KR0160631B1 - Digital image data compressor - Google Patents

Abstract

Disclosed is a video data compression apparatus employing a neural network in order to increase the computational speed when calculating a compression constant.

The image data compression device according to the present invention has an effect of implementing an efficient and simple image data compression device by adopting a neural network to calculate an appropriate compression constant considering the data transmission capability of the transmission system and the complexity of the image data.

Description

Digital video data compression device

1 is a block diagram showing an embodiment of a digital image data compression apparatus according to the present invention.

2 is a block diagram showing another embodiment of the digital image data compression apparatus according to the present invention.

3 is a flowchart showing the operation of the neural network shown in FIGS. 1 and 2.

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a device for compressing digital image data, such as a digital VRL, and more particularly, to a digital image data compression apparatus employing a neural network in order to improve the accuracy of the compression constant calculation.

In conventional digital image data compression using a quantizer and a variable length encoder, it is necessary to keep the bit rate of the image data constant in consideration of the power factor of the transmission line. To this end, the compression factor (SF) in the quantization unit is appropriately adjusted according to the complexity of the image.

The complexity of the image is calculated by summing the degree of change of the DCT transform blocks constituting the original image, and the degree of change of the DCT transform block is an average value of the differences between adjacent pixels.

However, in the prior art, there was a problem in that the complexity of the image and the compression compression corresponding to the variously changing images could not be accurately analyzed.

An object of the present invention is to provide a digital image data compression apparatus which is designed to solve the above problems and accurately calculates the complexity and compression constant of an image.

In accordance with another aspect of the present invention, there is provided a digital image data compression apparatus comprising: a block division unit for dividing image data into DCT transform blocks and sequentially outputting respective DCT transform blocks; A DCT converter configured to input a DCT transform block output from the block splitter to perform DCT conversion; A quantization unit for performing quantization with a given quantization step size on the DCT transform coefficient of the DCT transform block output from the DCT transform block; An entropy encoder for entropy encoding and outputting DCT transform coefficients of the quantized DCT transform block in the quantizer; A buffer that stores and outputs a plurality of DCT transform coefficients encoded by an entropy encoder; A buffer control unit for outputting a status signal indicating a full state of the buffer; A neural network unit which is trained to calculate an appropriate compression constant according to characteristics of an input image, and outputs a compression constant corresponding to the complexity of the image data by inputting image data output from the block splitter; And an adjusting unit for mixing the state signal generated in the buffer control unit and the compression constant generated in the neural network unit to provide the quantization step size of the quantizer. Hereinafter, the configuration and operation of the present invention will be described in detail with reference to the accompanying drawings.

1 is a block diagram showing the configuration of a digital image data compression apparatus according to the present invention. The apparatus shown in FIG. 1 includes a color separation unit 10 and a color separation unit 10 that separate R, G, and B color signals of an original image into luminance signals Y and color difference signals RY and BY. The block divider 12 and the DCT transform output from the block divider 12 for dividing the generated luminance signal Y and the color difference signals RY and BY into DCT transform blocks and sequentially outputting respective ACT transform blocks. DCT converter 14 for inputting a block to perform DCT transform, quantizer 16 and quantizer for performing quantization with a given quantization step size for DCT transform coefficients of DCT transform block output from DCT converter 14. An entropy encoder 18 for entropy encoding and outputting the DCT transform coefficients of the quantized DCT transform block in (16), and a buffer for storing and outputting a plurality of encoded DCT transform blocks in the entropy encoder 18 ( 20, a buffer control unit 2 for outputting a status signal indicating the fullness of the buffer 20 2), the state signal generated by the neural network unit 24 and the buffer controller 22 which inputs the image data from the block divider 12 and calculates and outputs a compression constant SF corresponding to the complexity of the image data. And an adjusting unit 26 that mixes the compression constant SF generated by the neural network unit 24 and provides it as a quantization step size of the quantization unit 16.

In the apparatus shown in FIG. 1, the neural network section 24 is constructed using neural networks. As is well known, the neural network is an information processing system modeled by the neural tissues of living organisms and has a parallel, distributed connection structure of simple elements, and generates a necessary output by generating a dynamic response to an input from the outside. . Such neural networks have the advantage of being able to simulate complex mathematical modeling by repeating simple learning processes.

The neural network unit 24 is trained to calculate an appropriate compression constant according to the characteristics of the input image.

In the subsequent application process, when the image data from the block dividing unit 12 is input, the neural network unit 24 determines the characteristics of the image data and outputs a compression constant corresponding thereto.

The compression constant output from the neural network unit 24 is input to the adjusting unit 28 and mixed with the state signal from the buffer control unit 22. The quantization step size output from the adjustment unit 26 is applied to the quantization unit 16 to adjust the quantization step size of the DCT transform coefficients, thereby controlling the bit amount of the DCT transform block.

FIG. 2 is a block diagram showing another embodiment of the digital image data compression apparatus according to the present invention. The same reference numerals are used for elements performing the same functions as those shown in FIG.

The neural network portion shown in FIG. 2 includes a digital neural network. Digital neural networks consist of a combination of multiple multipliers and step function operators, resulting in a significant time difference between the input and the output.

Due to this time difference, a delay time is generated between the system through the quantization unit 16 and the system through the neural network, making it difficult to process the image in real time. In the embodiment shown in FIG. 2, by using the analog neural network 31b for the neural network, the adjustment unit 26 is simultaneously analogized to shorten the time for calculating the compression constant and the quantization step size. In addition, the adjusting unit 26 mixes the compression constant and the state signal using the operational amplifier 31c, and converts the state signal from the buffer control unit 22 to the analog resistance value by the voltage control resistor 32b to convert all circuits. Analogization.

3 is a flowchart showing the operation of the neural network shown in FIGS. 1 and 2. As shown in FIG. 3, the neural network is applied through a learning process and an application process.

In the learning process 300, the neural network is trained to output corresponding compression constants for images of various complexity. This learning process is performed by adding a plurality of image data whose complexity and compression constant are predefined and outputting a compression constant corresponding thereto. The typical learning method for this is the error back propagation method. In this error backpropagation method, the neural network learns to output the optimal compression constant by repeating the process of re-adjusting the corresponding compression constant and the calculated compression constant when any image data is inputted. .

Here, the complexity is defined as the average value of the waves between the pixels constituting the image and their sum, and the compression constant is determined by the complexity and the prescribed bit amount.

The application process 310 calculates a compression integer using the neural network learned in the learning process 300. In the learning process 300, since the neural network is combined to output a corresponding compression constant for image data having a certain complexity, the neural network may calculate the compression constant in real time.

As described above, the digital image data compression device according to the present invention has an effect of implementing a simple image data compression device by adopting a neural network to calculate an appropriate compression constant considering the data transmission capability of the transmission system and the complexity of the image data. .

Claims (2)

A block dividing unit 12 dividing the image data into DCT transform blocks and sequentially outputting each DCT transform block; A DCT converter 14 for inputting a DCT transform block output from the block divider 12 to perform DCT conversion; A quantization unit (16) for performing quantization with a given quantization step size for the DCT transform coefficients of the DCT transform block output from the DCT transform unit (14); An entropy encoder 18 for entropy encoding and outputting DCT transform coefficients of the quantized DCT transform block in the quantizer 16; A buffer 20 for storing and outputting a plurality of DCT transform coefficients encoded by the entropy encoder 18; A buffer controller 22 for outputting a status signal indicating the fullness of the buffer 20; The neural network unit is trained to calculate an appropriate compression constant according to the characteristics of the input image, and outputs a compression constant (SF) corresponding to the complexity of the image data by inputting the image data output from the block splitter 12. (24); And an adjusting unit (26) for mixing the state signal generated by the buffer control unit (22) and the compression constant generated by the neural network unit (24) to provide the quantization step size of the quantizer. The neural network unit 24 of claim 1, wherein the neural network unit 24 converts the image data provided from the block divider 12 into an analog signal and the D / A converter 31a. An analog neural network 31b for inputting image data converted into an analog signal and outputting a compressed integer in an analog state; The adjusting unit 26 converts the state signal from the buffer control unit 22 into an analog signal and converts the analog state signal output from the D / A converter 32a into an analog resistance value corresponding thereto. The operational amplifier 31c and the operational amplifier, which amplify the compressed constant output from the analog control network 32b and the analog neural network 31b and convert the gain to the gain by the resistance value of the voltage control resistor 32b. And an A / D converter (32d) provided as a step size of (32c).