TWI771250B - Device and method for reducing data dimension, and operating method of device for converting data dimension - Google Patents

Abstract

The present disclosure provides a method for reducing a dimension of data, including the following steps: providing a base data database including a plurality of base data, wherein each of the base data has a first dimension; operating each of the base data according to a first weighting function , to form a dimension-reduced base data having a second dimension; operating each dimension-reduced base data according to a second weight function to restore each dimension-reduced base data to a restoration base data having the first dimension; adjusting the first and the second weighting functions according to an error between the restored base data and the base data, wherein in the base data database: each of the base data corresponds to a space vector; and the pluralty of space vectors are all orthogonal to each other.

Description

Apparatus and method for reducing data dimension, operation method of apparatus for converting data dimension

The present invention relates to a method and device for reducing data dimension, and more particularly, to a method and device for reducing data dimension that are independent of data.

As network cloud services become more and more prosperous, the data that needs to be transmitted and stored becomes larger and more complex, and the number of data centers increases. The market demand for data compression should be higher and higher. Data dimensionality reduction can be applied to data storage of cloud services. In many countries with a large number of data centers, there is a great demand for a large amount of data compression storage. Data dimensionality reduction can be applied to video streaming services. Video streaming services in various countries are very prosperous, and there is a high demand for compressed transmission of a large amount of streaming data.

In Patent Publication No. (CN105678265A), it discloses a manifold-based data dimensionality reduction method. Y_i is obtained according to the formula Y_i=W_i^T X_i, wherein Y_i is the low-dimensional representation of X_i, and W_i is obtained by the maximum distance criterion and The locality preserving projection algorithm is derived. The disadvantage of this method is that if the original dimension of the data is higher or there are more data sets, it will take more time to calculate the W_i matrix, and it can only be applied to the data in the space generated by the data set, and the versatility is low. .

In the prior art document 1 (Nonlinear Dimensionality Reduction by Locally Linear Embedding.), Locally Linear Embedding (LLE), which discloses a nonlinear dimensionality reduction method, LLE finds a low-dimensional projection of data, and its projection preserves the data The distance of a point from its local neighbor data point. The disadvantage of LLE is that the optimal value of the required number of adjacent data points K needs to be manually adjusted or obtained by cross-validation, which is time-consuming. It can only be applied to data in the space created by the data set.

In the prior art document 2 (Reducing the Dimensionality of Data with Neural Networks), it is disclosed to use a neural network-based autoencoder to reduce the data dimension. Its main disadvantage is that the effective target needs to be in the space generated by the training data set, and it can usually only act on the same type of data as the data set. resulting in a doubling of training time.

In view of the above limitations, the present disclosure proposes a data dimensionality reduction method based on a spatial basis. The space is usually a vector space, all data points have coordinates or coefficients in the space, and the basis can expand the entire vector space. Data points can be assembled from these smaller bases. This disclosure transforms the base into a training dataset for training neural network-based autoencoders. The trained autoencoder can be universally applied to all data in the space expanded by the aforementioned base. In this way, the dimensionality reduction model can be trained with a smaller amount of data, but it has a high degree of versatility. It only needs to train the model once and can be applied to most usage scenarios, saving a lot of training time.

The neural network-based data dimensionality reduction technology proposed in this disclosure can be used in these data centers or in consumer electronics to accelerate neural network computing chips to improve performance, making this disclosure more competitive than traditional technologies. .

Based on the above concept, the present disclosure provides a method for reducing the dimension of image data, including the following steps: generating a base image data matrix, wherein: the base image data matrix includes N × N base image data; each base image data has N × N dimensions; and applying a specific coefficient to each of the N × N base image data and combining them, it is sufficient to represent a to-be-processed image data. Each of the base image data is operated with a first weight function, and a residual down-sampling is performed on the base image data to form a reduced-dimensional base image data having a dimension smaller than N × N . Residual upsampling is performed on the reduced-dimension base image data with a second weight function to restore the reduced-dimension base image data into a restored base image data of N × N dimensions, wherein the restored base image data and the base There is an error between the image data. The first and the second weighting functions are adjusted according to the error.

Based on the above concept, the present disclosure provides an operation method of an apparatus for converting data dimensions, wherein the apparatus includes a first feature extraction unit and a second feature extraction unit, and the method includes the following steps: providing a base data database, A plurality of base data are included, wherein each of the base data has a first dimension. Each of the base data is operated according to a first weight function to form a dimension-reduced base data having a second dimension. operating each of the dimension-reduced base data according to a second weight function to restore each of the dimension-reduced base data into a restored base data with the first dimension, wherein: there is an error between the restored base data and the base data; and adjusting the first and second weight functions according to the error, wherein in the base database: each of the base data corresponds to a space vector; and the complex space vectors are mutually orthogonal to each other.

Based on the above concept, the present disclosure provides an apparatus for reducing data dimension, including a base data database and an artificial intelligence module. The base data database includes a plurality of base data orthogonal to each other, and each base data has a first dimension. The artificial intelligence module includes an encoding unit and a decoding unit, wherein: the encoding unit performs residual down-sampling on each of the base data according to a first weighting function to form a dimension-reduced dimension with a second dimension Basic information. The decoding unit performs residual up-sampling on the reduced-dimension base data according to a second weight function, so as to restore the reduced-dimension base data into a restored base data with the first dimension, wherein the restored base data and the base data There is an error therebetween, and the artificial intelligence module adjusts the first and the second weight function according to the error.

Based on the above concept, the present disclosure provides a method for reducing the dimension of data, including the following steps: providing a base data database, including a plurality of base data, wherein each base data has a first dimension; operating according to a first weight function each of the base data to form a dimension-reduced base data with a second dimension; operate each of the dimension-reduced base data according to a second weight function to restore each of the dimension-reduced base data to have the first dimension A restored base data of a dimension; the first and the second weighting functions are adjusted according to an error between the restored base data and the base data, wherein in the base data database: each base data corresponds to a space vector; and the complex space vectors are mutually orthogonal to each other.

The present disclosure uses a neural network model, which is non-linear dimensionality reduction, can have lower reconstruction errors, and can be used for any type of image data.

The method and device for reducing the dimension of the data disclosed in the present disclosure can be applied to general data, such as different types of databases, and especially can be applied to signal analysis with more complex raw data, such as spectral images, computed tomography and other biomedical images. The disclosed neural network model can converge to a very low reconstruction error with little manual tuning of parameters, and can be used on any dataset.

The present disclosure uses the spatial base as the training data, and only needs a small amount of data to obtain a model with good generality, which greatly reduces the training time.

For further description and advantages of the present disclosure, reference may be made to the subsequent drawings and embodiments for a clearer understanding of the technical solutions of the present disclosure.

10: A device for reducing the dimension of data

101: Base Database

101D: Complex base image data

101D11, 101D12, 101D1N, 101D21, 101DN1, 101DN2, 101DNN: base image data

102: Artificial Intelligence Modules

102A: Autoencoder Models

102AD: Decoding unit

102AE: Coding Unit

102ADC: Deconvolution calculation unit

102AEC: Convolution Computation Unit

102ADR: second activation unit

102AER: the first activation unit

D1: Characteristic data

D2: Characteristic data processed by activation unit

D1', D2': restore feature data

102AO: Optimization Layer

102CNN: Convolutional Neural Networks

Figure 1: A schematic diagram of multiple base image data according to a preferred embodiment of the present disclosure.

FIG. 2 is a schematic diagram of an apparatus for reducing data dimension according to a preferred embodiment of the present disclosure.

FIG. 3A: The distribution diagram of the test image data in the low-dimensional space according to the preferred embodiment of the present disclosure.

FIG. 3B is an enlarged distribution diagram of the image data to be processed in the low-dimensional space according to the preferred embodiment of the present disclosure.

FIG. 4 is a schematic diagram of a reconstructed image sample according to a preferred embodiment of the present disclosure.

FIG. 5 is a schematic diagram of a method for reducing the dimension of image data according to a preferred embodiment of the present disclosure.

FIG. 6 is a schematic diagram of an operation method of an apparatus for converting data dimensions according to a preferred embodiment of the present disclosure.

FIG. 7 is a schematic diagram of a method for reducing data dimension according to another preferred embodiment of the present disclosure.

Please read the following detailed description with reference to the accompanying drawings of the present disclosure, which are by way of example to introduce various embodiments of the present disclosure and to provide an understanding of how to implement the present disclosure. The embodiments of the present disclosure provide sufficient content for those skilled in the art to implement the disclosed embodiments of the present disclosure, or to implement the embodiments derived from the disclosed contents of the present disclosure. It should be noted that these embodiments are not mutually exclusive, and some embodiments can be appropriately combined with one or more other embodiments to form new embodiments, that is, the implementation of the present disclosure is not limited to Examples disclosed below. In addition, for the sake of brevity and clarity, the relevant details are not excessively disclosed in each embodiment, and even if specific details are disclosed, they are only exemplified to make the readers understand, and the relevant specific details in the various embodiments are not intended to be limiting. disclosure of the case.

Please refer to FIG. 1 , which is a schematic diagram of a plurality of base image data 101D according to a preferred embodiment of the disclosure. The complex base image data 101D includes base image data 102D11, 101D12, . . . 101D1N, 101D21, . . . 101D2N, . . . 101DNN . Please refer to FIG. 2 , which is a schematic diagram of an apparatus 10 for reducing data dimension according to a preferred embodiment of the present disclosure, which includes a base data database 101 as shown in FIG. 1 and an artificial intelligence module 102 as shown in FIG. 2 . The intelligent module 102 is, for example, an auto-encoder model 102A based on a physical substrate. The artificial intelligence module 102 includes an encoding unit 102AE and a decoding unit 102AD. Taking the dimensionality reduction of grayscale image data as an example, the goal is to reduce the original data dimension to 32×32=1024 and reduce it to 2 dimensions, and this 2-dimensional data can be restored to the original 1024 dimensions by the artificial intelligence module 102 data of. The data dimensionality reduction of the autoencoder model 102A based on the physical base mainly includes two parts, one is the generation of the complex base image data 101D for training, and the other is the training of the autoencoder model 102A based on the neural network 102CNN.

In FIG. 1, for the dimensionality reduction of gray-scale image data, in one embodiment, the complex base image data 101D of the discrete cosine transform (DCT) is used as the base image data for training. For any piece of image data F of N × N size or resolution to be processed, it can be expressed as a linear combination of complex base image data 101D

，其中s _u 及s _v 均係具有1×N維度的向量並各自對應於N×N維度的離散餘弦轉換(DCT)矩陣之列向量的其中之一，1≦u≦N且1≦v≦N，且該N×N個S _u,v 分別代表不同空間頻率的基底函數。 T ( u,v ) is the coefficient applied to Su _,v , each of the complex base image data 101D is defined as a base image matrix

, where both s _u and s _v are vectors with 1× N dimensions and each corresponds to one of the column vectors of an N × N -dimensional discrete cosine transform (DCT) matrix, 1≦ u ≦ N and 1≦ v ≦ N , and the N × N S _u,v respectively represent basis functions of different spatial frequencies.

的編碼以 及解碼功能，其中編碼單元102AE執行函數Encoder(x)=z的編碼功能，解碼單元102AD執行函數

的解碼功能。自動編碼模型102A的輸入訓練影像資料為x，輸出重建影像資料為

，其中經過編碼後的影像資料code z _k 為一2維的向量，即為降完維度後的影像資料。訓練自動編碼模型102A的方法為最小化模型的重建誤差

，其中M=1024，代表訓練資料的數量。當自動編碼模型102A訓練完成後，則可輸入待處理影像資料F至編碼單元102AE以執行函數Encoder(x)的編碼功能來降低待處理影像資料F的維度，以產生低維度的經過編碼後的影像資料code z。當需要重建原始的資訊時，只需要將code z輸入至解碼單元102AD以執行函數Decoder(z)的解碼功能，即可重建高維度的資訊。 In Figure 2, the auto-encoding model 102A performs the function

The encoding and decoding function, wherein the encoding unit 102AE performs the encoding function of the function Encoder ( x )= z , and the decoding unit 102AD performs the function

decoding function. The input training image data of the auto-encoding model 102A is x , and the output reconstructed image data is

, wherein the encoded image data code z _k is a 2-dimensional vector, that is, the image data after dimension reduction. The method for training the auto-encoding model 102A is to minimize the reconstruction error of the model

, where M = 1024, representing the number of training data. After the training of the automatic coding model 102A is completed, the to-be-processed image data F can be input to the encoding unit 102AE to execute the encoding function of the function Encoder ( x ) to reduce the dimension of the to-be-processed image data F, so as to generate a low-dimensional encoded image Image data code z . When the original information needs to be reconstructed, it is only necessary to input the code z to the decoding unit 102AD to execute the decoding function of the function Decoder ( z ), so that the high-dimensional information can be reconstructed.

In Figure 2, in addition to image base data, any type of base data can also be used as input base data in the training phase, such as audio base data, and they can come from a cloud disk, an audio-video streaming service, At least one of a computed tomography scan, a video compression and storage, an image preprocessing, and a data mining. The base data may be base vector data, which is a space base data and forms a vector space, all data points have coordinates or coefficients in the vector space, the space base data is expanded to form the vector space, and All data points in the vector space are formed from complex basis vector data to convert each of the basis data into a training data set. The complex base image data 101D in FIG. 1 forms a DCT matrix, and can also form a Hadmard matrix.

The basis data database 101 may include complex basis data that are orthogonal to each other, such as complex basis data 101D formed by applying T ( u,v ) as coefficients that s _u and s _v are orthogonal to each other, and each of the basis The data has a first dimension, such as 32×32×1 in Figure 2, which represents a base data with a dimension of 32×32, and each base image data 102D11, 101D12,...101D1N, 101D21,... The order of .101D2N,...101DNN is a random order. The encoding unit 102AE performs residual down sampling on each of the base data according to a first weight function, so as to form a dimension-reduced base data code z with a second dimension, such as in FIG. 2 1×1×2, which represents the two base data with reduced dimension of 1×1. The decoding unit 102AD performs residual upsampling on the reduced-dimensional base data codez according to a second weight function, so as to restore the reduced-dimensional base data codez into a restored base image data 101D11 with the first dimension ', wherein there is an error err between the restored base image data 101D11' and the base image data 101D11, and the artificial intelligence module 102 adjusts the first and second weight functions according to the error err.

In FIG. 2, the coding unit in the convolutional neural network 102CNN can be a residual down-sampling block, including a convolution calculation unit 102AEC, a first activation unit 102AER, and a fully connected layer (not shown) . The convolution calculation unit 102AEC extracts a characteristic data D1 for each base data 101D according to the first weight function. The first activation unit 102AER processes the characteristic data D1 according to a first nonlinear combination rule ReLU. Each base data 101D may be processed by the convolution calculation unit 102AEC and the first activation unit 102AER in two or more layers, and then input to the fully connected layer. The fully connected layer converts the feature data D2 processed by the activation unit 102AER into the reduced dimension base data code z. The decoding unit 102AD can be a residual up-sampling block in the convolutional neural network 102CNN, including a deconvolution calculation unit 102ADC, a second activation unit 102ADR, and an optimization layer 102AO. The deconvolution calculation unit 102ADC restores each reduced dimension data code z into a restored feature data D1 ′ according to the second weight function. The second activation unit 102ADR processes the restored characteristic data D1' according to a second nonlinear combination rule ReLU' to form a restored characteristic data D2'. The dimension-reduced base data code z may be processed by the deconvolution calculation unit 102ADC and the second activation unit 102ADR in two or more layers, and then input to the optimization layer 102AO. The optimization layer 102AO minimizes the error err to reconstruct the restored feature data D2' as the restored base data 101D11'. The optimization layer 102AO may be a residual block, which may have the same structure as the residual downsampling.

Please refer to FIG. 3A , which is a distribution diagram of test image data in a low-dimensional space according to a preferred embodiment of the present disclosure. FIG. 3B is an enlarged distribution diagram of the image data to be processed in the low-dimensional space according to the preferred embodiment of the disclosure. The horizontal axis represents the dimension 1 of the image data, and the vertical axis represents the dimension 2 of the image data. The circular points represent natural images, the triangular points represent handwritten digital images, and the square points represent clothing images. From Figure 3A and Figure 3B, it can be seen that the image data contains different types of data sets, the same type of data set is represented by the same shape point, and the same shape point scattered in different positions represents further classification in the same type material. After the training of the auto-encoding model 102AE is completed, any image can be input to the encoding unit 102AE to perform the encoding function of the function Encoder ( x ) to reduce the dimension and generate the low-dimensional base image data code z. When the original information needs to be reconstructed, it is only necessary to input the low-dimensional base image data code z into the decoding unit 102AD to execute the decoding function of the function Decoder ( z ), and then the high-dimensional information can be reconstructed. In this example, the automatic coding model 102A 10 epochs are trained. Using different types of test sets of natural images, handwritten digital images, and clothing images as test data, Figure 3 shows the distribution of some images in low-dimensional space. From Figures 3A and 3B, it can be seen that any image in space is Independent coordinate points, indicating that the entire high-dimensional space is converted into a low-dimensional space.

Please refer to FIG. 4 , which is a schematic diagram of a reconstructed image sample according to a preferred embodiment of the present disclosure. From the mean square error (MSE) between the test image data and the reconstructed image samples, the range of the mean square error is in the range of 3.8*10 ^-5 ~10.8*10 ^-5 , showing a good reconstructed image error.

Please refer to FIG. 5 , which is a schematic diagram of a method S10 for reducing the dimension of image data according to a preferred embodiment of the present disclosure. The method S10 includes the following steps: Step S101 , generating a base image data matrix, wherein: the base image data matrix includes: N × N base image data; each of the base image data has N × N dimensions; and each of the N × N base image data is applied with a specific coefficient and combined, which is sufficient to represent a to-be-processed image data. Step S102 , operating each of the base image data with a first weight function, and performing a residual down-sampling on the base image data to form a reduced-dimensional base image data with a dimension smaller than N × N . Step S103, performing residual upsampling on the reduced-dimension base image data with a second weight function, so as to restore the reduced-dimension base image data into a restored base image data of N × N dimensions, wherein the restored base image data There is an error with the base image data. Step S104, adjusting the first and the second weighting functions according to the error.

，其中s _u 及s _v 均係具有1×N維度的向量並各自對應於N×N維度的離散餘弦轉換(DCT)矩陣之列向量的其中之一，1≦u≦N且1≦v≦N，且該N×N個S _u,v 分別代表不同空間頻率的基底函數。該待處理影像資料係表示為

，

，其中T(u,v)為施予S _u,v 之係數。該方法更包括以調整後之該第一權重函數操作該待處理影像資料，且對該待處理影像資料進行一殘差下取樣，以形成具有一維度小於N×N的一降維度影像資料。 In any embodiment of the present disclosure, each of the base image data is defined as a base image matrix

, where both s _u and s _v are vectors with 1× N dimensions and each corresponds to one of the column vectors of an N × N -dimensional discrete cosine transform (DCT) matrix, 1≦ u ≦ N and 1≦ v ≦ N , and the N × N S _u,v respectively represent basis functions of different spatial frequencies. The to-be-processed image data is represented as

,

, where T ( u , v ) is the coefficient applied to S _u,v . The method further includes operating the to-be-processed image data with the adjusted first weight function, and performing a residual downsampling on the to-be-processed image data to form a reduced-dimensional image data having a dimension smaller than N × N .

In any embodiment of the present disclosure, the step of operating each of the base image data with the first weight function operates on each of the base image data in a random order. The N DCT matrix column vectors are obtained by sampling the values of the cosine function at different frequencies.

Please refer to FIG. 6 , which is a schematic diagram of an operation method S20 of an apparatus for converting data dimensions according to a preferred embodiment, wherein the apparatus includes a first feature extraction unit and a second feature extraction unit, and the method S20 includes the following steps: Step S201 , providing a base data database, including a plurality of base data, wherein each base data has a first dimension. Step S202 , operating each of the base data according to a first weight function to form a reduced-dimensional base data having a second dimension with a reduced dimension. Step S203, operate each of the reduced-dimensional base data according to a second weight function, and restore each of the reduced-dimensional base data into a restored base data with the first dimension, wherein: the restored base data and the base data have an error; and adjusting the first and second weighting functions according to the error, wherein in the basis database: each of the basis data corresponds to a space vector; and the complex space vectors are mutually orthogonal to each other.

表示，其中，S _u,v 為基底影像矩陣，

，s _u 與s _v 各自對應於離散傅立葉(DCT)矩陣之列向量的其中之一，該N×N個S _u,v 分別代表具有不同空間頻率的基底函數，T(u,v)施予S _u,v 之係數，1≦u≦N且1≦v≦N；以調整後之該第一權重函數操作該待處理資料，以形成維度等於該第二維度的一降維度資料；以及以調整後之該第二權重函數操作該降維度資料，以將該降維度資料還原成一還原資料。 In any embodiment of the present disclosure, the method for reducing or converting data dimensions further includes: inputting data to be processed, the data to be processed is

represents, where S _u,v is the base image matrix,

, s _u and s _v each correspond to one of the column vectors of the discrete Fourier transform (DCT) matrix, the N × N S _u,v respectively represent basis functions with different spatial frequencies, T ( u,v ) gives The coefficients of S _{u, v} , 1≦ u ≦ N and 1≦ v ≦ N ; operate the data to be processed with the adjusted first weight function to form a dimension-reduced data with a dimension equal to the second dimension; and The adjusted second weight function operates on the dimension-reduced data to restore the dimension-reduced data into restored data.

In any embodiment of the present disclosure, the first weight function includes L layers of first sub-weight functions, wherein L ≧1, each of the first sub-weight functions corresponds to a first input data, and the first layer of The first input data of the first sub-weight function is each of the base data; the second weight function includes L layers of second sub-weight functions, where L ≦1, each of the second sub-weight functions corresponds to a second input data, and the second input data of the second sub-weight function of the first layer is each of the dimension-reduced basis data. The step of operating each of the base data according to the first weighting function includes the following steps: extracting a characteristic data of the input data corresponding to the first sub-weighting function according to the first sub-weighting function; according to a first nonlinear combination rule, Processing the feature data, wherein the feature data is the input data of the first sub-weight function of the next layer, and the feature data corresponding to the first sub-weight function of the Lth layer is processed by the first nonlinear combination rule Overlaid with the base data in a first specific manner to obtain the dimension-reduced base data. The step of operating each of the dimension-reduced base data according to the second weight function includes the following steps: according to the second sub-weight function, restoring the dimension-reduce base data corresponding to the second sub-weight function into a restored feature data; according to a second sub-weight function A nonlinear combination rule for processing the restored feature data, wherein the restored feature data is the second input data of the second sub-weight function of the next layer, and the restored feature data corresponding to the second sub-weight function of the Lth layer After being processed by the second nonlinear combination rule, the reduced-dimensional base data is superimposed in a second specific manner to obtain the restored base data; and the error is minimized to optimize the first and second weight functions .

，其中s _u 及s _v 分別具有1×M及1×N維度維度的向量並分別對應於M×M及N×N維度的正交轉換矩陣之列向量的其中之一，1≦u≦M且1≦v≦N，且該M×N個S _u,v 分別代表不同空間頻率的基底函數，上述 的正交轉換矩陣可以是Hadamard,DFT,DCT,Haar,Slant矩陣。該編碼單元以調整後之該第一權重函數操作該待處理資料，且對該待處理資料進行一殘差下取樣，以形成具有一維度小於M×N的一降維度資料。 In any embodiment of the present disclosure, each of the base data is defined as a base matrix

, where s _u and s _v have 1× M and 1× N -dimensional vectors, respectively, and correspond to one of the column vectors of the M × M and N × N -dimensional orthogonal transformation matrices, respectively, 1≦ u ≦ M And 1≦ v ≦ N , and the M × N S _{u, v} respectively represent basis functions of different spatial frequencies, the above-mentioned orthogonal transformation matrix can be Hadamard, DFT, DCT, Haar, Slant matrix. The encoding unit operates the data to be processed with the adjusted first weight function, and performs a residual downsampling on the data to be processed to form a reduced-dimensional data having a dimension smaller than M × N .

Please refer to FIG. 7 , which is a schematic diagram of a method S30 for reducing data dimension according to another preferred embodiment. The method S30 includes the following steps: Step S301 , providing a base data database, including a plurality of base data, wherein each The base data has a first dimension. Step S302 , operating each of the base data according to a first weight function to form a dimension-reduced base data having a second dimension with a reduced dimension. Step S303 , operating each of the dimension-reduced base data according to a second weight function to restore each of the dimension-reduced base data to a restored base data having the first dimension. Step S304, adjusting the first and the second weighting functions according to an error between the restored base data and the base data, wherein in the base data database: each base data corresponds to a space vector; and the Complex space vectors are all orthogonal to each other.

To sum up, the present disclosure proposes a data dimensionality reduction method based on a spatial basis. The space is usually a vector space, all data points have coordinates or coefficients in the space, and the basis can expand the entire vector space. Data points can be assembled from these smaller bases. This disclosure transforms the base into a training dataset for training neural network-based autoencoders. The trained autoencoder can be universally applied to all data in the space expanded by the aforementioned base. In this way, the dimensionality reduction model can be trained with a smaller amount of data, and at the same time, it has a high degree of versatility. The model only needs to be trained once and can be applied to most usage scenarios, saving a lot of training time. The method for reducing the dimension of image data, the operation method of the apparatus for converting the dimension of the data, the apparatus for reducing the dimension of the data, and the method for reducing the dimension of the data disclosed in the present disclosure can be applied to related fields that need to process complex data, For example, video compression and storage, image preprocessing and data mining, etc. The traditional autoencoder data dimensionality reduction method usually requires a large data set, and the training model takes a lot of time; or it can only be applied to a small range of data types, and when it is used for different types of data, it needs to be retrained every time. model, it will take a lot of time. Significant time savings can be achieved by thus disclosing the proposed method.

Although the present invention is disclosed in the above-mentioned several embodiments or examples, it is not intended to limit the present invention. Anyone who is familiar with this technique can make some changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention should be determined by the scope of the appended patent application.

Steps of S301~S304:S30

S30: Methods for reducing data dimension

Claims (10)

A method for reducing the dimension of image data, comprising the following steps: A matrix of base image data is generated, where: The base image data matrix includes N × N base image data; each of the base image data has N × N dimensions; and After each of the N × N base image data is applied with a specific coefficient and combined, it is sufficient to represent a to-be-processed image data; operating each of the base image data with a first weight function, and performing a residual down-sampling on the base image data to form a reduced-dimensional base image data having a dimension less than N × N ; Residual upsampling is performed on the reduced-dimension base image data with a second weight function to restore the reduced-dimension base image data into a restored base image data of N × N dimensions, wherein the restored base image data and the base There is an error between the image data; and The first and the second weighting functions are adjusted according to the error. The method of claim 1, wherein: Each of the base image data is defined as a base image matrix

, where both s _u and s _v are vectors with 1× N dimensions and each corresponds to one of the column vectors of an N × N -dimensional discrete cosine transform (DCT) matrix, 1≦ u ≦ N and 1≦ v ≦ N , and the N × N S _u,v represent the basis functions of different spatial frequencies respectively; The to-be-processed image data is represented as

,

, where T ( u,v ) is the coefficient applied to S _u,v ; and The method further includes operating the to-be-processed image data with the adjusted first weight function, and performing a residual downsampling on the to-be-processed image data to form a reduced-dimensional image data having a dimension smaller than N × N . The method of claim 1, wherein: the step of operating each of the base image data with the first weight function operates on each of the base image data in a random order; and The N DCT matrix column vectors are obtained by sampling the values of the cosine function at different frequencies. A method for operating an apparatus for converting data dimensions, wherein the apparatus includes a first feature extraction unit and a second feature extraction unit, and the method includes the following steps: providing a base data database, including a plurality of base data, wherein each base data has a first dimension; operating each of the base data according to a first weight function to form a reduced-dimensional base data having a second dimension with a reduced dimension; and Operate each of the dimension-reduced base data according to a second weight function, and restore each of the dimension-reduced base data to a restored base data with the first dimension, wherein: There is an error between the restored base data and the base data; and The first and second weight functions are adjusted according to the error, wherein in the base database: each of the base data corresponds to a space vector; and The complex space vectors are mutually orthogonal to each other. The method according to claim 4, further comprising the following steps: Enter a pending data that starts with

represents, where S _u,v is the base image matrix,

, s _u and S _v each correspond to one of the column vectors of the discrete Fourier transform (DCT) matrix, the N × N S _u,v respectively represent basis functions with different spatial frequencies, T ( u,v ) gives The coefficient of S _u,v , 1≦ u ≦ N and 1≦ v ≦ N ; operating the to-be-processed data with the adjusted first weight function to form a dimension-reduced data with a dimension equal to the second dimension; and The dimension-reduced data is operated with the adjusted second weight function to restore the dimension-reduced data into restored data. A method as claimed in claim 4, wherein: The first weighting function includes L-layer first sub-weighting functions, where L ≧1, each of the first sub-weighting functions corresponds to a first input data, and the first sub-weighting function of the first layer the input data is each of the base data; The second weighting function includes L layers of second sub-weighting functions, where L ≧1, each of the second sub-weighting functions corresponds to a second input data, and the second sub-weighting function of the first layer The input data is the base data of each dimension reduction; The step of operating each of the base data according to the first weight function includes the following steps: extracting a characteristic data of the input data corresponding to the first sub-weight function according to the first sub-weight function; According to a first nonlinear combination rule, the feature data is processed, wherein the feature data is the input data of the first sub-weight function of the next layer, and the feature data corresponding to the first sub-weight function of the Lth layer is processed by The first nonlinear combination rule is processed and superimposed with the base data in a first specific manner to obtain the dimension-reduced base data; and The step of operating each of the dimension-reduced base data according to the second weight function includes the following steps: According to the second sub-weight function, the reduced dimension base data corresponding to the second sub-weight function is restored to a restored feature data; processing the restored feature data according to a second nonlinear combination rule, wherein the restored feature data The data is the second input data of the second sub-weight function of the next layer, and the restored feature data corresponding to the second sub-weight function of the L-th layer is processed by the second nonlinear combination rule and the dimension reduction basis The data are superimposed in a second specific manner to obtain the restored base data; and The error is minimized to optimize the first and second weighting functions. A device for reducing the dimension of data, comprising: a base data database including a plurality of base data orthogonal to each other, and each of the base data has a first dimension; and An artificial intelligence module, including an encoding unit and a decoding unit, wherein: The encoding unit performs residual down-sampling on each of the base data according to a first weight function to form a reduced-dimensional base data having a second dimension with a reduced dimension; and The decoding unit performs residual up-sampling on the reduced-dimension base data according to a second weight function, so as to restore the reduced-dimension base data into a restored base data with the first dimension, wherein the restored base data and the base data There is an error therebetween, and the artificial intelligence module adjusts the first and the second weight function according to the error. The apparatus of claim 7, wherein: The coding unit contains: a convolution computing unit, extracting a characteristic data from each base data according to the first weight function; a first activation unit for processing the feature data according to a first nonlinear combination rule; and a fully-connected layer that converts the feature data processed by the activation unit into the dimension-reduced base data; and The decoding unit contains: a deconvolution computing unit, according to the second weight function, to restore each dimension-reduced data to a first restored feature data; a second activation unit for processing the first restored characteristic data according to a second nonlinear combination rule to form a second restored characteristic data; and An optimization layer minimizes the error to reconstruct the second restored feature data as the restored base data. The apparatus of claim 7, wherein: Each of the base data is defined as a base matrix

, where s _u and s _v have 1× M and 1× N -dimensional vectors, respectively, and correspond to one of the column vectors of the M × M and N × N -dimensional orthogonal transformation matrices, respectively, 1≦ u ≦ M and 1≦ v ≦ N , and the M × N S _u,v represent basis functions of different spatial frequencies respectively; and The encoding unit operates the data to be processed with the adjusted first weight function, and performs a residual downsampling on the data to be processed to form a reduced-dimensional data having a dimension smaller than M × N . A method for reducing the dimensionality of data, including the following steps: providing a base data database, including a plurality of base data, wherein each base data has a first dimension; operating each of the base data according to a first weight function to form a dimension-reduced base data having a second dimension; operating each of the dimension-reduced base data according to a second weight function to restore each of the dimension-reduced base data to a restored base data with the first dimension; The first and the second weighting functions are adjusted according to an error between the restored base data and the base data, wherein in the base data database: each of the base data corresponds to a space vector; and The complex space vectors are mutually orthogonal to each other.

Images