Deep hash image retrieval method based on hierarchical multi-scale feature fusion

Abstract

The invention provides a deep hash image retrieval method based on hierarchical multi-scale feature fusion. The method comprises the following steps: firstly, adopting a Nested Hierarchical Transform as a backbone feature extraction network, and synchronously capturing global context information and local detail features of an image through a nested local self-attention mechanism and a hierarchical feature aggregation module; then, constructing a multi-scale feature fusion module, extracting visual features of different receptive fields in parallel by using a multi-branch unequal-ratio expansion rate dilated convolution kernel, and adaptively weighting important features in combination with attention to generate discriminative multi-scale fusion features; and finally, designing a mixed loss function, optimizing intra-class feature compactness through center similarity loss, and realizing efficient and semantic similarity keeping Hash code generation in combination with a quantization loss constraint Hash code discretization process. According to the method, full expression of image features is realized through a collaborative mechanism of Transform and multi-scale feature fusion, and the precision and efficiency of large-scale image retrieval are improved through an end-to-end deep hash learning framework.

Description

Technical Field

[0001] The present invention relates to image processing technology, and particularly to a deep hashing image retrieval method for hierarchical multi-scale feature fusion. Technical Background

[0002] The deep hashing image retrieval method is an important research direction in the cross-field of computer vision and information retrieval. Its core idea is to map high-dimensional image data into low-dimensional binary codes through deep hashing technology, significantly improving the retrieval efficiency while maintaining the semantic similarity of images. With the rapid development of the digital age, the scale of image data has shown an explosive growth. Traditional content-based image retrieval methods face problems such as high computational complexity and insufficient retrieval accuracy when dealing with large-scale data. However, deep hashing technology, through feature dimensionality reduction and binary coding, significantly improves the retrieval efficiency and reduces the storage overhead, showing great application potential in the field of image retrieval and having important value in practical scenarios such as medical image retrieval and e-commerce image search.

[0003] However, the existing deep hashing image retrieval methods still have the following key problems: First, the limitation of feature extraction. Although the method based on Convolutional Neural Network (CNN) can effectively extract local features, it is difficult to model the global correlation of image features; while the method based on Vision Transformer (ViT) can capture long-range dependencies, but its ability to extract multi-scale features is insufficient, resulting in the loss of detailed information. Second, insufficient multi-scale feature fusion. Existing methods usually only focus on single-scale or global features and fail to fully explore the complementarity between different-scale features, restricting the discriminative ability of hash codes. Third, high computational complexity. Traditional Transformer models have a large computational overhead when processing high-resolution images and are difficult to meet the requirements of real-time retrieval in practical applications.

[0004] In response to the above problems, in recent years, researchers have proposed various improvement schemes. For example, CSQ uses a CNN architecture combined with quantization loss to optimize hash code generation, DPN uses a pyramid network to extract multi-scale features, and TransHash first introduces ViT into hash learning to enhance the global modeling ability. Although these methods have made certain progress, there are still obvious deficiencies in the fine-grained modeling of multi-scale feature fusion, the balance between computational efficiency and retrieval accuracy, etc. Specifically, the existing methods fail to effectively combine local details and global context information, and lack targeted optimization of multi-level semantics in the feature interaction process, resulting in insufficient discriminability of the generated hash codes. Therefore, designing an efficient, lightweight and deep hashing image retrieval method that can fully fuse multi-scale features has important theoretical significance and application value. Summary of the Invention

[0005] The purpose of the present invention is to solve the problems of insufficient feature extraction, insufficient multi-scale information fusion, and insufficient preservation of global semantic similarity in existing deep hashing image retrieval methods. To achieve the above purpose, the present invention provides a deep hashing image retrieval method based on hierarchical multi-scale feature fusion, which realizes the efficient extraction and fusion of local and global features of images by introducing a hierarchical Transformer structure and a multi-scale dilated convolution fusion network, and combines the optimization of the central similarity quantization loss function to generate binary hash codes with strong discriminability, so as to achieve efficient and accurate image retrieval.

[0006] It mainly includes the following six parts: The first part is to preprocess the image dataset; the second part is to perform hierarchical feature extraction through the Nest Transformer module; the third part is to perform downsampling and multi-scale information fusion on the feature map through the multi-scale dilated convolution network; the fourth part is to input the fused feature map into the hash layer to generate hash codes; the fifth part is to optimize the generation of hash codes based on the central similarity quantization loss function; the sixth part is the training and testing of the deep hashing network model, specifically:

[0007] The first part includes two steps:

[0008] Step 1, download the public image datasets ImageNet, NUS_WIDE, and cifar_10, select N types of images from each dataset as training set samples, and the number of images in each type is not less than 500;

[0009] Step 2, uniformly adjust the training set images to a size of 256×256, randomly crop the images to a size of 224×224 after random flipping and rotation data augmentation operations, and convert them into tensors and then perform normalization processing to form the final training set samples;

[0010] The second part includes three steps:

[0011] Step 3, input the preprocessed training set sample images in Step 2 into the Nest Transformer module, and add position encoding after these feature vectors so that the model can understand the position information of different patches. First, divide the image into multiple non-overlapping image patches, map each image patch to an embedding vector of a fixed dimension through a linear projection layer, and then divide all the embedding vectors into several blocks to form the input;

[0012] Step 4, apply stacked hierarchical Transformer feature extraction to all image patches in parallel. Each Transformer block consists of a multi-head self-attention and a fully connected network, including skip connections and layer normalization, so as to achieve independent local self-attention operations on each image patch and preliminarily extract the local features of the image;

[0013] Step 5: Hierarchically fuse the local features obtained in Step 4 through a block aggregation module. Downsample using 3×3 convolution and max pooling, merge 4 adjacent blocks into larger blocks to form a multi-scale feature pyramid, and output the final multi-scale feature representation. The features at different levels respectively correspond to semantic information at different scales, providing a basis for subsequent multi-scale feature fusion.

[0014] The third part includes one step:

[0015] Step 6: Input the multi-scale features obtained in Step 5 into a multi-scale feature fusion module, which consists of an average pooling module, one 1×1 ordinary convolution, three dilated convolutions with dilation rates of 1, 3, and 5, and a CBAM attention module. First, preliminarily transform and reduce the dimensionality of the input features through a 1×1 ordinary convolution while retaining key information. Then, pass through three dilated convolutions with dilation rates of 1, 3, and 5 and the average pooling module in parallel. The dilated convolutions with different dilation rates capture targets at different scales, enriching semantic information while expanding the receptive field. The average pooling module provides the global context information of the image, fusing the feature results of multi-branch processing to capture more abundant semantic information. Then, through the CBAM (Convolutional Block Attention Module) attention module, sequentially through the channel attention mechanism and the spatial attention mechanism, adaptively weight important features and suppress redundant information, so that the fused multi-scale features are more discriminative, providing a higher-quality feature representation for subsequent hash code generation.

[0016] The fourth part includes two steps:

[0017] Step 7: Perform global average pooling processing on the fused image features along the spatial dimension to obtain a global feature vector.

[0018] Step 8: Through a fully connected layer with a hyperbolic tangent function non-linearity, map the continuous image features to the interval [-1, 1], and then perform a threshold operation to finally convert them into a K-bit binary hash code, realizing the hash coding of the image features.

[0019] The fifth part includes one step:

[0020] Step 9: Input the hash code generated in Step 8 into a hash learning optimization module based on central similarity quantization. By introducing predefined hash centers, explicitly model the global similarity in the Hamming space, use binary cross-entropy to measure the Hamming distance between the hash code and the centers, and introduce a quantization loss based on a bimodal Laplace prior to construct the total optimization objective, thereby enhancing the discriminability of the hash code and better retaining

[0021] Semantic similarity between images;

[0022] The sixth part includes two steps:

[0023] Step 10: Input the training set samples in Step 2 into the network from Step 3 to Step 9, set appropriate network hyperparameters, including learning rate, number of training epochs, optimizer, and the loss function is a combination of central similarity quantization loss and quantization loss. Train the network to obtain the final deep hashing image retrieval model;

[0024] Step 11: Input the test set images into the trained model obtained in Step 10. The model outputs the corresponding binary hash codes, and realizes fast and accurate retrieval of the input images by calculating the Hamming distance between the hash codes.

[0025] The present invention provides a deep hashing image retrieval method based on hierarchical multi-scale feature fusion. First, in terms of feature extraction, to solve the problem of insufficient feature extraction in existing methods, a hierarchical Nest Transformer module is designed. This module constructs a multi-scale feature pyramid through image patch embedding and patch aggregation mechanisms, and adopts a combination of local self-attention and cross-patch feature fusion to efficiently capture local details and global semantics of images. Then, aiming at the problem of insufficient multi-scale information fusion, an improved multi-scale dilated convolution fusion network is designed. By parallelly applying dilated convolutions with different dilation rates and average pooling modules, a feature fusion structure with multi-level receptive fields is constructed, and then the key region features are further highlighted through the attention mechanism, effectively enhancing the feature expression ability of targets at different scales. Finally, to solve the problem of insufficient discriminability of hash codes, a central similarity quantization loss function is proposed. By introducing predefined hash centers to explicitly model the similarity in Hamming space and combining with the bimodal Laplacian quantization constraint, binary hash codes with strong discriminability are generated, significantly improving the accuracy and efficiency of image retrieval. Brief Description of the Drawings

[0026] Figure 1 It is the overall flowchart of the present invention;

[0027] Figure 2 It is the overall framework diagram of the network of the present invention;

[0028] Figure 3 It is the Nest Transformer feature extractor of the present invention;

[0029] Figure 4 It is the multi-scale dilated convolution fusion network of the present invention;

[0030] Figure 5 It is the attention module of the present invention;

[0031] Figure 6This is the retrieval effect diagram of the present invention. Specific embodiments

[0032] To better understand the present invention, the following provides a more detailed description of a hierarchical multi-scale feature fusion-based deep hashing image retrieval method of the present invention in conjunction with specific embodiments. In the following description, the detailed description of the currently existing technology may dilute the subject matter of the present invention, and these descriptions will be ignored here.

[0033] Step 1, the flowchart of the present invention is as Figure 1 shown. First, download the public image datasets ImageNet, NUS_WIDE, and cifar_10, and select N types of images from each dataset as training set samples, with the number of images in each category being no less than 500;

[0034] Step 2, uniformly adjust the training set images to a size of 256×256, randomly crop the images to a size of 224×224 after random flipping and rotation data augmentation operations, and convert them into tensors and then perform normalization processing to form the final training set samples;

[0035] Figure 2 This is the overall network framework diagram of the hierarchical multi-scale feature fusion-based deep hashing image retrieval method of the present invention. In this implementation, the following steps are carried out:

[0036] Step 3, input the preprocessed training set sample images in Step 2 into the Nest Transformer module, and add positional encoding after these feature vectors so that the model can understand the position information of different patches. First, divide the image into non-overlapping image patches of 4×4, and map each image patch to a 128-dimensional embedding vector through a linear projection layer to form an initial feature sequence;

[0037] Step 4, perform hierarchical Transformer feature extraction on the image patches, as Figure 3 shown in the network structure diagram. The specific implementation includes:

[0038] Step 4.1, stack 2 Transformer blocks (composed of multi-head self-attention MSA and multi-layer perceptron MLP, including skip connections and layer normalization) to achieve independent local self-attention operations on each image patch, and initially extract the local features of the image. Perform local self-attention calculation in the 128-dimensional space to extract low-level local features;

[0039] Step 4.2, perform downsampling on the feature map through the block aggregation module in Step 5, merge 4 adjacent blocks into larger blocks, increase the feature dimension to 256 dimensions, and stack 2 Transformer blocks for middle-level feature modeling;

[0040] Step 4.3, perform block aggregation again through Step 5, expand the feature dimension to 512 dimensions, stack 14 Transformer blocks for deep global feature learning, and form a multi-scale feature pyramid.

[0041] Step 5, perform hierarchical feature fusion through the block aggregation module, use 3×3 convolution and max pooling for downsampling, merge 4 adjacent blocks into a larger block to form a multi-scale feature pyramid, and output the final multi-scale feature representation. Among them, features at different levels correspond to semantic information at different scales, providing a basis for subsequent multi-scale feature fusion.

[0042] Step 6, input the multi-scale features output by the Nest Transformer in Step 5 into the multi-scale feature fusion module, which consists of an improved dilated convolution multi-scale feature extraction network and a CBAM attention module:

[0043] Step 6.1, the improved dilated convolution multi-scale feature extraction network has a structure as Figure 4 shown. First, perform preliminary transformation and dimensionality reduction on the input features through a 1×1 ordinary convolution while retaining key information. Then, pass through 3 dilated convolution and average pooling modules with dilation rates of 1, 3, and 5 in parallel. Through hierarchical cascading of 3×3 convolutional kernels with dilation rates of (1, 3, 5), multi-granularity feature capture from local details to global context is achieved, effectively expanding the receptive field while enriching semantic information. At the same time, the non-uniform dilation rate design alleviates the "gridding problem" caused by standard dilated convolution operations. The average pooling module can extract large-scale structures and semantic information in the image, reduce the spatial dimension of the features, and lower the computational complexity, helping the model better understand the overall content of the image. Add a BN+ReLU combination after each convolution branch to solve the problem of gradient explosion. Finally, perform feature fusion on the multi-branch processing results to capture richer semantic information.

[0044] Step 6.2, then pass through the CBAM (Convolutional Block Attention Module) attention module, with a structure as Figure 5 shown. Pass through the channel attention mechanism and the spatial attention mechanism in sequence. First, in the channel attention sub-module, through the dual-path aggregation of global average pooling and max pooling, learn the importance weights between channels. Subsequently, in the spatial attention sub-module, construct a spatial relationship model through a 7×7 convolution, highlight the features of key regions, and suppress redundant information, so that the fused multi-scale features are more discriminative, providing a higher-quality feature representation for subsequent hash code generation.

[0045] Step 7, perform global average pooling on the fused image features in step 6 along the spatial dimension to obtain a global feature vector;

[0046] Step 8, through a fully connected layer with the nonlinearity of the hyperbolic tangent function, map the continuous image features to the interval [-1, 1], and then perform a threshold operation to finally convert them into K-bit binary hash codes to achieve the hash encoding of the image features;

[0047] Step 9, input the hash codes generated in step 8 into the hash learning optimization module based on central similarity quantization, and the specific implementation is as follows:

[0048] Step 9.1, by introducing predefined hash centers, which are pre-computed according to the semantic categories of the image dataset, and each category corresponds to a hash center, which is used to represent the ideal representation of the images of that category in the hash space. By comparing the generated hash codes with these predefined hash centers, the similarity between the hash codes and the corresponding category centers can be measured. Specifically, the binary cross-entropy loss function is used to measure the Hamming distance between the hash codes and the hash centers. The Hamming distance is a commonly used metric to measure the difference between two binary hash codes, representing the number of different bits between the two hash codes. The binary cross-entropy loss function can effectively calculate the similarity between the hash codes and the hash centers, and its calculation formula is as follows:

[0049]

[0050] where K is the length of the hash code, h i,k represents the k-th bit of the hash code of the i-th data point, and c′ i,k represents the k-th bit of the hash center corresponding to the i-th data point;

[0051] Step 9.2, and introduce a quantization loss based on the double-peaked Laplacian prior. The purpose of the quantization loss is to ensure that the generated hash codes are as close as possible to binary values (i.e., 0 or 1), thereby improving the discreteness and discriminability of the hash codes. The calculation formula of the quantization loss is as follows:

[0052]

[0053] where the double-peaked Laplacian prior is a probability distribution that assumes that each bit of the hash code has a high probability density near 0 and 1, while the probability density of the intermediate values is low. This prior distribution can effectively push each bit of the hash code towards 0 or 1, thereby achieving a better quantization effect;

[0054] Step 9.3, construct the total optimization objective, combine the central similarity loss and the quantization loss, so as to enhance the discriminability of the binary hash codes and better retain the semantic similarity between images. The calculation formula of the total optimization objective is as follows:

[0055]

[0056] Where Θ is the set of all parameters learned by the deep hashing function, and λ1 is a hyperparameter used to balance the weights between the global similarity modeling loss and the quantization loss; by minimizing the total optimization objective, a strongly discriminative binary hash code can be generated in the Hamming space while better preserving the semantic similarity between images;

[0057] Step 10: Input the training set samples in Step 2 into the network from Step 3 to Step 9, and set appropriate network hyperparameters: the learning rate is 1×10 -5 , the number of epochs is 100, the batch size is 64, the optimizer is Adam, the loss function is a combination of the central similarity quantization loss and the quantization loss, and train the network to obtain the final pre-trained model for deep hashing image retrieval;

[0058] Step 11: Input the test set images into the trained model obtained in Step 10. The model outputs the corresponding binary hash codes. Using metrics such as the Hamming distance, calculate the similarity between the hash codes of the query image and the database images to achieve efficient similarity measurement; finally, sort the retrieval results according to the similarity and return the list of images most similar to the query image to complete the fast and accurate retrieval task of the images. Figure 6 This is a partial retrieval effect diagram of the present invention.

[0059] The present invention provides a deep hashing image retrieval method with hierarchical multi-scale feature fusion. First, in terms of feature extraction, to solve the problem of insufficient feature extraction in existing methods, a hierarchical Nest Transformer module is designed. This module constructs a multi-scale feature pyramid through image patch embedding and patch aggregation mechanisms, and adopts a combination of local self-attention and cross-patch feature fusion to efficiently capture local details and global semantics of images; then, to address the problem of insufficient multi-scale information fusion, an improved multi-scale dilated convolution fusion network is designed. By parallelly applying dilated convolutions with different dilation rates and average pooling modules, a feature fusion structure with multi-level receptive fields is constructed, and then the key region features are further highlighted through an attention mechanism to effectively enhance the feature expression ability of targets at different scales; finally, to solve the problem of insufficient discriminability of hash codes, a central similarity quantization loss function is proposed. By introducing a predefined hash center to explicitly model the similarity in the Hamming space and combining with a bimodal Laplacian quantization constraint, a strongly discriminative binary hash code is generated, significantly improving the accuracy and efficiency of image retrieval.

[0060] Although the above description of the illustrative embodiments of the present invention has been given, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are within the scope of protection.

Claims

1. A deep hashing image retrieval method for hierarchical multi-scale feature fusion, characterized in that It mainly includes the following six parts: The first part is to preprocess the image dataset; the second part is to perform hierarchical feature extraction through the Nest Transformer module; The third part is to perform downsampling and multi-scale information fusion on the feature map through a multi-scale dilated convolutional network; the fourth part is to input the fused feature map into the hash layer to generate hash codes; the fifth part is to optimize the generation of hash codes based on the central similarity quantization loss function; the sixth part is the training and testing of the deep hash network model, specifically: The first part includes two steps: Step 1, download the public image datasets ImageNet, NUS_WIDE, and cifar_10, and select N images from each dataset as training set samples, with the number of images in each category being no less than 500; Step 2, uniformly adjust the training set images to a size of 256×256, randomly crop the images to a size of 224×224 after random flipping and rotation data augmentation operations, and convert them into tensors and then perform normalization processing to form the final training set samples; The second part includes three steps: Step 3, input the preprocessed training set sample images in Step 2 into the Nest Transformer module, and add positional encoding after these feature vectors so that the model can understand the position information of different patches. First, divide the image into multiple non-overlapping image patches, map each image patch to an embedding vector of a fixed dimension through a linear projection layer, and then divide all the embedding vectors into several blocks to form the input; Step 4, apply stacked hierarchical Transformer feature extraction to all image patches in parallel. Each Transformer block consists of multi-head self-attention and a fully connected network, including skip connections and layer normalization, and initially extracts the local features of the image. The specific implementation includes: (1) Stack 2 Transformer blocks to achieve independent local self-attention operations on each image patch and initially extract the local features of the image. Perform local self-attention calculation in a 128-dimensional space to extract low-level local features; (2) Downsample the feature map through the block aggregation module in Step 5, merge 4 adjacent blocks into a larger block, increase the feature dimension to 256 dimensions, and stack 2 Transformer blocks for middle-level feature modeling; (3) Aggregate the blocks again through Step 5, expand the feature dimension to 512 dimensions, and then stack 14 Transformer blocks for deep global feature learning to form a multi-scale feature pyramid; Step 5, perform hierarchical feature fusion on the local features obtained in Step 4 through the block aggregation module, use 3×3 convolution and max pooling for downsampling, merge 4 adjacent blocks into a larger block to form a multi-scale feature pyramid, and output the final multi-scale feature representation, where features at different levels correspond to semantic information at different scales, providing a basis for subsequent multi-scale feature fusion; The third part includes one step: Step 6: Input the multi-scale features output by the Nest Transformer in Step 5 into the multi-scale feature fusion module, which consists of an improved dilated convolutional multi-scale feature extraction network and a CBAM attention module: (1) Improved dilated convolutional multi-scale feature extraction network: First, perform preliminary transformation and dimensionality reduction on the input features through a 1×1 ordinary convolution while retaining key information. Then, pass through three dilated convolutional and average pooling modules with dilation rates of 1, 3, and 5 in parallel. By hierarchically concatenating 3×3 convolutional kernels with dilation rates of (1, 3, 5), multi-granularity feature capture from local details to global context is achieved, effectively expanding the receptive field while enriching semantic information. At the same time, the non-uniform dilation rate design alleviates the "gridding problem" caused by standard dilated convolution operations. The average pooling module can extract large-scale structures and semantic information in the image, reduce the spatial dimension of the features, and lower the computational complexity, helping the model better understand the overall content of the image. After each convolutional branch, add a BN+ReLU combination to solve the gradient explosion problem and improve the training stability of the model. Finally, perform feature fusion on the multi-branch processing results to capture richer semantic information; (2) Then, pass through the CBAM (Convolutional Block Attention Module) attention module, which sequentially passes through the channel attention mechanism and the spatial attention mechanism. First, in the channel attention sub-module, learn the importance weights between channels through the dual-path aggregation of global average pooling and max pooling. Subsequently, in the spatial attention sub-module, construct a spatial relationship model through a 7×7 convolution to highlight the key region features and suppress redundant information, thereby making the fused multi-scale features more discriminative and providing a higher-quality feature representation for subsequent hash code generation; The fourth part includes two steps: Step 7: Perform global average pooling on the fused image features in Step 6 along the spatial dimension to obtain a global feature vector; Step 8: Through a fully connected layer with a hyperbolic tangent function non-linearity, map the continuous image features to the interval [-1, 1], and then perform a threshold operation to finally convert them into K-bit binary hash codes to achieve the hash encoding of the image features; The fifth part includes one step: Step 9: Input the hash codes generated in Step 8 into the hash learning optimization module based on central similarity quantization, and the specific implementation is as follows: (1) By introducing predefined hash centers, which are pre-computed according to the semantic categories of the image dataset, each category corresponds to a hash center, which is used to represent the ideal representation of the images of that category in the hash space; use the binary cross-entropy loss function to measure the similarity between the generated hash codes and these predefined hash centers; (2) And introduce a quantization loss based on the bimodal Laplacian prior, where the bimodal Laplacian prior is a probability distribution that assumes that each bit of the hash code has a high probability density near 0 and 1, while the probability density of the intermediate value is low; this prior distribution can effectively push each bit of the hash code towards 0 or 1, thereby achieving a better quantization effect, and thus improving the discreteness and discriminability of the hash code; (3) Construct the total optimization objective, combine the central similarity loss and the quantization loss, and by minimizing the total optimization objective, binary hash codes with strong discriminability can be generated in the Hamming space, while better retaining the semantic similarity between images; The sixth part includes two steps: Step 10, input the training set samples in step 2 into the network from step 3 to step 9, set appropriate network hyperparameters, including the learning rate, the number of training epochs, the optimizer, and the loss function is a combination of the central similarity quantization loss and the quantization loss, and train the network to obtain the final deep hash image retrieval model; Step 11, input the test set images into the trained model obtained in step 10, the model outputs the corresponding binary hash codes, use metric methods such as the Hamming distance to calculate the similarity between the hash codes of the query image and the database images, and achieve efficient similarity measurement; finally, sort the retrieval results according to the similarity and return the list of images most similar to the query image to complete the fast and accurate retrieval task of the images.

2. The hierarchical multi-scale feature fusion-based deep hashing image retrieval method according to claim 1, characterized in that In step 4, a hierarchical Transformer module is applied to all image patches in parallel. In the first layer, 2 Transformer blocks are stacked to extract low-level local features; in the second layer, 2 Transformer blocks are stacked for middle-level feature modeling; in the third layer, 14 Transformer blocks are stacked for deep global feature learning.

3. The hierarchical multi-scale feature fusion based deep hashing image retrieval method according to claim 1, wherein, In step 5, an efficient cross-block information transfer and hierarchical feature fusion are achieved through the block aggregation mechanism, where the features at different levels correspond to semantic information at different scales. Each layer merges 4 adjacent blocks into larger blocks to form a multi-scale feature pyramid to achieve global information integration.

4. A hierarchical multi-scale feature fusion-based deep hashing image retrieval method according to claim 1, wherein In step 6, an improved multi-scale dilated convolutional network is used to effectively expand the receptive field while maintaining the semantic richness of the features; mainly using hierarchical cascaded non-uniform dilation rate convolutional kernels to achieve multi-granularity feature capture from local details to global context, and alleviating the "checkerboarding problem" caused by standard dilated convolution operations; then through the attention module, adaptively weight important features, suppress redundant information, and output enhanced multi-scale features.

5. The hierarchical multi-scale feature fusion-based deep hashing image retrieval method according to claim 1, wherein The semantic preservation and quantization enhancement strategy adopted in step 9 can significantly improve the semantic consistency and coding quality of the hash codes, thereby achieving efficient and accurate image retrieval. This strategy endows the network with excellent feature discrimination ability and retrieval stability.

Images