CN114821308A - Furniture style identification and generation method - Google Patents
Furniture style identification and generation method Download PDFInfo
- Publication number
- CN114821308A CN114821308A CN202210373630.8A CN202210373630A CN114821308A CN 114821308 A CN114821308 A CN 114821308A CN 202210373630 A CN202210373630 A CN 202210373630A CN 114821308 A CN114821308 A CN 114821308A
- Authority
- CN
- China
- Prior art keywords
- furniture
- image
- style
- pictures
- classification
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Abstract
The invention relates to a furniture style identification and generation method, which comprises the following steps: the invention realizes the purpose through the following technical scheme: a furniture style identification and generation method comprises the following steps: (1) setting a furniture label database and a classification database, wherein the label database is provided with a plurality of label vocabularies, sets the existing styles of furniture in the classification database for classification, inputs a furniture picture or information, calls the label database and the classification database to perform initial classification on the furniture; (2) selecting a convolutional neural network to establish a plurality of image recognition classification models, (3) establishing a furniture style database; (4) carrying out batch matting on a plurality of furniture pictures to obtain the furniture pictures with pure white backgrounds; (5) inputting the processed furniture pictures into a Style GAN2 network for two times to generate results; (6) generating an image and video file; (7) observing and browsing the generated images and videos, and identifying a preferred scheme; (8) finishing the design and optimization of the product appearance design scheme, and presenting a design drawing and displaying the design drawing to an interface; the method of the invention has inexhaustible originality; training the thinking of the designer, and outputting efficiently and stably; the designed threshold is low, and more originality is excited; avoiding some legal risks or economic consequences, avoiding design thinking solidification and the like.
Description
Technical Field
The invention relates to the field of furniture design, in particular to a furniture style identification and generation method.
Background
The furniture industry of China develops for decades to form a certain industrial scale, the furniture production and export have already occupied a very important position in the international furniture industry, along with the development and wide application of the economic globalization trend and the Internet, the method brings more fierce competition for enterprises, and forces the enterprises to continuously seek new technology, new method and new means so as to improve the core competitiveness of the enterprises in the market and further obtain survival and development, under the background, the design capability is very important, and the design capability becomes a new driving force in some industries; nowadays, the furniture industry is also at the key node of transformation and upgrading, and how to make the old furniture industry generate new vitality by relying on a novel computer technology and a design theory becomes a problem which needs to be considered and solved urgently for enterprises and designers;
in the design method, because the modern design in China starts late, the creative method still has a large gap compared with the world advanced level, and the creative method is mainly expressed in the following aspects: firstly, designers need to spend a great deal of time and energy to select knowledge for inspiring creative inspiration from original case materials, so that the innovation capability is weak; secondly, the traditional design method depends on the personal experience of designers to a great extent, the designed works also have strong personal willingness of the designers, are easily influenced by the subjective emotion, ability and the like of the designers, have certain uncertainty and have high labor cost of the designers; finally, the aspects of mining user requirements, insights and design problems, scheme evaluation and examination and the like still stay in design and research methods such as questionnaires, interviews, behavior observation, brainstorming, subjective evaluation and the like, and subconscious ideas and feedback of the user are difficult to reach, so that a matching error exists between a design result and a real appeal of the user; meanwhile, with the development of the technology and the development of the design style towards simplification, the homogenization phenomenon of furniture products is more and more serious, and higher requirements are put forward on the innovation capability of designers.
Disclosure of Invention
The invention aims to provide a furniture style identification and generation method, which solves the problems.
The invention realizes the purpose through the following technical scheme: a furniture style identification and generation method comprises the following steps:
(1) appointing a piece of furniture, and marking the existing style of the furniture by adopting various label vocabularies;
(2) selecting a convolutional neural network to establish a plurality of image recognition classification models;
(3) establishing a furniture style database;
(4) carrying out batch matting on a plurality of furniture pictures to obtain the furniture pictures with pure white backgrounds;
(5) inputting the processed furniture pictures into a Style GAN2 network for two times to generate results;
(6) generating an image and video file;
(7) observing and browsing the generated images and videos, and identifying a preferred scheme;
(8) and finishing the design and optimization of the product appearance design scheme, and presenting a design drawing and displaying the design drawing to an interface.
Preferably, the furniture style database is subjected to clustering analysis by a systematic clustering method, 4-6 main categories are extracted, and then the categories are named.
The method comprises clustering all classes by a system clustering method, extracting main components by a main component analysis method as auxiliary classification basis to classify the classes, defining the distance relationship among samples by the system clustering method, clustering the samples closest to the samples into subclasses, combining the clustered subclasses according to the inter-class distance, continuing the clustering, and finally clustering the subclasses into a wholeThe clustering algorithm is connected in an average mode in a group, the distance calculation adopts the Euclidean distance to generate a pedigree diagram, and the calculation formula of the Euclidean distance is as follows:
wherein, in the step (A),
、
respectively represent the ith word, and n is the number of clusters.
Preferably, 4 principal components are extracted, 65.84% of information content in 14 variables is explained by the factors of the 4 principal components, the distance in the pedigree diagram is taken as 17, and the result of the clustering analysis is divided into 4 classes by the line drawing, wherein the classes are respectively as follows: streamline, parametric, industrial wind; modern, netred, simple, classical, northern; new Chinese, Japanese, traditional Chinese; european, American, Italian luxury, naming these 4 classes separately, resulting in the final style label vocabulary.
Preferably, the method selects RseNet50 to establish an image recognition classification model, and sets the following parameters: the image input is an RGB color 3 channel 224 x 224 image; the ratio of the training set to the test set is 4: 1; the activation function is ReLU; softmax regression classification; accelerating calculation by the GPU; the image adopts a data enhancement strategy (color enhancement, random angle, increasing cutting and horizontal random turning); batch _ size is 64; the number of training rounds is 30; the learning rate is 0.001, and the learning rate is set to be 0.002 in the Adam optimizer; the program has cleaned the unopened data.
Preferably, the step of building the database by using deep learning collected data is as follows:
(1) collecting a volume of source data from a network;
(2) cleaning other data which are not the target product by using the image recognition classification model;
(3) classifying labels of the target products by using an image recognition classification model;
(4) sorting the classification results and establishing a furniture style database.
As a preferred aspect of the present invention, the step of building a database using deep learning collected data is as follows
(1) Collecting a volume of source data from a network;
(2) cleaning other data which are not chairs by using the image recognition classification model, manually and simply browsing to judge the classification accuracy, and entering the next step if the data are accurate; if the result is inaccurate, checking the reason, and re-identifying and classifying;
(3) the method comprises the steps of utilizing an image to identify a label of a classification model chair, automatically labeling image data through deep learning, and manually rechecking after identification is finished;
(4) sorting the classification results and establishing a chair stylized database;
collecting pictures of a specific chair, and requiring that only one complete chair is contained in one picture; the image size is not less than 224 × 224 pixels; the number of pictures among all styles is equal; the chair in the same style only has two chairs with different materials, colors or shooting angles; in order to meet the requirement of deep learning on big data, the number of the data required for image acquisition is not less than 800;
printing style labels on pictures, selecting RseNet50 as a target network, establishing an image recognition classification model, and inputting the pictures into RGB color 3 channel 224 x 224 images; the ratio of the training set to the test set is 4: 1; the activation function is ReLU; softmax regression classification; accelerating calculation by the GPU; the image adopts color enhancement, random angle, increasing cutting and horizontal random turning; batch _ size is 64; the number of training rounds is 30; the learning rate is 0.001, and the learning rate is set to be 0.002 in the Adam optimizer; the program cleans unopened data;
selecting pictures of chairs and non-chairs from source data pictures, inputting the pictures into the established image recognition classification model to train the network, and enabling the network model to learn the characteristics of the chairs/non-chairs, namely realizing the batch classification of the images; statistically analyzing and correcting the erroneous image;
screening an image file from an oriental style database as an alternative image for the optimization, scratching off all background images and image shadows, inputting the obtained image into a StyleGAN2 network for training, and after 3000 steps (steps) of training, setting the FID score (FreechetInclusionDistance) to be 24.73;
selecting the step number to be 1-3000, downloading the training process image by taking 500 steps as a first file, and corresponding to the FID score change; and (3) intercepting partial images (3 multiplied by 3) to show a training process, naming the single group of 3 multiplied by 3 images of the upper image from left to right and from top to bottom, wherein the chair images are formed when the FID score is reduced to 68.66 in the training 1000 steps, and the chair is more truthfully and clearly presented in the subsequent 2000 steps and is finally slowly and completely transferred into the eastern style chair.
Preferably, 1391 piece of image data of furniture with a white background and a shadow-free specific style towards right is trained for 3000 steps, a corresponding FID score is obtained, training process images are downloaded for one file according to 500 steps, the situation that the furniture starts to be slowly changed into furniture from a car after the training of a furniture data set of about 500 steps is obtained, the shadow of the car can be seen, most of images are already formed when the FID score is reduced to 68.66 in the training of 1000 steps, the furniture is more truly and clearly shown in the subsequent 2000 steps, and finally the furniture is slowly and completely changed into the furniture with the preset style.
Preferably, the produced furniture image is divided into two parts according to an ideal sample and an error sample, and then the two parts are compared to obtain the front-facing furniture image, although the number of data sets is small, the structural characteristics of the furniture are basically fixed, and different design results can be generated.
As the optimization of the invention, the generated more perfect images mostly directly stimulate inspiration design, the whole shape is similar to the final design image, in the design method, the advantages in the original design are extracted through the experience of a designer, and the product is perfected through the design in the aspects of CMF and the like.
In the present invention, it is preferable that an image with incomplete or erroneous generation is generated for image association inspiring inspiration design, and since the original image information is seriously lost, a designer needs to exert his/her own imagination based on the original image to design the image, and in the image design, the styling ability of the designer and the comprehension ability of the multi-product itself are exerted more, and the designer needs to integrate his/her own design concept into the product as well as to pay attention to CMF and the like, and the generated image provides only abstract concepts or intention inspiration, and thus the assistance of styling is less.
Compared with the prior art, the invention has the following beneficial effects: the furniture is classified and summarized according to styles, and then the furniture style is successfully recognized by utilizing a furniture recognition and classification model of a convolutional neural network, and the furniture recognition and classification model is practically applied to data set collection and classification work; a set of data set collection method is established by utilizing a furniture identification classification model, and the effectiveness and the high efficiency of the furniture identification classification model are improved; the method of the invention has inexhaustible originality; training the thinking of the designer, and outputting efficiently and stably; the designed threshold is low, and more originality is excited; avoiding some legal risks or economic consequences, avoiding design thinking solidification and the like.
Drawings
FIG. 1 is a schematic diagram of a furniture style clustering lineage according to an embodiment of the invention;
FIG. 2 is a schematic flow chart of a data set collection method according to an embodiment of the invention;
FIG. 3 is a schematic diagram of a StyleGAN2 network training picture generation process preview representation according to an embodiment of the present invention;
FIG. 4 is a schematic diagram of a design flow based on generated images according to an embodiment of the present invention.
Detailed Description
The invention will be further described with reference to the accompanying drawings in which:
a furniture style identification and generation method, as shown in fig. 1-4, comprising the steps of:
(1) appointing a piece of furniture, and marking the existing style of the furniture by adopting a proper label vocabulary;
(2) selecting a convolutional neural network to establish an image recognition classification model;
(3) establishing a furniture style database;
(4) carrying out batch matting on a plurality of furniture pictures to obtain the furniture pictures with pure white backgrounds;
(5) inputting the processed furniture pictures into a Style GAN2 network for two times to generate results;
(6) generating an image and video file;
(7) observing and browsing the generated images and videos, and searching for inspiration;
(8) and finishing the appearance design of the product manually, and presenting the final effect.
Firstly, roughly classifying some furniture styles through some obvious classification standards, simplifying and summarizing the furniture styles, carrying out clustering analysis through a system clustering method, defining the distance relation between samples, generating a furniture style clustering pedigree diagram by adopting an Euclidean distance calculation formula, extracting 4-6 main categories, and then naming.
Selecting RseNet50 to establish an image recognition classification model, and setting the following parameters: the image input is an RGB color 3 channel 224 x 224 image; the ratio of the training set to the test set is 4: 1; the activation function is ReLU; softmax regression classification; accelerating calculation by the GPU; the image adopts a data enhancement strategy (color enhancement, random angle, increasing cutting and horizontal random turning); batch _ size is 64; the number of training rounds is 30; the learning rate is 0.001, and the learning rate is set to be 0.002 in the Adam optimizer; the program has cleaned the unopened data.
The steps of using deep learning to collect data to build a database are as follows:
1. collecting a volume of source data from a network;
2. cleaning other data which are not the target product by using the image recognition classification model;
3. classifying labels of the target products by using an image recognition classification model;
4. sorting the classification results and establishing a furniture style database.
Training 1391 pieces of image data of furniture with a white background and a certain shade and a right style for 3000 steps, acquiring a corresponding FID score, downloading training process images by taking 500 steps as a file, and starting to slowly change the furniture from a car after approximately 500 steps of training of a furniture data set, wherein the shadow of the car can be seen, most of images are formed when the FID score is reduced to 68.66 after 1000 steps of training, furniture images are formed, the furniture is more truly and clearly displayed in the subsequent 2000 steps, and finally the furniture is slowly and completely transferred into the furniture with a preset style.
The produced furniture image is divided into two parts according to an ideal sample and an error sample, and then the two parts are compared to obtain a front-facing furniture image, although the number of data sets is small, the furniture structure characteristics are basically fixed, and different design results can be generated.
Most of generated perfect images directly stimulate inspiration design, the overall shape is similar to the final design image, in the design method, advantages in the original design are extracted through the experience of a designer, and the product is perfect through the design in the aspects of CMF and the like.
The method is characterized in that an image with incomplete image defect or wrong image generation is generated for image association inspiring inspiration design, because original image information is seriously lost, a designer can design the image by exerting self imagination based on the original image, in the image design, the modeling capability of the designer and the understanding capability of multiple products are exerted more, the designer not only pays attention to aspects such as CMF (China Mobile switching) and the like, but also needs to integrate self design concepts into the products, the generated image only provides abstract concepts or intention inspiration, and the modeling is less helped.
In use, taking a chair as an example, 40 furniture styles are selected, and then the 40 styles are reduced to 14 representative styles in a generalized way, namely: modern, new chinese, light style luxury, classic northern europe, netred, industrial wind, extreme simple, streamlined, parameterized, european, simple european, american, traditional chinese, japanese;
the tested subject is required to group and classify the 14 words describing the chair style according to the modeling characteristics of the chair, the grouping number and the number of each group are not limited, the frequency of each word in the same group is counted after data are recycled, a 13 x 13 similarity matrix is made, the numerical value in the matrix represents the similarity degree between the words, and the larger the numerical value is, the higher the similarity degree is; extracting a principal component as a reference by adopting a factor analysis-principal component analysis method; analyzing clustering results by adopting clustering;
the clustering method is a method of cluster analysis, and comprises defining the distance relationship between samples, and clustering the samples with the closest distance into small samplesAnd class, merging the aggregated subclasses according to the distance between the classes, continuing in this way, finally clustering the subclasses into a large class, applying average connection in a group by using a clustering algorithm, generating a pedigree diagram by adopting Euclidean distance in distance calculation, wherein the calculation formula of the Euclidean distance is as follows:
wherein, in the step (A),
、
respectively representing ith words, wherein n is the number of clusters; 4 main components are extracted, namely: the 4 factors can explain 65.84% of information quantity in 14 variables, and by combining the characteristics of deep learning and market conditions, the distance in the pedigree diagram is taken as 17, and the result of the clustering analysis is divided into 4 classes by dividing lines, wherein the 4 classes are respectively as follows: streamline, parametric, industrial wind; modern, netred, simple, classical, northern; new Chinese, Japanese, traditional Chinese; european, american, and artistic luxury. Naming the 4 classes respectively to obtain final style label vocabularies;
the method comprises the following steps of collecting image data by using a web crawler, inducing and classifying the image data, establishing a database, and establishing the database by using a deep learning collected data set:
(1) collecting a volume of source data from a network;
(2) other data that are not the target product (chair) are purged using the image recognition classification model. Such as sofas, fabrics, patterns, indoor designs and the like, and the accuracy of classification is judged by simply browsing manually. If the accuracy is correct, the next step is carried out; if the result is inaccurate, checking the reason, and re-identifying and classifying;
(3) the labels (styles) of the target products (chairs) are classified using an image recognition classification model. Namely, the image data is automatically labeled through deep learning. Because deep learning has certain uncertainty, manual rechecking is also needed after recognition is finished;
(4) sorting the classification results and establishing a (chair stylized) database;
collecting pictures of a specific chair, and requiring that only one complete chair is contained in one picture; the image size is not less than 224 × 224 pixels; the number of pictures among all styles is approximately equal; the chair in the same style only has two chairs with different materials, colors or shooting angles and the like; in order to meet the requirement of deep learning on big data, the number of image acquisition is as much as possible (each style of image data is not less than 800);
printing style labels on pictures, selecting RseNet50 as a target network, establishing an image recognition classification model, and inputting the pictures into RGB color 3 channel 224 x 224 images; the ratio of the training set to the test set is 4: 1; the activation function is ReLU; softmax regression classification; GPU acceleration calculation; the image adopts a data enhancement strategy (color enhancement, random angle, increasing cutting and horizontal random turning); batch _ size is 64; the number of training rounds is 30; the learning rate is 0.001, and the learning rate is set to be 0.002 in the Adam optimizer; the program cleans unopened data;
selecting pictures of chairs and non-chairs from a large number of source data pictures, inputting the pictures into the established image recognition classification model to train the network, and enabling the network model to learn the characteristics of the chairs/non-chairs, namely realizing the batch classification of the images; statistically analyzing and correcting the erroneous image;
screening a batch of image files with simple backgrounds, easy cutout and obvious style characteristics from an oriental style database as alternative images for optimization, carrying out batch cutout, removing all background images and image shadows, inputting the obtained images into a StyleGAN2 network for training, and after 3000 steps (steps) of training, obtaining a FID (Freechet inclusion distance) of 24.73;
selecting the step number to be 1-3000, and downloading the training process image by taking 500 steps as a first file, wherein the step number just corresponds to the FID score change; intercepting partial images (3 multiplied by 3) to show a training process, naming a single group of 3 multiplied by 3 images of an upper image as figures 1-9 from left to right and from top to bottom, wherein the images can be clearly seen from a generation process, the generation process applies transfer learning, learning pre-training data is an automobile, the automobile starts to be slowly changed into a chair from the automobile after training of chair data sets of 500 steps approximately, the shadow of the automobile can be seen at the moment, the chair images are already formed when most of the images are trained for 1000 steps and the FID score is reduced to 68.66, the chair is more truthfully and clearly presented in the subsequent 2000 steps, and finally the chair is slowly and completely changed into the chair with the oriental style;
finding inspiration from the finally generated picture, comprising the steps of:
(1) the method is characterized in that the inspiration is directly excited by the image, the inspiration is obtained from the original image, and the product is finally designed and the generated image is not separated by fine adjustment of the original generated image and improvement of the structure and the details;
(2) the image association stimulates inspiration, abstract concepts or intentions are obtained from generated images firstly, then the concepts or intentions are associated, the abstractions are converted into the imagination, and a great number of new design schemes are presented in the process; the method is mainly used for acquiring inspiration from images which are not complete enough or generate errors;
(3) the method comprises the steps of generating an image prototype through hand drawing, and manually drawing the outline and key parts of an original generated image once, so that the characteristics of the generated image are more clearly understood, and the hand drawing process is a thinking and redesigning process;
(4) and (4) performing hand-drawing deduction, and after copying the original to generate an image, performing hand-drawing deduction design. Different from the traditional reference design case images, the reference images based on the generated images have no complete design cases for reference in the design, and the design results are thousands of people, so that the phenomenon of design thinking solidification is not easy to occur;
(5) modeling deduction, the process pays more attention to knowledge and experience of designers, and the original fuzzy generated draft is visualized;
(6) and rendering and proofing to finish the final effect presentation, including aspects such as CMF and the like.
And finally, combining the chair style picture generated by the method by a designer, and further refining and designing the finished product.
The foregoing illustrates and describes the principles, general features, and advantages of the present invention. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, which are described in the specification and illustrated only to illustrate the principle of the present invention, but that various changes and modifications may be made therein without departing from the spirit and scope of the present invention, which fall within the scope of the invention as claimed.
Claims (9)
1. A furniture style identification and generation method is characterized by comprising the following steps:
(1) appointing a kind of furniture, and marking the existing style of the furniture by adopting various label vocabularies;
(2) selecting a convolutional neural network to establish a plurality of image recognition classification models;
(3) establishing a furniture style database;
(4) carrying out batch matting on a plurality of furniture pictures to obtain the furniture pictures with pure white backgrounds;
(5) inputting the processed furniture pictures into a Style GAN2 network for two times to generate results;
(6) generating an image and video file;
(7) observing and browsing the generated images and videos, and identifying a preferred scheme;
(8) and finishing the design and optimization of the product appearance design scheme, and presenting a design drawing and displaying the design drawing to an interface.
2. The furniture style identification and generation method of claim 1, wherein the furniture style database is clustered by a systematic clustering method, and 4-6 main categories are extracted and named.
3. The method of claim 2, wherein the categories are clustered by a systematic clustering method, and then the main components are extracted by a main component analysis method to be used as an auxiliary classification basis for classifying the categories, the systematic clustering method first defines the distance relationship between samples, the samples with the closest distance are first grouped into subclasses, and then the grouped subclasses are combined according to the distance between the subclasses, such asAnd finally, clustering the small sub-classes into a large class, wherein the clustering algorithm uses average connection in a group, the distance calculation adopts the Euclidean distance to generate a pedigree diagram, and the calculation formula of the Euclidean distance is as follows:
wherein, in the step (A),
、
respectively represent the ith word, and n is the number of clusters.
4. The furniture style identification and generation method according to claim 3, wherein 4 principal components are extracted, the 4 principal components have factors explaining 65.84% of information content in 14 variables, the distance in the pedigree diagram is taken as 17, and the division lines divide the results of the cluster analysis into 4 categories, which are respectively: streamline, parametric, industrial wind; modern, netred, simple, classical, northern; new Chinese, Japanese, traditional Chinese; and 4, naming the 4 classes respectively to obtain a final style label vocabulary.
5. The furniture style recognition and generation method according to claim 1, wherein RseNet50 is selected to build an image recognition classification model, and the following parameters are set: the image input is an RGB color 3 channel 224 x 224 image; the ratio of the training set to the test set is 4: 1; the activation function is ReLU; softmax regression classification; accelerating calculation by the GPU; the image adopts a data enhancement strategy (color enhancement, random angle, increasing cutting and horizontal random turning); batch _ size is 64; the number of training rounds is 30; the learning rate is 0.001, and the learning rate is set to be 0.002 in the Adam optimizer; the program cleans the unopened data.
6. The furniture style identification and generation method of claim 1, wherein the step of building the database by using deep learning collected data comprises:
collecting a volume of source data from a network;
cleaning other data which are not the target product by using the image recognition classification model;
classifying labels of the target products by using an image recognition classification model;
sorting the classification results and establishing a furniture style database.
7. The furniture style identification and generation method of claim 6,
(1) collecting a volume of source data from a network;
(2) cleaning other data which are not chairs by using the image recognition classification model, manually and simply browsing to judge the classification accuracy, and entering the next step if the data are accurate; if the result is inaccurate, checking the reason, and re-identifying and classifying;
(3) the method comprises the steps of automatically labeling image data through deep learning by utilizing labels of an image recognition classification model chair, and manually rechecking after recognition is finished;
(4) sorting the classification results and establishing a chair stylized database;
collecting pictures of a specific chair, and requiring that only one complete chair is contained in one picture; the image size is not less than 224 × 224 pixels; the number of pictures among all styles is equal; the chair in the same style only has two chairs with different materials, colors or shooting angles; in order to meet the requirement of deep learning on big data, the number of the data required for image acquisition is not less than 800;
printing style labels on pictures, selecting RseNet50 as a target network, establishing an image recognition classification model, and inputting the pictures into RGB color 3 channel 224 x 224 images; the ratio of the training set to the test set is 4: 1; the activation function is ReLU; softmax regression classification; accelerating calculation by the GPU; the image adopts color enhancement, random angle, increasing cutting and horizontal random turning; batch _ size is 64; the number of training rounds is 30; the learning rate is 0.001, and the learning rate is set to be 0.002 in the Adam optimizer; the program cleans unopened data;
selecting pictures of chairs and non-chairs from source data pictures, inputting the pictures into the established image recognition classification model to train the network, and enabling the network model to learn the characteristics of the chairs/non-chairs, namely realizing the batch classification of the images; statistically analyzing and correcting the erroneous image;
screening an image file from an oriental style database as an alternative image for the optimization, scratching off all background images and image shadows, inputting the obtained image into a StyleGAN2 network for training, and after 3000 steps (steps) of training, setting the FID score (FreechetInclusionDistance) to be 24.73;
selecting the step number to be 1-3000, downloading the training process image by taking 500 steps as a first file, and corresponding to the FID score change; the training process is shown by intercepting partial images (3 x 3), the single group of 3 x 3 images of the upper image is named from left to right and from top to bottom, the chair images are formed when the FID score is reduced to 68.66 in the training 1000 steps, and the chair is more clearly shown in the subsequent 2000 steps and finally transferred to the eastern style chair.
8. The furniture style identification and generation method of claim 1, wherein a plurality of images of furniture with white background and no shadow of a specific style towards right are collected, a plurality of steps are trained, corresponding FID scores are obtained, and a picture of a production process is collected.
9. The furniture style identification and generation method of claim 1, wherein the produced furniture image is divided into two parts according to an ideal sample and an error sample, and the comparison and the error correction are performed.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210373630.8A CN114821308A (en) | 2022-04-11 | 2022-04-11 | Furniture style identification and generation method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210373630.8A CN114821308A (en) | 2022-04-11 | 2022-04-11 | Furniture style identification and generation method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114821308A true CN114821308A (en) | 2022-07-29 |
Family
ID=82534715
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210373630.8A Pending CN114821308A (en) | 2022-04-11 | 2022-04-11 | Furniture style identification and generation method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114821308A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116052137A (en) * | 2023-01-30 | 2023-05-02 | 北京化工大学 | Deep learning-based classical furniture culture attribute identification method and system |
-
2022
- 2022-04-11 CN CN202210373630.8A patent/CN114821308A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116052137A (en) * | 2023-01-30 | 2023-05-02 | 北京化工大学 | Deep learning-based classical furniture culture attribute identification method and system |
| CN116052137B (en) * | 2023-01-30 | 2024-01-30 | 北京化工大学 | Deep learning-based classical furniture culture attribute identification method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109977413B (en) | Emotion analysis method based on improved CNN-LDA | |
| CN110378985B (en) | Animation drawing auxiliary creation method based on GAN | |
| CN108629338B (en) | Face beauty prediction method based on LBP and convolutional neural network | |
| CN110532912B (en) | Sign language translation implementation method and device | |
| CN109241255A (en) | A kind of intension recognizing method based on deep learning | |
| CN109871851B (en) | Chinese character writing normalization judging method based on convolutional neural network algorithm | |
| CN108717856A (en) | A kind of speech-emotion recognition method based on multiple dimensioned depth convolution loop neural network | |
| CN111144448A (en) | Video barrage emotion analysis method based on multi-scale attention convolutional coding network | |
| CN108596051A (en) | A kind of intelligent identification Method towards product style image | |
| CN107038154A (en) | A kind of text emotion recognition methods and device | |
| CN105930416A (en) | Visualization processing method and system of user feedback information | |
| CN112883709B (en) | Method for automatically generating antithetical couplet by natural language processing technology | |
| CN109492105A (en) | A kind of text sentiment classification method based on multiple features integrated study | |
| CN109359296B (en) | Public opinion emotion recognition method and device and computer readable storage medium | |
| CN109213853A (en) | A kind of Chinese community's question and answer cross-module state search method based on CCA algorithm | |
| CN107145514A (en) | Chinese sentence pattern sorting technique based on decision tree and SVM mixed models | |
| CN110532421A (en) | A kind of gender based on people, the music recommended method and system of mood and age | |
| CN115294427A (en) | Stylized image description generation method based on transfer learning | |
| CN114821308A (en) | Furniture style identification and generation method | |
| CN116012492A (en) | Prompt word intelligent optimization method and system for character generation image | |
| Chi | Research on the application of k-means clustering algorithm in student achievement | |
| CN109543749A (en) | Drawing sentiment analysis method based on deep learning | |
| Zhou et al. | Evaluation and design method for product form aesthetics based on deep learning | |
| CN109471930B (en) | Emotional board interface design method for user emotion | |
| CN113886562A (en) | AI resume screening method, system, equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |