Furniture main body multi-target identification method, device and equipment and storage medium

Abstract

The invention relates to the technical field of furniture image recognition, in particular to a furniture main body multi-target recognition method and device, equipment and a storage medium. Pre-training a preset convolutional neural network model based on a preset furniture picture annotation library and a preset instance segmentation model to obtain a deep convolutional neural network segmentation model; optimizing the deep convolutional neural network segmentation model based on a preset model tuning parameter to obtain a furniture image main body identification model; acquiring a furniture image data set; performing similarity retrieval on the furniture image data set based on the furniture image main body identification model and a preset vector database to obtain a furniture main body multi-target identification result; by fusing the advantages of the instance segmentation model, the furniture image recognition efficiency and reasoning speed are improved through pre-training and tuning, multi-scene application is supported, and digital transformation of the home industry is assisted.

Description

Technical Field

[0001] This invention relates to the field of furniture image recognition technology, specifically to a method, apparatus, device, and storage medium for multi-target recognition of furniture subjects. Background Technology

[0002] With the rapid development of smart homes and digital interior design, the demand for furniture recognition technology in areas such as intelligent space planning, virtual home decoration design, and home storage management is increasing. However, existing image recognition technologies face multiple technical challenges when processing furniture scenes: First, complex textures (such as wallpaper and carpet patterns), lighting variations, and clutter in the indoor environment can easily dilute the information of the main furniture subject due to background interference, leading to feature extraction bias and target misjudgment. Second, in complex scenarios where multiple pieces of furniture coexist in living rooms, bedrooms, etc., traditional algorithms struggle to accurately identify multiple core targets such as sofas, tables, chairs, and wardrobes simultaneously, resulting in issues like target overlap and missed detections, category confusion, and an inability to meet the refined furniture layout analysis requirements of intelligent systems. Third, facing massive furniture image databases, vector retrieval technology suffers from the curse of dimensionality and excessively high feature similarity, leading to a significant decrease in retrieval efficiency and accuracy, making it difficult to achieve rapid matching and intelligent recommendation of furniture styles. Therefore, there is an urgent need to develop a multi-target furniture recognition method that can resist background interference, accurately identify multiple targets, and adapt to large-scale data retrieval, in order to promote the intelligent upgrade of the home furnishing industry. Summary of the Invention

[0003] To address the shortcomings of the existing technology, this invention proposes a multi-target recognition method for furniture components.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for multi-target recognition of furniture subjects includes: pre-training a pre-defined convolutional neural network model based on a pre-defined furniture image annotation library and a pre-defined instance segmentation model to obtain a deep convolutional neural network segmentation model; optimizing the deep convolutional neural network segmentation model based on pre-defined model tuning parameters to obtain a furniture image subject recognition model; acquiring a furniture image dataset; and performing similarity retrieval on the furniture image dataset based on the furniture image subject recognition model and a pre-defined vector database to obtain the furniture subject multi-target recognition result.

[0005] Furthermore, the step of performing similarity retrieval on the furniture image dataset based on the furniture image subject recognition model and a preset vector database to obtain multi-target furniture subject recognition results includes: processing the furniture image dataset based on the furniture image subject recognition model and a preset confidence threshold to obtain a high-confidence dataset; generating a combined mask set based on the high-confidence dataset and a preset value range threshold interval; performing a broadcast operation on the high-confidence dataset based on a preset binarization threshold and the combined mask set to obtain segmentation mask region information; performing full feature extraction on the segmentation mask region information based on a preset residual neural network model to obtain a vector feature set; obtaining a classification label set based on the high-confidence dataset; and performing similarity retrieval on the vector feature set and the classification label set based on the vector database to obtain multi-target furniture subject recognition results.

[0006] Furthermore, the step of processing the furniture image dataset based on the furniture image subject recognition model and a preset confidence threshold to obtain a high-confidence dataset includes: setting the size and configuring the channels of the furniture image dataset to obtain a configured furniture image dataset; performing subject segmentation on the configured furniture image dataset based on the furniture image subject recognition model to obtain a furniture subject image dataset; analyzing the furniture subject image dataset based on the confidence threshold to obtain a confidence analysis result; and filtering the furniture subject image dataset according to the confidence analysis result to obtain a high-confidence dataset.

[0007] Further, the step of generating a combined mask set based on the high-confidence dataset and a preset value range threshold interval includes: obtaining a segmentation mask set based on the high-confidence dataset; performing value range analysis on the segmentation mask set according to the value range threshold interval to obtain the value range analysis result; when the value range analysis result indicates that there are segmentation masks whose value range is within the value range threshold interval, then obtaining all segmentation masks whose value range is within the value range threshold interval from the segmentation mask set to obtain a mask set to be processed; performing normalization processing on each mask to be processed in the mask set to be processed to obtain a normalized dataset; obtaining the remaining dataset from the segmentation mask set based on the normalized dataset; and generating a combined mask set based on the normalized dataset and the remaining dataset.

[0008] Furthermore, the step of broadcasting the high-confidence dataset according to a preset binarization threshold and a combined mask set to obtain segmentation mask region information includes: binarizing each mask in the combined mask set according to the binarization threshold to obtain a binarized mask set; performing three-dimensional expansion on the binarized mask set to obtain a three-dimensional mask set; and broadcasting the high-confidence dataset according to the three-dimensional mask set to obtain segmentation mask region information.

[0009] Furthermore, the step of extracting full features from the segmented mask region information based on a preset residual neural network model to obtain a vector feature set includes: obtaining the fully connected layer of the residual neural network model; performing a layer-by-layer removal operation on the fully connected layer according to a preset classification layer category to obtain an updated connected layer; configuring the residual neural network model according to the updated connected layer to obtain a configured residual neural network model; pre-training the configured residual neural network model according to a preset image database to obtain a pre-trained residual neural network model; and extracting full features from the segmented mask region information based on the pre-trained residual neural network to obtain a vector feature set.

[0010] Furthermore, the step of performing similarity retrieval on the vector feature set and classification label set based on a vector database to obtain the multi-target recognition result of the furniture subject includes: normalizing the vector feature set based on a preset vector normalization algorithm to obtain a normalized feature set; classifying the normalized feature set according to the classification label set to obtain a classification feature set; partitioning the classification feature set based on the vector database to obtain a partitioned feature set; converting the classification label set according to the encoding method and a preset vector space to obtain a label feature set; jointly processing the label feature set and the partitioned feature set to obtain a joint vector feature set; performing similarity retrieval on the joint vector feature set to obtain a similarity feature set; reordering the similarity feature set according to preset label logic rules to obtain a reordered feature set; and generating the multi-target recognition result of the furniture subject based on the reordered feature set.

[0011] Furthermore, a multi-target furniture subject recognition device includes: a model pre-training module for pre-training a preset convolutional neural network model based on a preset furniture image annotation library and a preset instance segmentation model to obtain a deep convolutional neural network segmentation model; a model optimization module for optimizing the deep convolutional neural network segmentation model based on preset model tuning parameters to obtain a furniture image subject recognition model; a data acquisition module for acquiring a furniture image dataset; and a target subject recognition module for performing similarity retrieval on the furniture image dataset based on the furniture image subject recognition model and a preset vector database to obtain multi-target furniture subject recognition results.

[0012] Furthermore, a furniture body multi-target recognition device includes: a memory and at least one processor, wherein the memory stores instructions; at least one processor invokes the instructions in the memory to cause the furniture body multi-target recognition device to perform the various steps of a furniture body multi-target recognition method as described in any one of the above descriptions.

[0013] Furthermore, a computer-readable storage medium stores instructions that, when executed by a processor, implement the steps of a multi-target furniture subject recognition method as described in any one of the above descriptions.

[0014] The beneficial effects of the multi-target recognition method for furniture main body of the present invention are as follows: By leveraging the architectural advantages of instance segmentation models, object detection and pixel-level segmentation tasks are deeply integrated. Pre-training with a furniture image annotation library enables the pre-trained deep convolutional neural network segmentation model to simultaneously detect target locations and generate segmentation masks, significantly improving target image recognition efficiency. Model optimization based on parameter tuning ensures the model continuously converges to a low-loss state during training. This refined training strategy not only enhances the furniture image subject recognition model's sensitivity to details such as furniture outlines and materials but also reduces computational overhead, improving the model's inference speed. The trained furniture image subject recognition model can quickly locate target products from massive images, supporting applications such as similar style recommendations and style classification. Furthermore, based on a general architecture, the furniture image subject recognition model supports transfer learning, adapting to different furniture images and even other home furnishing fields. It also supports multi-format export, facilitating deployment on various devices, providing technical support for the automation upgrade of furniture production, sales, and design, reducing labor costs, and driving the digital transformation of the home furnishing industry. Attached Figure Description

[0015] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a first flowchart of a multi-target identification method for furniture main body provided in an embodiment of the present invention; Figure 2 This is a second flowchart of a multi-target identification method for furniture main body provided in an embodiment of the present invention; Figure 3 This is a third flowchart of a multi-target identification method for furniture main body provided in an embodiment of the present invention; Figure 4 This is a fourth flowchart of a multi-target identification method for furniture main body provided in an embodiment of the present invention; Figure 5 The fifth flowchart of a multi-target identification method for furniture main body provided in an embodiment of the present invention; Figure 6 The sixth flowchart of a multi-target identification method for furniture main body provided in an embodiment of the present invention; Figure 7 This is a seventh flowchart of a multi-target identification method for furniture main body provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of a multi-target recognition device for furniture bodies provided in an embodiment of the present invention; Figure 9 This is a structural schematic diagram of a multi-target recognition device for furniture bodies provided in an embodiment of the present invention. Detailed Implementation

[0016] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0017] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0018] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 An embodiment of a multi-target recognition method for furniture bodies according to the present invention includes: 101. Based on the preset furniture image annotation library and the preset instance segmentation model, the preset convolutional neural network model is pre-trained to obtain a deep convolutional neural network segmentation model; In this embodiment, the instance segmentation model is the yolov8-seg instance segmentation model developed using the yolov8 architecture. The convolutional neural network model is pre-trained using a furniture image annotation library and the instance segmentation model, so that the trained deep convolutional neural network segmentation model can simultaneously complete object detection and pixel-level segmentation tasks. 102. Optimize the deep convolutional neural network segmentation model based on preset model tuning parameters to obtain a furniture image subject recognition model; In this embodiment, the deep convolutional neural network segmentation model is optimized by adjusting the model tuning parameters until the loss value generated during the optimization process tends to 0 or equal to 0, thereby training a deep convolutional neural network segmentation model with high accuracy and faster inference speed. 103. Obtain the furniture image dataset; 104. Based on the furniture image subject recognition model and the preset vector database, perform similarity retrieval on the furniture image dataset to obtain the multi-target recognition results of the furniture subject; In this embodiment, by combining the architectural advantages of the instance segmentation model, object detection and pixel-level segmentation tasks are deeply integrated. Pre-training with a furniture image annotation library enables the pre-trained deep convolutional neural network segmentation model to simultaneously detect target locations and generate segmentation masks, significantly improving target image recognition efficiency. Model optimization based on parameter tuning allows the model to continuously converge to a low-loss state during training. This refined training strategy not only improves the sensitivity of the furniture image subject recognition model to details such as furniture outlines and materials but also optimizes and reduces computational overhead, increasing the inference speed of the furniture image subject recognition model. The trained furniture image subject recognition model can quickly locate target products from massive images, supporting applications such as similar style recommendations and style classification. Furthermore, the furniture image subject recognition model is based on a general architecture, supports transfer learning, and can adapt to different furniture images and even other home furnishing fields. It also supports multi-format export, facilitating deployment on various devices, providing technical support for the automation upgrade of furniture production, sales, and design, reducing labor costs, and promoting the digital transformation of the home furnishing industry.

[0019] Please see Figure 2 A second embodiment of a multi-target recognition method for furniture bodies according to the present invention includes: 201. Based on the furniture image subject recognition model and the preset confidence threshold, the furniture image dataset is processed to obtain a high-confidence dataset; In this embodiment, furniture images are filtered using a confidence threshold to ensure input data quality and reduce noise interference; 202. Generate a combined mask set based on the high-confidence dataset and the preset value range threshold interval; 203. Broadcast the high-confidence dataset according to the preset binarization threshold and combined mask set to obtain segmentation mask region information; In this embodiment, the mask is binarized using a uniform binarization threshold, and a broadcast mechanism is used to efficiently process batch data, thereby improving computational efficiency. 204. Based on the preset residual neural network model, perform full feature extraction on the segmented mask region information to obtain a vector feature set; In this embodiment, by using a residual neural network model to extract deep features from the mask region, the gradient vanishing problem faced by traditional neural networks in deep training is effectively solved, and richer semantic information (such as material and texture) can be captured. 205. Obtain the classification label set based on the high-confidence dataset; 206. Similarity retrieval is performed on the vector feature set and classification label set based on the vector database to obtain the multi-target recognition results of the furniture subject; In this embodiment, a confidence threshold is used to filter furniture image datasets, filtering low-quality data, reducing noise interference, and ensuring high-quality input data. Binarization thresholding and a broadcasting mechanism are used to efficiently process batch data, improving overall computational efficiency. A residual neural network model is used to extract deep features from masked regions, effectively overcoming the gradient vanishing problem and deeply mining semantic information such as furniture material and texture, enhancing feature representation capabilities. Similarity retrieval is performed using a vector database to quickly and accurately obtain multi-target recognition results for furniture subjects. This solution achieves full-process optimization from data processing to feature extraction and target recognition, improving processing speed while ensuring recognition accuracy, providing efficient technical support for furniture e-commerce, intelligent design, and other fields.

[0020] Please see Figure 3 A third embodiment of a multi-target recognition method for furniture bodies according to the present invention includes: 301. Set the size and configure the channels of the furniture image dataset to obtain a configured furniture image dataset; In this embodiment, the furniture image data is preprocessed, the image size is uniformly adjusted to 640*640, and the image channel format is set to RGB to ensure the consistency of the data format and provide standardized input for subsequent model processing; 302. Based on the furniture image subject recognition model, perform subject segmentation on the configuration furniture image dataset to obtain the furniture subject image dataset; In this embodiment, the main body segmentation of the furniture image dataset is performed, eliminating the influence of redundant background vector features on similarity retrieval, optimizing the strategy of using the entire image to extract vector features, improving feature discriminativeness, and enhancing the accuracy of data retrieval in big data. 303. Analyze the furniture subject image dataset based on the confidence threshold to obtain the confidence analysis results; 304. Based on the confidence analysis results, filter the furniture subject image dataset to obtain a high-confidence dataset; In this embodiment, an image subject recognition model is used to identify furniture subjects, filtering out subjects with a confidence level below a custom threshold and selecting furniture subjects with a high confidence level. In this embodiment, during the data preprocessing stage, the furniture images are standardized in size and channel format to eliminate data format differences, providing standardized input for model processing and reducing computational losses caused by inconsistent data specifications. Subject segmentation is performed based on the furniture image subject recognition model, eliminating background interference and abandoning the full-image processing strategy, focusing on extracting furniture subject features, enhancing feature discrimination power, and effectively improving the accuracy of big data retrieval. The segmented furniture subject images are rigorously screened using a confidence threshold to filter low-confidence data, ensuring that all data entering subsequent processes are high-reliability samples, further optimizing the recognition results, and providing efficient and accurate technical support for applications such as precise recommendations in furniture e-commerce and intelligent warehouse management.

[0021] Please see Figure 4 The fourth embodiment of a multi-target recognition method for furniture bodies in this invention includes: 401. Obtain the segmentation mask set based on the high-confidence dataset; 402. Perform value range analysis on the segmentation mask set based on the value range threshold interval to obtain the value range analysis results; In this embodiment, the segmentation mask M in the segmentation mask set ∈ ℝH×W, with a value range of [0, 255] or [0.0, 1.0]. ℝ represents the real number set, which is a set of all rational and irrational numbers in mathematics. Here, it is explained that the element values ​​of M are real numbers. 403. When the value range analysis result indicates that the value range of a segmentation mask is within the value range threshold interval, then obtain all segmentation masks whose value range is within the value range threshold interval from the segmentation mask set to obtain the mask set to be processed. 404. Normalize each mask in the mask set to be processed to obtain a normalized dataset; In this embodiment, when the value range of the segmentation mask M is within the threshold range of [0, 255], the segmentation mask M needs to be normalized; the normalization formula is: , This is the normalized segmentation mask; 405. Obtain the remaining dataset from the segmentation mask set based on the normalized dataset; 406. Generate a combined mask set based on the normalized dataset and the remaining dataset; In this embodiment, the mask set is analyzed by value range threshold intervals to filter out masks that meet specific value range requirements, forming a mask set to be processed, thus avoiding interference from invalid or abnormal data. The masks to be processed are normalized to unify the data scale, ensuring that masks from different value ranges are comparable in subsequent calculations and enhancing data stability. By distinguishing between the normalized dataset and the remaining dataset, a combined mask set is generated, which not only retains the effective information of the original mask but also optimizes the numerical expression of the mask. This allows the final combined mask set to more accurately reflect the outline and details of the furniture, providing high-quality data support for subsequent image analysis, target recognition, and other tasks.

[0022] Please see Figure 5 The fifth embodiment of a multi-target recognition method for furniture main body in this invention includes: 501. Binarize each mask in the combined mask set according to the binarization threshold to obtain a binarized mask set; In this embodiment, the purpose of binarizing the mask is to ensure that subsequent segmentation of the mask region information only retains the foreground region (main body). The mask is binarized by setting a threshold T (e.g., 0.5), and the expression for binarizing the mask is as follows: , This is the binarized result, used to distinguish between foreground and background regions. : Represents the index of the mask in the vertical direction (height direction), used to traverse each row of pixels, with a value range of [1,H], that is, from row 1 to row H; : Represents the index of the mask in the horizontal direction (width direction), used to traverse each column of pixels, with a value range of [1,W], that is, from the 1st column to the Wth column, resulting in a binary mask B∈{0,1}H×W; 502. Perform a three-dimensional extension on the binary mask set to obtain a three-dimensional mask set; In this embodiment, the two-dimensional mask B (i.e., the binarized mask) is extended to three dimensions for subsequent broadcast operations. The extension operation is as follows: In the formula, For a three-dimensional mask, the dimensions of the expanded three-dimensional mask are [H,W,3]; For tensor expansion (copying) functions, For binary masking, It is a real tensor with a height of H, a width of W, and 3 channels; 503. Broadcast the high-confidence dataset based on the 3D mask set to obtain segmentation mask region information; In this embodiment, the broadcast operation applies the mask from the 3D mask set to each channel of the high-confidence dataset (image), with the aim of retaining only the pixels of the main body region in the high-confidence data and setting the rest to zero. The expansion operation is as follows: element-wise multiplication (Hadamard product): ; In order to broadcast the image by element-wise multiplication (⊙), I is the high-confidence image data in the high-confidence dataset, I∈ℝH×W×3; The high-confidence image data after broadcasting can be used to obtain segmentation mask region information. In this embodiment, a binarization threshold is used to transform the combined mask into a binary mask set, effectively removing background information and retaining only the foreground subject. A three-dimensional expansion operation is used to upgrade the 2D mask to fit the RGB channel structure of the image, laying the foundation for broadcast operations. Finally, through broadcast operations, the 3D mask is applied element-wise to a high-confidence image to locate and extract the main furniture pixels, eliminating redundant background interference. This scheme simplifies data features, reduces invalid calculations, and enhances the saliency of subject features, providing clean and efficient data input for subsequent feature extraction and target recognition, thus improving the overall performance of furniture image analysis.

[0023] Please see Figure 6 The sixth embodiment of a multi-target recognition method for furniture bodies in this invention includes: 601. Obtain the fully connected layers of the residual neural network model; In this embodiment, the residual neural network model includes the ResNet-50 residual neural network model (a deep residual neural network), which consists of the following parts: Conv1: 7×7 convolution, stride = 2, MaxPooling layer (its function is to downsample the feature map output by Conv1); Conv1, Conv2_x, Conv3_x, Conv4_x, Conv5_x: each of the four different residual modules (each residual module includes multiple Bottleneck structure residual blocks); GlobalAveragePooling (GAP); fully connected layer (separator, usually outputting 1000 classes), which simplifies the network structure by removing layers that are irrelevant to the classification layer categories and removing redundant parts. This operation reduces unnecessary parameters and computations, allowing the network to focus on features related to furniture image recognition and avoid consuming resources on irrelevant tasks; 602. Perform a layer removal operation on the fully connected layer according to the preset classification layer category to obtain the updated connected layer; In this embodiment, the last fully connected layer (fc) is removed, while retaining the output characteristics of GlobalAveragePooling: 603. Configure the residual neural network model according to the updated connection layer to obtain the configured residual neural network model; 604. Pre-train the configured residual neural network model according to the preset image database to obtain the pre-trained residual neural network model; In this embodiment, the residual neural network model is pre-trained using the ImageNet image database as a subsequent feature extractor; ImageNet is one of the most influential image databases in the field of computer vision, aiming to promote research on tasks such as image classification, object detection, and semantic segmentation; 605. Based on a pre-trained residual neural network, perform full feature extraction on the segmentation mask region information to obtain a vector feature set; In this embodiment, the model focuses on extracting effective features of the main body of the furniture based on the general knowledge learned from pre-training and the simplified network structure, and generates vector features for subsequent identification and retrieval tasks. In another embodiment, the 2048-dimensional vector features can be further reduced to 1024 dimensions using a linear dimensionality reduction layer or a component analysis method (PCA principal component analysis). In this embodiment, by removing layers irrelevant to furniture recognition and retaining the output of the global average pooling layer, the network structure is simplified, reducing redundant parameters and invalid computations. This allows the model resources to focus on learning key features of furniture images, reducing computational overhead while enhancing task relevance. A large-scale image database is used for pre-training, leveraging its rich image semantics and feature patterns to endow the model with powerful general feature learning capabilities, effectively accelerating model convergence and improving generalization performance. The optimized pre-trained residual neural network, based on its simplified structure and pre-training knowledge, can extract deep semantic features such as material and texture of the furniture body, generating a highly discriminative vector feature set. This provides a solid data foundation for tasks such as multi-target recognition and similarity retrieval of furniture images, significantly improving the overall performance of the furniture image analysis system.

[0024] Please see Figure 7 The seventh embodiment of a multi-target recognition method for furniture bodies in this invention includes: 701. Normalize the vector feature set based on the preset vector normalization algorithm to obtain a normalized feature set; In this embodiment, the 2048-dimensional vector features need to be normalized using the L2 vector normalization algorithm before they can be entered into the Milvus vector database (which supports multimodal vector query combinations: image feature vector + classification label joint query) for similarity retrieval. The L2 vector normalization algorithm formula is as follows: In the formula, x = , where x1, x2, ..., xn are the vector features in the vector feature set, and x1, x2, ..., xn are the values ​​of each component in the vector features, and ||x||2 = the L2 norm (length) of the feature vector. Normalized features (length 1); 702. Classify the normalized feature set according to the classification label set to obtain the classification feature set; 703. Partition the classification feature set based on the vector database to obtain the partitioned feature set; In this embodiment, the classification label serves as a constraint, filtering before vector similarity matching to improve semantic accuracy. After classification, images of different categories can be stored in different partitions of the Milvus vector database, improving query efficiency and load balancing. The following is the expression for the partitioning operation: In the formula, C is the classification feature set. This represents a partition operation on set C, and the output is a subset. Each It is a subset of C that satisfies: ,Right now Includes all tags in C The sample, The features of the sample can be images, text, vectors, etc. The label for the sample indicates the category to which the sample belongs; 704. Based on the encoding method and the preset vector space, perform type conversion on the classification label set to obtain the label feature set; 705. Perform joint processing on the label feature set and the partition feature set to obtain a joint vector feature set; In this embodiment, the classification label can be encoded into a vector ec∈ℝ^k using the cone-hot encoding method, and then concatenated or fused with the image vector f (the image vector in the partition feature set) to obtain a joint representation. The concatenation operation expression is as follows: In the formula, For joint vector features, For the dimensions of label features, The dimension of the partition feature. "The set of all real vectors of dimension d + k"; The MLP fusion operation expression is as follows: In the formula, It is a multilayer perceptron, consisting of multiple fully connected layers (FC) and activation functions (such as ReLU), used to perform linear transformations on the concatenated features; 706. Perform similarity retrieval on the joint vector feature set to obtain the similarity feature set; 707. Reorder the similarity feature set according to the preset label logic rules to obtain the reordered feature set; 708. Generate multi-target recognition results for furniture main body based on the reordered feature set; In this embodiment, a vector normalization algorithm is used to standardize the vector feature set, ensuring consistent vector lengths and laying the foundation for efficient similarity retrieval in the vector database, thus improving the accuracy of multimodal queries (such as image features + classification labels). Based on the classification labels, the normalized features are partitioned and stored, placing furniture images of different categories in independent partitions of the vector database. This achieves semantic filtering during queries (filtering by label constraints before matching vectors) and optimizes load balancing, shortening the response time for large-scale data retrieval. Classification labels are converted into semantic vectors using an encoding method and concatenated or fused with image features to generate joint vectors. This effectively combines visual features (material, texture) with semantic information (category, style), enabling retrieval not only based on pixel-level similarity but also capturing semantic associations. The retrieval results are reordered using label logic rules to ensure that the output results meet the priorities of the business scenario, ultimately forming a multi-target furniture subject recognition result with both visual accuracy and semantic understanding, providing precise data support for e-commerce recommendations, intelligent warehousing, and other scenarios.

[0025] The above describes a method for multi-target recognition of furniture main body in an embodiment of the present invention. The following describes a device for multi-target recognition of furniture main body in an embodiment of the present invention. Please refer to [link / reference]. Figure 8 One embodiment of the multi-target recognition device for furniture bodies according to the present invention includes: Model pre-training module 1 is used to pre-train a preset convolutional neural network model based on a preset furniture image annotation library and a preset instance segmentation model, so as to obtain a deep convolutional neural network segmentation model. Model optimization module 2 is used to optimize the deep convolutional neural network segmentation model based on preset model tuning parameters to obtain a furniture image subject recognition model; Data acquisition module 3 is used to acquire furniture image datasets; The target subject recognition module 4 is used to perform similarity retrieval on the furniture image dataset based on the furniture image subject recognition model and the preset vector database to obtain the multi-target recognition results of the furniture subject; In this embodiment, by combining the architectural advantages of the instance segmentation model, object detection and pixel-level segmentation tasks are deeply integrated. Pre-training with a furniture image annotation library enables the pre-trained deep convolutional neural network segmentation model to simultaneously detect target locations and generate segmentation masks, significantly improving target image recognition efficiency. Model optimization based on parameter tuning allows the model to continuously converge to a low-loss state during training. This refined training strategy not only improves the sensitivity of the furniture image subject recognition model to details such as furniture outlines and materials but also optimizes and reduces computational overhead, increasing the inference speed of the furniture image subject recognition model. The trained furniture image subject recognition model can quickly locate target products from massive images, supporting applications such as similar style recommendations and style classification. Furthermore, the furniture image subject recognition model is based on a general architecture, supports transfer learning, and can adapt to different furniture images and even other home furnishing fields. It also supports multi-format export, facilitating deployment on various devices, providing technical support for the automation upgrade of furniture production, sales, and design, reducing labor costs, and promoting the digital transformation of the home furnishing industry.

[0026] Figure 9 This is a schematic diagram of the structure of a furniture body multi-target recognition device 900 provided in an embodiment of the present invention. This furniture body multi-target recognition device 900 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 910 (e.g., one or more processors) and a memory 920, and one or more media 930 (e.g., one or more mass storage devices) storing application programs 933 or data 932. The memory 920 and media 930 can be temporary or persistent storage. The program stored in the media 930 may include one or more modules (not shown in the diagram), each module may include a series of instruction operations on the furniture body multi-target recognition device 900. Furthermore, the processor 910 may be configured to communicate with the media 930 and execute a series of instruction operations in the media 930 on the furniture body multi-target recognition device 900 to implement the steps of the furniture body multi-target recognition method provided in the above-described method embodiments.

[0027] A multi-target recognition device 900 for furniture bodies may further include one or more power supplies 940, one or more wired or wireless network interfaces 950, one or more input / output interfaces 960, and / or one or more operating systems 931, such as Windows Server, MacOSX, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 9 The illustrated structure of a multi-target recognition device for furniture bodies does not constitute a limitation on a multi-target recognition device 900 for furniture bodies. It may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0028] A computer-readable medium storing instructions that, when executed by a processor, implement the steps of a multi-target furniture subject recognition method as described above.

[0029] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention. The actual content is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A method for multi-target recognition of furniture main body, characterized in that, include: Based on a pre-defined furniture image annotation library and a pre-defined instance segmentation model, a pre-defined convolutional neural network model is pre-trained to obtain a deep convolutional neural network segmentation model. The deep convolutional neural network segmentation model is optimized based on preset model tuning parameters to obtain a furniture image subject recognition model. Obtain a dataset of furniture images; Based on the furniture image subject recognition model and the pre-set vector database, similarity retrieval is performed on the furniture image dataset to obtain multi-target recognition results of furniture subjects.

2. The method for multi-target recognition of furniture main body as described in claim 1, characterized in that, The similarity retrieval of the furniture image dataset based on the furniture image subject recognition model and a preset vector database to obtain multi-target furniture subject recognition results includes: The furniture image dataset is processed based on the furniture image subject recognition model and a preset confidence threshold to obtain a high-confidence dataset; Generate a combined mask set based on the high-confidence dataset and the preset value range threshold interval; The high-confidence dataset is broadcast based on a preset binarization threshold and a combined mask set to obtain segmentation mask region information; Based on a pre-defined residual neural network model, full feature extraction is performed on the segmented mask region information to obtain a vector feature set; The classification label set is obtained from the high-confidence dataset; Similarity retrieval is performed on vector feature sets and classification label sets based on a vector database to obtain multi-target identification results for furniture subjects.

3. The method for multi-target recognition of furniture main body as described in claim 2, characterized in that, The process of processing the furniture image dataset based on the furniture image subject recognition model and a preset confidence threshold to obtain a high-confidence dataset includes: The furniture image dataset is sized and configured with channels to obtain a configured furniture image dataset; Based on the furniture image subject recognition model, the configuration furniture image dataset is segmented to obtain the furniture subject image dataset; The confidence analysis results were obtained by analyzing the furniture subject image dataset based on the confidence threshold. The furniture subject image dataset was filtered based on the confidence analysis results to obtain a high-confidence dataset.

4. The method for multi-target recognition of furniture main body as described in claim 2, characterized in that, The step of generating a combined mask set based on the high-confidence dataset and a preset value range threshold includes: The segmentation mask set is obtained from the high-confidence dataset; The segmentation mask set is analyzed based on the value range threshold interval to obtain the value range analysis results; If the value range analysis result indicates that the value range of a segmentation mask is within the value range threshold interval, then all segmentation masks whose value range is within the value range threshold interval are obtained from the segmentation mask set to obtain the mask set to be processed. Normalize each mask in the mask set to be processed to obtain a normalized dataset; The remaining dataset is obtained from the segmentation mask set based on the normalized dataset; Generate a combined mask set based on the normalized dataset and the remaining dataset.

5. The method for multi-target recognition of furniture main body as described in claim 2, characterized in that, The step of broadcasting the high-confidence dataset according to a preset binarization threshold and a combined mask set to obtain segmented mask region information includes: Binarize each mask in the combined mask set according to the binarization threshold to obtain a binarized mask set; The binary mask set is extended in three dimensions to obtain a three-dimensional mask set; Broadcast the high-confidence dataset based on the 3D mask set to obtain segmentation mask region information.

6. The method for multi-target recognition of furniture main body as described in claim 2, characterized in that, The preset residual neural network model performs full feature extraction on the segmented mask region information to obtain a vector feature set, including: Obtain the fully connected layers of the residual neural network model; The fully connected layer is removed according to the preset classification layer category to obtain the updated connected layer; The residual neural network model is configured based on the updated connection layers to obtain the configured residual neural network model; The configuration residual neural network model is pre-trained based on a preset image database to obtain a pre-trained residual neural network model; Full feature extraction of segmentation mask region information is performed based on pre-trained residual neural network to obtain vector feature set.

7. The method for multi-target recognition of furniture main body as described in claim 2, characterized in that, The similarity retrieval based on the vector database of vector feature sets and classification label sets to obtain multi-target identification results for furniture subjects includes: The vector feature set is normalized based on a preset vector normalization algorithm to obtain a normalized feature set. The normalized feature set is classified according to the classification label set to obtain the classification feature set; The classification feature set is partitioned based on a vector database to obtain a partitioned feature set; The classification label set is type-converted according to the encoding method and the preset vector space to obtain the label feature set; The label feature set and the partition feature set are jointly processed to obtain a joint vector feature set; Similarity retrieval is performed on the joint vector feature set to obtain the similarity feature set; The similarity feature set is reordered according to the preset label logic rules to obtain the reordered feature set; The furniture main body multi-target recognition results are generated based on the reordered feature set.

8. A multi-target recognition device for furniture bodies, characterized in that, include: The model pre-training module is used to pre-train a pre-defined convolutional neural network model based on a pre-defined furniture image annotation library and a pre-defined instance segmentation model, so as to obtain a deep convolutional neural network segmentation model. The model optimization module is used to optimize the deep convolutional neural network segmentation model based on preset model tuning parameters to obtain a furniture image subject recognition model. The data acquisition module is used to acquire furniture image datasets; The target subject recognition module is used to perform similarity retrieval on the furniture image dataset based on the furniture image subject recognition model and the preset vector database to obtain multi-target recognition results of furniture subjects.

9. A multi-target recognition device for furniture bodies, characterized in that, include: A memory and at least one processor, wherein the memory stores instructions; At least one of the processors invokes the instructions in the memory to cause the furniture body multi-target recognition device to perform the steps of the furniture body multi-target recognition method as claimed in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the various steps of the furniture body multi-target recognition method as described in any one of claims 1-7.

Images