremental dataset.

[0089] Fifth, the joint loss function is defined for three-domain joint optimization: in, For cross-entropy loss, , A hybrid training strategy is employed: the CNN and adapter are optimized on the GPU; quantum circuit parameters... Updated on quantum simulators / quantum hardware; end-to-end differentiable training achieved through quantum natural gradient descent and zero-order gradient estimation.

[0090] Experimental results: On the COCO dataset validation set, the method of this invention achieves 47.3% mAP, which is 4.7% higher than the ResNet-50 baseline (42.6% mAP); the average inference energy consumption is 0.38 J / image, which is 62% lower than traditional CNN; the quantum entropy index can distinguish error-prone categories, and active learning on low-confidence samples can further improve generalization performance.

[0091] Compared with the prior art, the embodiments of this application have the following beneficial effects:

[0092] First, high-dimensional image features are significantly compressed using a learnable sparse projection matrix, and structured sparse constraints are applied to remove redundant information. This technique effectively reduces the computational and storage overhead of the model. Experiments have verified that it can reduce inference energy consumption by up to 62%, making it particularly suitable for resource-constrained edge computing devices.

[0093] Second, a large language model with frozen parameters is introduced, and a lightweight adapter is used to extract deep semantic features from images. This transfers the powerful semantic understanding capabilities of the large language model to visual tasks, and through dynamic fusion with visual features, significantly improves the model's classification accuracy and semantic generalization ability in complex scenes.

[0094] Third, the fused semantic features are encoded into quantum states, and the prediction uncertainty is quantified by the evolution and measurement of parameterized quantum circuits. The quantum Shannon entropy uncertainty estimate provided enhances the transparency and interpretability of the model decision-making process.

[0095] Fourth, a three-domain collaborative training and inference framework consisting of matrix projection, semantic adapter, and quantum circuits was designed. This architecture allows the model to flexibly and efficiently schedule and allocate tasks among edge devices, cloud devices, and quantum coprocessors according to task requirements, thereby realizing a high-performance, low-latency, and highly scalable target recognition system in a heterogeneous hardware environment.

[0096] The following provides a possible implementation of an image target recognition device, which performs the various execution steps and corresponding technical effects of the image target recognition method shown in the above embodiments and possible implementations. The device includes:

[0097] The conversion module is used to convert the input image into a high-dimensional visual feature vector, and to reduce the dimensionality of the high-dimensional visual feature vector by using a learnable sparse projection matrix to obtain a low-dimensional embedding representation.

[0098] The fusion module is used to map low-dimensional embedding representations to the embedding space of a large language model with frozen parameters through a lightweight adaptation network, extract semantic enhancement features, and fuse the semantic enhancement features with the visual features output by the visual feature extraction network to generate fused features.

[0099] The encoding module is used to encode the fused features into quantum states, evolve and measure the quantum states through parameterized quantum circuits, obtain the probability distribution of the target category, and calculate the prediction uncertainty based on the probability distribution.

[0100] The collaborative optimization module is used to collaboratively optimize the parameters of the sparse projection matrix, lightweight adaptation network, and parameterized quantum circuit based on the constructed joint loss function, thereby minimizing target recognition error and prediction uncertainty.

[0101] This preferred embodiment provides a computer device that can implement the steps of any embodiment of the image target recognition method provided in this application. Therefore, it can achieve the beneficial effects of the image target recognition method provided in this application. For details, please refer to the previous embodiments, which will not be repeated here.

[0102] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor. Therefore, embodiments of this application provide a storage medium storing multiple instructions that can be loaded by a processor to execute the steps of any embodiment of the image target recognition method provided in this application.

[0103] The storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0104] Since the instructions stored in the storage medium can execute the steps in any of the image target recognition method embodiments provided in this application, the beneficial effects that any of the image target recognition methods provided in this application can achieve can be realized. For details, please refer to the previous embodiments, which will not be repeated here.

[0105] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for image target recognition, characterized in that, The method includes: The input image is converted into a high-dimensional visual feature vector, and the high-dimensional visual feature vector is reduced in dimensionality using a learnable sparse projection matrix to obtain a low-dimensional embedding representation. A lightweight adapting network is used to map low-dimensional embedding representations to the word embedding space of a large language model with frozen parameters, generating semantically enhanced features. Extracting the first category logical value from the decoding end of the large language model Simultaneously, the second category logical value is obtained from the visual feature extraction network. ; First category logical value With second category logical value Perform dynamic gating fusion to generate fused category logical values. ; The dynamic gating fusion method uses a weighted sum of the first-class and second-class logical values ​​through gating weights. The gating weights are determined by the first-class logical value... Second category logical value It was calculated together; The semantic enhancement features are encoded as quantum states, and the quantum states are evolved and measured through parameterized quantum circuits to obtain the probability distribution of the target category. The prediction uncertainty is then calculated based on the probability distribution. Based on the constructed joint loss function, the parameters of the sparse projection matrix, lightweight adaptation network, and parameterized quantum circuit are jointly optimized to minimize target recognition error and prediction uncertainty. The joint loss function is: , For classifying losses, To control the weights of sparse regularization, To control quantum entropy regularity, To control the distribution divergence, For real labels, The merged category logical value, For group sparse regularization terms, For Shannon entropy, Let KL divergence be the KL divergence. This represents the probability distribution of the quantum circuit output. This represents the probability distribution of the output of a traditional CNN path. Based on the constructed joint loss function, the parameters of the sparse projection matrix, lightweight adaptation network, and parameterized quantum circuit are jointly optimized to minimize the target recognition error and prediction uncertainty. The steps include: An alternating training strategy is adopted between GPU and quantum coprocessor. Quantum natural gradient descent and zero-order gradient estimation methods are used to jointly optimize the sparse projection matrix parameters, lightweight adaptation network parameters and parameterized quantum circuit parameters, so as to simultaneously minimize the target recognition error and prediction uncertainty.

2. The image target recognition method as described in claim 1, characterized in that, The steps of converting the input image into a high-dimensional visual feature vector and then using a learnable sparse projection matrix to reduce the dimensionality of the high-dimensional visual feature vector to obtain a low-dimensional embedding representation include: Perform superpixel segmentation or block processing on the input image to form an image tensor. , For the set of real numbers, Image height, Image width, This refers to the number of image channels; Expand the image tensor into a one-dimensional vector. , For tensor flattening operations; Through a learnable sparse random projection matrix For a one-dimensional vector Compression is performed to obtain low-dimensional embeddings. , It is the target dimension after dimensionality reduction.

3. The image target recognition method as described in claim 2, characterized in that, After obtaining the low-dimensional embedding, it also includes: For low-dimensional embedding representation Apply constraints: ,in For group sparse regularization terms, The sparsity threshold is adjustable.

4. The image target recognition method as described in claim 1, characterized in that, The steps of encoding semantically enhanced features into quantum states, evolving and measuring these quantum states using parameterized quantum circuits to obtain the probability distribution of the target category, and calculating the prediction uncertainty based on the probability distribution include: The semantic enhancement features are encoded into quantum states, and the quantum states are evolved through parameterized quantum circuits to obtain the evolved quantum states; The evolved quantum state is repeatedly sampled under a preset measurement basis to obtain the probability distribution of the target category; The quantum Shannon entropy is calculated based on the probability distribution and used as the model's prediction uncertainty for the sample. If the prediction uncertainty exceeds the preset uncertainty threshold, an active learning strategy is triggered, and the sample is added to the next training set.

5. An image target recognition device, characterized in that, The device includes: The conversion module is used to convert the input image into a high-dimensional visual feature vector, and to reduce the dimensionality of the high-dimensional visual feature vector by using a learnable sparse projection matrix to obtain a low-dimensional embedding representation. The fusion module is used to map low-dimensional embedding representations to the word embedding space of a large language model with frozen parameters through a lightweight adaptation network, generating semantically enhanced features. Extracting the first category logical value from the decoding end of the large language model Simultaneously, the second category logical value is obtained from the visual feature extraction network. ; First category logical value With second category logical value Perform dynamic gating fusion to generate fused category logical values. ; The dynamic gating fusion method uses a weighted sum of the first-class and second-class logical values ​​through gating weights. The gating weights are determined by the first-class logical value... Second category logical value It was calculated together; The encoding module is used to encode semantic enhancement features into quantum states, evolve and measure the quantum states through parameterized quantum circuits, obtain the probability distribution of the target category, and calculate the prediction uncertainty based on the probability distribution. The collaborative optimization module is used to collaboratively optimize the parameters of the sparse projection matrix, lightweight adaptation network, and parameterized quantum circuit based on the constructed joint loss function, minimizing the target recognition error and prediction uncertainty. The joint loss function is: , For classifying losses, To control the weights of sparse regularization, To control quantum entropy regularity, To control the distribution divergence, For real labels, The merged category logical value, For group sparse regularization terms, For Shannon entropy, Let KL divergence be the KL divergence. This represents the probability distribution of the quantum circuit output. The probability distribution of the output of the traditional CNN path is used. An alternating training strategy is adopted between the GPU and the quantum coprocessor. Quantum natural gradient descent and zero-order gradient estimation methods are used to jointly optimize the sparse projection matrix parameters, lightweight adaptation network parameters and parameterized quantum circuit parameters, so as to simultaneously minimize the target recognition error and prediction uncertainty.

6. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing a computer program, which is loaded and executed by the processor to implement the image target recognition method as described in any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which is loaded and executed by a processor to implement the image target recognition method as described in any one of claims 1-4.

Images