TWM567900U - Data reduction and computer system for establishing data identification model - Google Patents

Abstract

A computer system for data reduction and data identification model, a combination generation module generates rules according to a rule of selecting Y(1≦Y≦X-1) algorithms from X(X≧2) kinds of dithering algorithms A combination of Z algorithms, wherein the data reduction module separately performs a dithering process on an image data represented by a bit in the Z algorithm, and converts the image data into Z types in b bits. Representing the reduced image data of the pixel, wherein 1≦b≦a-1, and a model building module respectively trains a deep neural network with the Z-reduced image data to correspondingly generate Z training depths. The neural network model, the training result, and an error between the training result and a desired result, and selecting the deep neural network model that generates the smallest error among the Z errors as a data recognition model.

Description

Data reduction and computer system for establishing data identification model

The present invention relates to a deep neural network training system, and more particularly to a computer system for applying a deep neural network to establish a data recognition model.

Artificial intelligence (AI) is one of the technologies that has been in full swing in recent years. For example, image recognition and data analysis used by autonomous vehicles or various types of robots require AI technology. In the development of AI technology, Deep Neural Network (DNN) has always played an important role. Deep neural network is a deep learning method in machine learning. It simulates the mathematical model of the biological nervous system and continuously provides a large amount of data for multiple calculations and trainings of different classes and architectures to find the best. The most effective and deep learning model. Therefore, providing high-quality and large amounts of data to perform computational training on deep neural networks can help improve the training accuracy of deep neural networks.

However, a large amount of data means that deep neural networks require a large amount of computation and generate relatively large parameters, and these parameters will occupy a large amount of memory space, making the resulting deep neural network model not easy to apply to the memory. In electronic devices or electronic devices where space is limited. Therefore, it is necessary to moderately reduce the parameters in the deep neural network model to reduce the memory space occupancy in practical applications.

Therefore, the purpose of the present invention is to provide a computer system for data reduction and data identification models.

Thus, the computer system for reducing the amount of data and establishing a data identification model includes a storage device and a processing device, the storage device storing an image data and an X (X≧2) type of color-shifting algorithm, the processing device and the storage device The device is electrically coupled to access the image data and the dithering algorithms, and includes a combination generation module, a data reduction module and a model creation module; wherein the combination generation module is based on X (X) ≧ 2) Selecting the rules of Y(1≦Y≦X-1) algorithms in the dithering algorithm to generate Z algorithm combinations; the data decrement module combines the Z algorithms into a The bit indicates that the image data of the pixel is dithered, and the image data is converted into Z kinds of reduced image data whose pixels are represented by b bits, wherein 1≦b≦a-1; the model Establishing a module to prepare a deep neural network, and training the deep neural network with the Z-reduced image data to generate a deep neural network model and a training result after training for each reduced image data. And the results of the training and a desired result Error, and selecting the combination of the algorithms corresponding to the minimum error of the Z errors as a filter module for data decrement, and selecting the deep neural network model that generates the smallest error among the Z errors as one Data identification model.

In some implementations of the present invention, Y=2, and the data reduction module performs color dithering on the image data in one of each algorithm combination, and converts the pixel data of the image data. a first converted image data represented by m bits, and the data reduction module performs color dithering on the image data by another algorithm in each algorithm combination, and the image data is The pixel is converted into a second converted image data represented by n bits, and the first converted image data and the second converted image data constitute the reduced image data, m+n=b.

In some implementations of the present invention, the step of the model building module training the deep neural network with each of the converted image data comprises: (C1) inputting the reduced image data into the deep neural network to output The training result; (C2) comparing the training result with the expected result to generate the error; (C3) inputting the error into the deep neural network; and (C4) repeating steps (C1) to (C3) until the If the error does not change, the trained deep neural network model and the error are output.

In some embodiments of the present invention, the storage device and the processing device are integrated into a computer device, the storage device being a storage unit of the computer device, and the processing device is a processing unit of the computer device.

In some embodiments of the present invention, the storage device and the processing device are electrically coupled to each other via a wired or wireless network to communicate with each other.

The utility model has the following advantages: the data reduction module generates corresponding Z-reduced image data according to the original image data and the Z algorithm combination, and the model is used to create the module to reduce the Z image. The data is trained on the deep neural network, and correspondingly, the Z-trained deep neural network model and its error are generated, and the deep neural network model with the smallest error is selected as the data recognition model, thereby The effect of decrementing the parameters in the data recognition model is achieved.

Before the present invention is described in detail, it should be noted that in the following description, similar elements are denoted by the same reference numerals.

Referring to FIG. 1, an embodiment of a computer system for reducing data and establishing a data identification model includes a method for storing an image data and storing X(X≧2) dithering algorithms. The device 1 and the processing device 2 are electrically coupled to the storage device 1 to access the image data and the dithering algorithm, and include a combination generation module 21 and a data reduction module 22 And a model building module 23, and in the embodiment, the three modules are implemented by means of program software, but not limited thereto, the three modules can also be combined with firmware or software and hardware. The way to achieve.

Therefore, in this embodiment, after the program including the combination generation module 21, the data reduction module 22, and the model creation module 23 is executed by the processing device 2, the method flow shown in FIG. 2 is completed. The image data refers to a certain amount of original image data (or raw data). The storage device 1 and the processing device 2 can communicate via a wired or wireless network to transmit data. Alternatively, the storage device 1 and the processing device 2 can be integrated into the same computer device. A storage unit of the device, the processing device 2 being a processing unit of the computer device, such as a central processing unit.

Therefore, first, as shown in step S1 of FIG. 2, the combination generation module 21 generates Z according to the rule of selecting Y(1≦Y≦X-1) algorithms from X(X≧2) kinds of dithering algorithms. The algorithm combination is provided, and the algorithm combination is provided to the data reduction module 22; with X=9, that is, 9 kinds of dithering algorithms, Y=2, that is, 2 out of 9 dithering algorithms are selected. For example, the combination generation module 21 can generate Z=36, that is, 36 algorithm combinations according to the above rules. In this embodiment, the nine color-shifting algorithms may be, for example, "Floyd-Steinberg", "Jarvis, Judice & Ninke", "Stucki", "Burkes", "Sierra", "Two-row Sierra", "Sierra Lite", "Atkinson", "Gradient-based", etc., but not limited to this.

Then, in step S2 of FIG. 2, the data reduction module 22 performs a color dithering process on the image data whose pixels are represented by a (a ≧ 8) bits by using the Z algorithm combinations. The image data is converted into Z kinds of deduced image data whose pixels are represented by b bits, where 1≦b≦a-1; specifically, as shown in FIG. 3, it is assumed that each algorithm combination is composed of two types. (ie, Y=2) is composed of algorithms f1 and f2, and the data decrement module 22 performs dither processing on the image data D by one of the algorithms f1 of the algorithm combination, and the image data D is The pixel is converted into a first converted image data D1 represented by m bits, and the data reduction module 22 performs color dithering on the image data D by another algorithm f2 combined by the algorithm. The pixel of the image data D is converted into a second converted image data D2 represented by n bits, and the first converted image data D1 and the second converted image data D2 constitute the reduced image. Data D', and m+n=b. Of course, the above Y may also be other positive integers greater than 2 but less than X.

For example, suppose the pixel of the image data is represented by a=24, that is, 24-bit, and the algorithm f1 converts the pixel of the image data D into the first converted image data represented by m=1 bits. D1, and the algorithm f2 converts the pixel of the image data D into the second converted image data D2 represented by n=1 bits, so the pixel of the reduced image data D′ is b=2 Bit representation. Thereby, the image data D originally represented by the 24-bit representation pixel can be reduced to the reduced-length image data D' in which the pixels are represented by 2 bits. Of course, the above b, m, n may be other A positive integer, as long as m + n = b and b ≦ a-1 is the effect of data reduction. Therefore, after the image data D is processed by the data reduction module 22, Z types of decremented image data D' representing the pixels in b bits corresponding to the combination of the Z algorithms are generated.

Then, in step S3 of FIG. 2, the model building module 23 trains the prepared deep neural network 24 with the Z-reduced image data D', so that each of the reduced image data D' is generated after training. a deep neural network model and a training result (eg, a recognition rate), and an error between the training result and a desired result (eg, a desired recognition rate); specifically, as shown in step S31 of FIG. 4, The model building module 23 inputs each of the reduced image data D' into the deep neural network 24, and the deep neural network 24 trains the reduced image data D' to output the training result; then, as shown in FIG. Step S32, the model establishing module 23 compares the training result with the expected result to generate the error; then, in step S33, the model establishing module 23 determines whether the current error is the first error (ie, the first The generated error)? If yes, proceed to step S34, input the error feedback to the deep neural network 24, and repeat the above steps S31 and S2, and if not (that is, at least the second error has been generated), proceed to step S35. Judging Is the secondary error equal to the previous error? If not, indicating that the training result of the deep neural network 24 has not stabilized yet, step S34 is performed, the error is fed back into the deep neural network 24, and the above steps S31 to S35 are repeated. Until the error no longer changes, indicating that the training result of the deep neural network 24 has stabilized, the trained deep neural network model and the error are output and stored in the storage device 1. Thereby, the model building module 23 generates Z trained deep neural network models and training results corresponding to the Z reduced image data D′, and errors of the Z training results and the expected results.

Then, in step S4 of FIG. 2, the model building module 23 selects, according to the Z errors, a combination of the algorithms corresponding to the minimum error as a filter module for data decrement, and selects and The deep neural network model corresponding to the minimum error is used as a data recognition model. And the algorithm in the combination of the algorithms corresponding to the minimum error (ie, the filter module) can retain more feature details in the image data, and the algorithms have a better complementary relationship with each other. The combination allows the reduced image data D' to retain the most feature details (ie, the best sensitivity to the image data) in all combinations, so that it can correspond to the post-training deep neural network model (ie The data recognition model has a minimum error and a relatively high recognition accuracy. Therefore, the filter module and the data identification model obtained by the embodiment can be applied to a subsequent model compression program and related applications.

In summary, the above embodiment uses the data reduction module 22 to generate Z-weighted image data D′ according to the original image data D and the Z algorithm combination, and the model building module 23 uses the Z species. After the reduction, the image data D' trains the deep neural network, and correspondingly generates the Z-trained deep neural network model and its error, and selects the deep neural network model with the smallest error as the data recognition model. Thereby, the effect of reducing the parameters in the data identification model is achieved, and the efficacy and purpose of the novel are indeed achieved.

However, the above is only the embodiment of the present invention, and when it is not possible to limit the scope of the present invention, all the simple equivalent changes and modifications according to the scope of the patent application and the contents of the patent specification are still This new patent covers the scope.

1‧‧‧Storage device

2‧‧‧Processing device

21‧‧‧Combined production module

22‧‧‧ Data Reduction Module

23‧‧‧Model building module

24‧‧‧Deep neural network

D‧‧‧Image data

D'‧‧‧Reduced image data

F1, f2‧‧‧ algorithm

D1‧‧‧First converted image data

D2‧‧‧Second converted image data

S1~S4‧‧‧ steps

S31~S36‧‧‧Steps

Other features and effects of the present invention will be clearly shown in the embodiments with reference to the drawings, wherein: FIG. 1 is a main device and a module block of an embodiment of a computer system for reducing data and establishing a data identification model of the present invention. Figure 2 is a main flow chart of the embodiment; Figure 3 is a schematic diagram showing the function of the data reduction module and the model building module of the embodiment; and Figure 4 is the model building mode of the embodiment. A flowchart of the training of the deep neural network.

Claims (5)

A computer system for reducing data and establishing a data identification model includes: a storage device for storing an image data and an X (X≧2) type of dithering algorithm; and a processing device electrically coupled to the storage device for storing Taking the image data and the dithering algorithms, and comprising a combination generation module, a data reduction module and a model creation module; wherein the combination generation module is based on the X (X≧2) dithering calculation In the method, the rules of Y(1≦Y≦X-1) algorithms are selected to generate Z algorithm combinations; the data reduction module combines the Z algorithms to represent the pixels in a bit. The image data is subjected to dithering processing, and the image data is converted into Z kinds of decremented image data whose pixels are represented by b bits, wherein 1≦b≦a-1; the model building module prepares a deep nerve Networking, and training the deep neural network with the Z-reduced image data to generate a post-training deep neural network model and a training result for each type of reduced image data, and the training result and an expectation An error in the result, and the Z algorithm is selected The combination of the algorithm corresponding to the minimum error of the Z errors is used as a filter module for data decrement, and the deep neural network model that generates the smallest error among the Z errors is selected as a data identification. model. A computer system for reducing data and establishing a data identification model as claimed in claim 1, wherein Y=2, and the data reduction module is combined in each algorithm One of the algorithms performs color dithering on the image data, and converts the pixels of the image data into a first converted image data represented by m bits, and the data reduction module is used in each algorithm. Another algorithm in the combination performs color dithering on the image data, and converts the pixels of the image data into a second converted image data represented by n bits, and the first converted image data And the second converted image data constitutes the reduced image data, m+n=b. The computer system for reducing data and creating a data identification model according to claim 1 or 2, wherein the step of the model building module training the deep neural network with each converted image data comprises: (C1) reducing the image after the reduction Data is input to the deep neural network to output the training result; (C2) the training result is compared with the expected result to generate the error; (C3) the error is input to the deep neural network; and (C4) is repeated Steps (C1) to (C3) output the trained deep neural network model and the error until the error no longer changes. A computer system for reducing data and establishing a data identification model according to claim 1, wherein the storage device and the processing device are integrated in a same computer device, and the storage device is a storage unit of the computer device, and the processing device is A processing unit of the computer device. A computer system for reducing data and creating a data identification model as claimed in claim 1, wherein the storage device and the processing device are electrically coupled to each other via a wired or wireless network to communicate with each other.