Anaphora image segmentation method based on space-frequency dual tuning

Abstract

The invention relates to the technical field of image segmentation, in particular to an anaphora image segmentation method based on space-frequency dual tuning, and the method comprises the steps: dynamically generating a frequency domain learning prompt and a space domain learning prompt through a space frequency domain dual prompt technology; splicing interaction is carried out by utilizing spatial domain prompts and encoder features, and semantic representation of an anaphora target is enhanced; recovering high-frequency details by adopting a Laplacian pyramid and a LoRA-based frequency domain encoder, and enhancing key features through multiplication operation prompted by a frequency domain; and finally, double-domain optimization features are fused, and an accurate boundary segmentation result is generated through a mask decoder, so that the problems of detail degradation and boundary fuzziness caused by lack of frequency domain information in a traditional method are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of image segmentation technology, specifically to a method for representing image segmentation based on spatial frequency duality optimization. Background Technology

[0002] With the rapid development of multimodal large-scale models, parameter-efficient fine-tuning methods have become the dominant technical approach in the field of image segmentation due to their ability to fully utilize the general prior knowledge in pre-trained models while significantly reducing computational and storage costs. Current mainstream method paradigms can be summarized into four categories: The first is full fine-tuning, which updates all parameters of the entire multimodal model during the fine-tuning process. While achieving significant results, it incurs huge computational overhead. The second is parameter-efficient independent encoder fine-tuning, which uses techniques such as adapters, side networks, or cue-based tuning to perform independent lightweight tuning of the visual encoder and text encoder, respectively. The third is parameter-efficient interactive encoder fine-tuning, which focuses on efficiently updating the parameters of interaction modules between visual and text features (such as cross-attention mechanisms) to enhance intermodal fusion. The fourth is parameter-efficient segmentation token decoding, which directly generates a special segmentation token, similar to a language token, to drive the mask decoder, achieving pixel-level segmentation.

[0003] While the aforementioned paradigms based on efficient parameter fine-tuning have achieved remarkable results, they all share a fundamental flaw: all these methods are entirely confined to feature learning and optimization in the spatial domain. Spatial domain processing primarily focuses on pixel-level visual information such as color, texture, and shape. However, during the layer-by-layer propagation of deep networks, high-frequency details (such as object edges, sharp corners, and subtle texture patterns) are easily smoothed and degraded, leading to a gradual "blurring" of feature maps. More critically, these methods completely ignore the rich discriminative information contained in the frequency domain. In the frequency domain, high-frequency components precisely carry the fine details of object boundaries, while low-frequency components preserve overall structural information and background context. This neglect of frequency domain information prevents the model from utilizing the potential and powerful complementarity and synergy between the spatial and frequency domains (e.g., the spatial domain provides semantic context, and the frequency domain provides contour details), thus limiting the model's ability to deeply understand the referent and ultimately resulting in coarse or imprecise object boundaries at the pixel-level segmentation stage.

[0004] Therefore, developing a new method that can synergistically utilize spatial and frequency domain information has become a key technical challenge for improving the performance of referential image segmentation. Summary of the Invention

[0005] The purpose of this invention is to provide a method for indexical image segmentation based on space-frequency duality tuning, in order to solve the problems mentioned in the background section. Specifically, the technical problem includes how to design a lightweight tuning method that can effectively mine and fuse frequency domain information and space-frequency interaction cues, thereby addressing the issues of high-frequency detail degradation and inaccurate segmentation boundaries caused by existing indexical image segmentation techniques being limited to the spatial domain.

[0006] To achieve the above objectives, the present invention aims to solve the problems of high-frequency degradation and inaccurate segmentation boundaries in the denotation image segmentation method based on space-frequency duality tuning, specifically including the following method steps: S1. Acquire the input image and dynamically generate two types of learnable cues, including frequency domain cues and spatial domain cues, using spatial-frequency domain dual cueing techniques. The specific process for dynamically generating these two types of learnable cues includes: The system receives an input image and a textual representation expression, where the input image is processed by a visual encoder to extract spatial feature maps, and the textual representation expression is processed by a semantic encoder to generate a text embedding vector. Dynamically generate frequency domain cueing and spatial domain cueing using spatial-frequency domain dual cueing techniques; The frequency domain cue is integrated from three components through a concatenation operation, including naive frequency domain cue, spatially conditionally guided frequency domain cue, and semantically guided frequency domain cue. The three components of the frequency domain cue are generated through corresponding cue control networks and cue guidance networks. The naive frequency domain cue consists of a learnable parameter matrix; the spatially conditionally guided frequency domain cue is generated by processing spatial feature maps through a spatial-to-frequency domain cue control network; and the semantically guided frequency domain cue is generated by processing text embedding vectors through a semantic-to-frequency domain cue guidance network. The spatial domain cue is integrated from three components through a concatenation operation, including a naive spatial domain cue, a frequency-conditional guided spatial domain cue, and a semantically guided spatial domain cue. The three components of the spatial domain cue are generated through corresponding cue control networks and cue guidance networks. The naive spatial domain cue is composed of a learnable parameter matrix; the frequency-conditional guided spatial domain cue is generated by processing frequency domain features through a frequency-to-spatial domain cue control network; and the semantically guided spatial domain cue is generated by processing text embedding vectors through a semantic-to-spatial domain cue guidance network.

[0007] Step S1 dynamically generates two types of learnable cues using spatial-frequency dual cueing technology. Its effect is to utilize visual and semantic foundations to construct frequency-domain cues that integrate spatial conditions and semantic guidance, as well as spatial-domain cues that integrate frequency-domain conditions and semantic guidance. This effectively uncovers spatial-frequency interaction cues and provides accurate guidance information for subsequent feature optimization.

[0008] S2. Input the input image into the spatial domain encoder to extract basic spatial domain features, and interact with the spatial domain cue. Token overlay is achieved through a concatenation operation to generate optimized spatial domain features, specifically including: The input image is fed into a pre-trained spatial domain encoder to extract basic spatial domain feature maps. The basic spatial domain feature map interacts with the spatial domain cue. By concatenating the spatial domain cue as a learnable cue token with the token sequence of the basic spatial domain feature map, an extended token sequence is formed. The extended token sequence is forward-propagated through subsequent layers of the encoder, exchanges information through self-attention computation, and finally outputs optimized spatial domain features.

[0009] Step S2 involves concatenating spatial domain cues with basic spatial domain features. This process enables the cue token and image token to exchange information and perform attention calculations within the encoder, thereby adaptively modulating and optimizing the original spatial features, enhancing the semantic features of the referent, and mitigating the feature degradation problem caused by spatial domain limitations.

[0010] S3. Perform a Laplacian transform on the input image to generate a multi-scale Laplacian pyramid, and process the multi-scale Laplacian pyramid using a LoRA-based high-frequency auxiliary enhancement module. The generation process of the multi-scale Laplacian pyramid specifically includes: A Gaussian pyramid is constructed, with the input image at the bottom layer. By continuously applying Gaussian blur and downsampling operations, images with resolution halved layer by layer are generated. Each layer of the multi-scale Laplacian pyramid is generated by calculating the difference between the image of each Gaussian pyramid layer and the image of the layer above it after upsampling.

[0011] By inputting a multi-scale Laplacian pyramid into a frozen frequency domain encoder, high-frequency details are recovered using a parameter-efficient fine-tuning technique. The frequency domain cue is then applied to the frequency domain features using a multiplication operation, resulting in a compensated frequency domain feature output. The specific process of recovering high-frequency details using the parameter-efficient fine-tuning technique includes: The multi-scale Laplacian pyramid is fed into the frozen frequency domain encoder; the LoRA technique is used to fine-tune the frozen encoder, and the original weights are kept unchanged by introducing a learnable low-rank matrix, and the initial frequency domain features are output.

[0012] The specific process by which the multiplication operation applies frequency domain hints to frequency domain features includes: Tensor operations are performed between the frequency domain cue and the initial frequency domain features to generate a weighted graph. After the weighted graph is activated by an activation function, it is multiplied element-wise with the initial frequency domain features to output the compensated frequency domain features.

[0013] Step S3 constructs a multi-scale Laplacian pyramid and uses a LoRA-based frequency domain encoder for high-frequency enhancement. Its effect is to extract and restore high-frequency detail information of the image from multiple scales, and further enhance key edges and texture features through the multiplication selection mechanism of frequency domain cueing, which significantly improves the problem of blurred segmentation boundaries caused by loss of high-frequency information.

[0014] S4. The optimized spatial domain features are fused with the compensated frequency domain features. The fusion process specifically includes: The optimized spatial domain features and compensated frequency domain features are aligned using a feature alignment operation. The matching of the two feature dimensions is ensured through size adjustment and channel projection. The feature tensor is then fused using a channel dimension splicing strategy to form a fused feature tensor.

[0015] The fused feature tensor is input into the mask decoder using a feature integration strategy to generate the target segmentation mask. Specifically, this includes: The fused feature tensor is input into the mask decoder, which is built on a lightweight architecture. Inside the mask decoder, the fused features are reshaped into a token sequence and cross-attention is calculated with the text embedding features; After processing by the network layers, the image is gradually upsampled and transformed to the original input image resolution; By compressing the channels of the prediction head and mapping them through the activation function, a pixel-level probability map is obtained, which is then binarized to generate the final target segmentation mask.

[0016] Step S4 integrates the optimized spatial domain features with the compensated frequency domain features. Its effect is to complement and integrate the semantically rich spatial features with the detailed frequency domain features. Through cross-modal interaction and upsampling reconstruction in the mask decoder, a clear and accurate target segmentation mask is finally generated, realizing the efficient fusion and synergistic enhancement of spatial and frequency information under lightweight optimization.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention dynamically generates learnable cues that combine semantic guidance and inter-domain conditional constraints through spatial-frequency dual cueing technology, effectively mining and integrating spatial-frequency interaction cues. In particular, it employs LoRA technology to efficiently fine-tune the parameters of the frozen frequency-domain encoder, and by introducing a trainable low-rank matrix, it accurately recovers high-frequency details while maintaining the integrity of pre-trained knowledge, significantly improving the edge blurring problem caused by spatial domain limitations. Finally, through dual-domain feature fusion and lightweight optimization strategies, it achieves more accurate and detailed segmentation of the target boundary while significantly reducing computational overhead. Attached Figure Description

[0018] Figure 1This is a schematic diagram of the overall method steps of the present invention; Figure 2 This is the core flowchart of step S3 of the present invention. Detailed Implementation

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Next, please refer to Figure 1 The purpose of this embodiment is to provide a method for representing image segmentation based on space-frequency dual tuning, which includes the following steps: S1. Acquire the input image and dynamically generate two types of learnable cues, including frequency domain cues and spatial domain cues, using spatial-frequency domain dual cueing techniques. The specific process includes: The input image and textual representation expression are received, where the spatial feature map of the input image is extracted by a visual encoder (denoted as...). The text reference expression generates a text embedding vector (denoted as ) through a semantic encoder. This provides a visual and semantic foundation for dual-domain prompts; Two types of learnable cues, frequency domain cues and spatial domain cues, are dynamically generated using spatial-frequency domain dual cueing techniques, wherein: Frequency domain cues are used to capture frequency domain features (such as high-frequency details and low-frequency structure). A frequency domain cue consists of three components, which are combined through a concatenation operation to form naive frequency domain cues, spatially guided frequency domain cues, and semantically guided frequency domain cues. The `concat` function is a concatenation function used to concatenate naive frequency domain cues, spatially guided frequency domain cues, and semantically guided frequency domain cues. The vector representation of the naive frequency domain cues is as follows: ;in This represents the basic unit of the prompt vector. This represents the 1st, 2nd, and so on up to the Nfpth naive frequency domain hint token; This represents the set of real numbers, indicating that the prompt consists of real numbers; This represents the channel dimension of each naive frequency domain cue token, for example, in the Transformer model, the dimension of the corresponding feature vector (such as 512-dimensional or 768-dimensional). This indicates the number of naive frequency domain cue tokens, defining the length of the cue vector or the number of tokens. Let represent a matrix with Nfp rows and C columns, which is a trainable parameter matrix with shape [Nfp, C]. The vector representation of the frequency domain cue guided by spatial conditions is as follows: ;in This represents the basic unit of the prompt vector. Indicates the 1st, 2nd, and so on up to the 1st. A spatial condition-guided frequency domain prompt token; It is a lightweight spatial-to-frequency domain cue control network; The vector representation of semantically guided frequency domain cues is as follows: To indicate the quantity, among which This represents the basic unit of the prompt vector. Indicates the 1st, 2nd, and so on up to the 1st. A semantically guided frequency domain prompt token; It is a semantic-to-frequency domain cueing and guidance network.

[0021] Spatial domain cues are also composed of three types of components: naive spatial domain cues, frequency-domain condition-guided spatial domain cues, and semantically guided spatial domain cues, namely: The vector representation hinted at by the naive space domain is: ,in This represents the basic unit of the prompt vector. Indicates the 1st, 2nd, and so on up to the 1st. A simple spatial domain prompt token; This represents the dimension of each naive space domain hint token; The vector representation of frequency domain conditional guidance spatial domain hints is as follows: ,in This represents the basic unit of the prompt vector. Indicates the 1st, 2nd, and so on up to the 1st. A frequency domain conditional guidance spatial domain prompt token; It is a lightweight frequency-domain to spatial-domain cue control network; The vector representation of semantically guided spatial domain hints is as follows: ,in This represents the basic unit of the prompt vector. Indicates the 1st, 2nd, and so on up to the 1st. A semantic guidance space domain prompt token; It is a semantic-to-spatial domain prompting and guidance network.

[0022] S2. Input the input image into the spatial domain encoder to extract basic spatial domain features, and interact with spatial domain prompts. Token overlay is achieved through a concatenation operation to generate optimized spatial domain features, specifically including: The input image is fed into a pre-trained spatial domain encoder, which extracts visual features from the image layer by layer, and finally outputs a basic spatial domain feature map tensor, denoted as . The spatial domain feature map tensor carries the semantic and structural information of the image in the spatial domain. Spatial domain feature map tensor Spatial Domain Hints The interaction is performed through a splicing operation, the process of which is as follows: In the Transformer architecture, the spatial domain feature map tensor A sequence of tokens represented as a series of image patches; Spatial domain hints It is viewed as a series of learnable cue tokens, directly related to the spatial domain feature map tensor. The token sequences are concatenated along the sequence length dimension to form a new, extended token sequence; this is equivalent to injecting additional, controllable guiding information into the model. This information integrates cues from the frequency domain and the semantic domain, thereby guiding the spatial domain encoder to pay more attention to the target region features related to text reference in subsequent self-attention and other computational processes. After the above splicing operation, the extended token sequence will continue to propagate forward through the subsequent Transformer layers in the spatial domain encoder. In these layers, the cue token and the original image token will exchange information and perform attention calculations, ultimately adaptively modulating and optimizing the original spatial domain features. Finally, the features obtained from the encoder output are the spatial domain features after spatial domain cue tuning, i.e., the tuned spatial domain features, denoted as... This feature not only preserves the original spatial information but also enhances the semantic features related to the referent, laying a solid foundation for subsequent fusion with frequency domain features and accurate segmentation.

[0023] S3, please refer to Figure 2 The Laplacian transform is performed on the input image to generate a multi-scale Laplacian pyramid, specifically including: First, construct a Gaussian pyramid, starting from the bottom layer (layer 0). Given an input image, a series of images with progressively halved resolution are generated by continuously applying Gaussian blur and downsampling operations. , ,..., ;in Represents the downsampling function; The first of the Gaussian pyramids layer; Then, each layer of the multi-scale Laplacian pyramid is generated by calculating the difference between each layer of the Gaussian pyramid image and the upsampled image of the layer above it. The specific calculation formula is as follows: ,in Indicates the upsampling function; The first representing the Pyramid of Laplace Layers, each layer Both contain high-frequency detail information (edges, textures, etc.) of the input image at a specific scale, as well as some residual information; the larger the scale ( The smaller the size, the more refined the details; the smaller the scale ( The larger the value, the coarser the details.

[0024] The multi-scale Laplacian pyramid is processed using a LoRA (low-rank adaptive)-based high-frequency auxiliary enhancement module. By inputting the multi-scale Laplacian pyramid into a frozen frequency domain encoder, high-frequency details are recovered using efficient parameter fine-tuning techniques. Specifically, this includes: Images of each layer of the multi-scale Laplacian pyramid As input, a frozen frequency domain encoder is fed in. This frequency domain encoder has the same pre-trained weight architecture as the spatial domain encoder in step S2, but its weights are frozen in this process, i.e. they do not participate in gradient updates, so as to make full use of its prior knowledge of extracting general features and ensure parameter efficiency. To efficiently fine-tune the parameters of the frozen frequency domain encoder and adapt it for high-frequency detail extraction tasks, this invention employs LoRA (Low-Rank Adaptation) technology, specifically a LoRA-based high-frequency auxiliary enhancement module. In particular, learnable low-rank matrices B and A are introduced into the linear layer within the frequency domain encoder. For an original frozen weight... The frozen layer, its forward propagation process is modified as follows: ,in ,in The first frequency domain block Layer representation; The first frequency domain block Layer representation; Indicates that the weights are frozen; This represents the result of introducing a learnable low-rank matrix multiplication; achieved through training only. and And maintain This significantly reduces the number of trainable parameters while effectively guiding the frozen encoder to focus on recovering and extracting key high-frequency detail features from the Laplacian input, the output of which is considered as preliminary frequency domain features.

[0025] Frequency domain hints dynamically generated in step S1 The interaction is applied to the preliminary frequency domain features obtained in the previous step, specifically through a multiplication operation, as follows: First, frequency domain hints A weighted graph is generated by performing tensor operations with the frequency domain features. After the weighted graph is activated by the Sigmoid function, it is multiplied element-wise with the initial frequency domain features. This operation is equivalent to an attention mechanism or gating mechanism, which allows frequency domain cues to come from different parts of adaptively weighted frequency domain features, filters and enhances frequency domain information that is beneficial to the segmentation of the referent target, and suppresses irrelevant or noisy information.

[0026] After the feature extraction and frequency domain cue multiplication selection operations fine-tuned by the LoRA technique described above, the final output is the compensated frequency domain feature, denoted as... This feature effectively incorporates fine edge and texture information captured by a multi-scale Laplacian pyramid and adaptively selected by the model.

[0027] S4. The optimized spatial domain features and the compensated frequency domain features are fused, and the fusion strategy is used to input the feature integration strategy into the mask decoder to generate the final target segmentation mask. Specifically, this includes: The optimized spatial domain features output from step S2 Compensated frequency domain characteristics output from step S3 Before fusion, feature alignment is required to ensure that the two feature tensors are dimensionally matched in terms of spatial resolution and number of channels. Spatial size is adjusted by upsampling or downsampling, and the number of channels is adjusted by linear projection such as 1x1 convolution. The fusion strategy employed is channel-dimensional stitching, which involves combining optimized spatial domain features. and compensated frequency domain characteristics The concatenation is performed along the channel dimension C to form a fused feature tensor. This operation preserves all information from the spatial and frequency domains to the greatest extent possible, providing the decoder with a composite feature representation that combines global semantic understanding and local detail resolution.

[0028] fusion features after splicing The input will be fed into a mask decoder, which is built on a lightweight Transformer architecture, for example, consisting of 3 Transformer layers (each containing 8 attention heads) and a feedforward network (512 dimensions); The feature integration strategy is reflected in the internal processing flow of the mask decoder, and includes: Serialization: Feature Fusion First, it is reshaped into a sequence of tokens; Interaction and Enhancement: The token sequence undergoes cross-attention computation with the text embedding features in the Transformer layer, which allows the visual features to be further focused and adjusted according to the semantics of the text, ensuring that the decoding process closely revolves around the target referred to by the language; Upsampling and reconstruction: After processing by the Transformer layer, the image is passed through the feedforward network of the mask decoder and gradually upsampled and transformed, finally mapped to a spatial size with the same resolution as the original input image; The output of the mask decoder is a feature map with the same spatial size as the input image. This feature map is passed through a prediction head, such as a convolutional layer, to compress the number of channels to 1. Finally, a sigmoid activation function is used to map each pixel value of the single-channel feature map to a range of 0 to 1, representing the probability that the pixel belongs to the target. The final output generated is a binarized target segmentation mask. Typically, a threshold (such as 0.5) is set, and pixels with a probability value greater than the threshold are identified as target objects (value 1), while the remaining pixels are identified as background (value 0), thus obtaining clear and accurate segmentation boundaries.

[0029] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for segmenting an image based on space-frequency duality tuning, characterized in that, The method comprises the following steps: S1, obtaining an input image, and dynamically generating two types of learnable prompts through a spatial frequency domain dual prompt technology, including a frequency domain prompt and a spatial domain prompt; S2, inputting the input image into a spatial domain encoder to extract basic spatial domain features, and interacting with the spatial domain prompt to realize token superposition through a splicing operation to generate optimized spatial domain features; S3, performing Laplace transform on the input image to generate a multi-scale Laplace pyramid, processing the multi-scale Laplace pyramid through a high-frequency auxiliary enhancement module based on LoRA, inputting the multi-scale Laplace pyramid into a frozen frequency domain encoder, restoring high-frequency details by using a parameter efficient fine-tuning technology, and outputting compensated frequency domain features by using a multiplication operation to apply the frequency domain prompt to the frequency domain features; S4, fusing the optimized spatial domain features and the compensated frequency domain features, and inputting a mask decoder through a feature integration strategy to generate a target segmentation mask.

2. The image segmentation method based on space-frequency duality tuning according to claim 1, wherein, The dynamic generation process of the two types of learnable prompts comprises: receiving an input image and a text representative expression, wherein the input image extracts a spatial feature map through a visual encoder, and the text representative expression generates a text embedding vector through a semantic encoder; dynamically generating a frequency domain prompt and a spatial domain prompt through a spatial frequency domain dual prompt technology; the frequency domain prompt is integrated by three components through a splicing operation, including a naive frequency domain prompt, a spatial condition guided frequency domain prompt, and a semantic guided frequency domain prompt; the spatial domain prompt is integrated by three components through a splicing operation, including a naive spatial domain prompt, a frequency domain condition guided spatial domain prompt, and a semantic guided spatial domain prompt.

3. The image segmentation method based on space-frequency duality tuning of claim 2, wherein, The three components of the frequency domain prompt are generated through corresponding prompt control networks and prompt guide networks, specifically including: the naive frequency domain prompt is composed of a learnable parameter matrix; the spatial condition guided frequency domain prompt is generated by processing the spatial feature map through a spatial domain to frequency domain prompt control network; the semantic guided frequency domain prompt is generated by processing the text embedding vector through a semantic to frequency domain prompt guide network.

4. The method of claim 2, wherein the method is based on space-frequency duality optimization. The three components of the spatial domain prompt are generated through corresponding prompt control networks and prompt guide networks, specifically including: the naive spatial domain prompt is composed of a learnable parameter matrix; the frequency domain condition guided spatial domain prompt is generated by processing the frequency domain features through a frequency domain to spatial domain prompt control network; the semantic guided spatial domain prompt is generated by processing the text embedding vector through a semantic to spatial domain prompt guide network.

5. The method of image segmentation by space-frequency duality tuning according to claim 1, wherein, The generation process of the optimized spatial domain features comprises: inputting the input image into a pre-trained spatial domain encoder to extract a basic spatial domain feature map; interacting the basic spatial domain feature map with the spatial domain prompt by splicing the spatial domain prompt as a learnable prompt token with the token sequence of the basic spatial domain feature map to form an extended token sequence; the extended token sequence is forward propagated through the subsequent layers of the encoder, exchanges information through self-attention calculation, and finally outputs the optimized spatial domain features.

6. The method of image segmentation by space-frequency duality tuning according to claim 1, wherein, The generation process of the multi-scale Laplace pyramid comprises: A Gaussian pyramid is constructed, with the bottom layer being the input image, and images with halved resolution are generated by continuously applying Gaussian blur and downsampling operations; Each layer of the multi-scale Laplacian pyramid is generated by calculating the difference between the current layer of the Gaussian pyramid image and the upsampled image of the previous layer.

7. The method of image segmentation by space-frequency duality tuning according to claim 1, wherein, The specific process of the parameter efficient fine-tuning technique to restore high-frequency details includes: The multi-scale Laplacian pyramid is input into the frozen frequency domain encoder. The LoRA technique is used to fine-tune the frozen encoder, which introduces a learnable low-rank matrix while keeping the original weights unchanged, and outputs the preliminary frequency domain features.

8. The method of image segmentation by space-frequency duality tuning according to claim 1, wherein, The specific process of the multiplication operation to apply the frequency domain prompt to the frequency domain features includes: The frequency domain prompt and the preliminary frequency domain features are subjected to tensor operation to generate a weight map. After the weight map is activated by an activation function, it is multiplied element by element with the preliminary frequency domain features to output the compensated frequency domain features.

9. The method of image segmentation by space-frequency duality tuning according to claim 1, wherein, The fusion process in step S4 specifically includes: The fine-tuned spatial domain features and the compensated frequency domain features are subjected to feature alignment operations to ensure that the dimensions of the two features match through size adjustment and channel projection. The fusion is performed using a channel dimension splicing strategy to form a fused feature tensor.

10. The method of image segmentation by space-frequency duality tuning according to claim 1, wherein, The generation process of the target segmentation mask specifically includes: The fused feature tensor is input into the mask decoder, which is built based on a lightweight architecture; Inside the mask decoder, the fused features are reshaped into a token sequence and cross-attention calculation is performed with the text embedding features; After network layer processing, the resolution is gradually upsampled to the original input image resolution; Through the prediction head compression channel and the activation function mapping, a pixel-level probability map is obtained, and the final target segmentation mask is generated after binarization.

Images