JPH02206971A - Picture encoding device - Google Patents

Abstract

PURPOSE:To always obtain an optimum prediction function by allowing a neural network to study the optimum state in each block being a division of a picture prior to coding and using the result of study as an output data. CONSTITUTION:The subject device is constituted of a color original 1, picture input section 2, picture segment section 4, 1st scanner 5, 2nd scanner 7, coder 11 and an output control section 15. While the 2nd scanner 7 scans the picture in a block prior to coding in each block, a true color signal at that time is given simultaneously to a neural network so as to allow the network to study its internal parameter adaptively and optimizingly, and then internal parameter of the neural network after studying and a data reference color are outputted for each block. Thus, an optimum prediction function is always obtained.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of Industrial Application The present invention is particularly applied to the field of deriving high-quality codes from a small amount of information, and is used, for example, as a color image storage device or a device for preprocessing image signals of a color facsimile. As a result, the code length can be compressed without losing the information. The present invention relates to an image encoding device suitable for encoding color images.

2. Description of the Related Art Conventionally, a method is known in which a color image is divided into all blocks and then signal compression is performed by estimating other components from a monochromatic signal within the block. (for example,
Hiroaki Kodera, Masayoshi Naka, Katsuhiro Kanamori, "Predictive reproduction of full-color images from monochromatic images using the correlation between three-color signals," IEICE National Conference, D-81, 1988) As shown in the figure, taking advantage of the fact that there is a strong correlation between the three color components within a block of a color image, for colors other than the reference color, function expansion is performed using the reference color, and the expansion coefficient and the reference color are Information is compressed by expressing a color image using color data.

Figure 5 shows the basic configuration of an image encoding apparatus using this method. The original image data 100 is divided into blocks by a block generator 101, resulting in block data 102 of an image of each primary color component consisting of n×mul elements. When one color of the block data 102 is used as reference color data, the computing unit 103 determines a coefficient for approximating data of another color at a corresponding position in the block using a predetermined function. A calculation is performed to provide a coefficient output 104. For example, the green component Gij is used as the reference color.
Suppose we choose . Here, 1. ” represents the position within the block. At this time, the blue component B, j and the red component Rij k are each polynomials with h terms, and the coefficients bk, rk (k=1.2・・・）
seek.

Problems to be Solved by the Invention However, in the method described in the prior art, an adaptive prediction function is obtained for each image block, but the shape of the function is a predetermined polynomial, and it is difficult to approximate it by a nonlinear combination of these functions. is not addressed, so it is not possible to optimize the approximation by images. Therefore, the prediction error largely depends on the properties of the image and is not necessarily the minimum.

Unfortunately, analytical means to optimize this function are not easily realized.

In view of the above-mentioned problems, the present invention provides an image encoding device that always obtains an optimal prediction function and can easily reproduce an original image from only color data.

Means for Solving the Problems Therefore, the present invention provides image cutting means for cutting out a block of n×m pixels from a target color image with respect to a characteristic reference color representing the image; A neural network model means (hereinafter referred to as a neural network) that estimates and outputs the color of a pixel at the same position of the block as a direct input of the representative color represented, and This problem can be solved by providing a first scanning means for scanning a block, a second scanning means for scanning each block pixel by pixel, and a parameter reading means for reading out state parameters inside the neural network. solve.

Function: In the image encoding device according to the present invention, prior to encoding,
The neural network is trained to the optimal state within the blocks into which the image is divided, and the learning results are used as output data, so the optimal prediction function is always found.

That is, in the image encoding apparatus according to the present invention, after the entire screen is divided into blocks of n×m pixels by the cutting means, the blocks are sequentially read by the first scanning means. While the block is repeatedly scanned by the second scanning means, the true color signal at that time is simultaneously given to the neural network to adaptively learn and optimize its internal parameters, and estimate the color signal from the reference color. Organize as you do.

Thereafter, for each block, the learned internal parameters of the neural network and the reference color data are used as output data. Therefore, functions can be optimized for each block, and all nonlinear prediction functions can be generated due to the characteristics of the neural network, making it easy to reduce prediction errors.

EXAMPLES Below, the present invention will be explained in detail with reference to the drawings. FIG. 1 shows the configuration of an image encoding device in an embodiment of the present invention. As shown in FIG. 1, a color original 1 is color-separated by an image input section 2 and sent as color image data 3 to an image cutting section 4. Image input section 2
is a device that extracts color information contained in an image (for each pixel forming an image), and is constructed using a color television camera, a mechanical color scanner, etc. 3 represents the color attributes of each pixel, and is a three-primary color signal of red, blue, and green, or an analog signal representing other chromaticity, or a digital signal having an appropriate number of bits to represent the color.

The first read address signal 6 from the first scanner 5 and the second read address signal 8 from the second scanner 7 are applied to the image cutout section 4 . The first scanner 5 provides an address for reading out the image in blocks of n×m pixels. The second scanner 7 also provides addresses for sequentially reading out the pixels in the block.

The image cutting section 4 outputs a reference color signal 9 and compressed color signals 1o of two normal colors. The reference color signal 9 is applied to the reference signal input terminal 12 of the encoder 11, and corresponds to one color of the color signals. Further, the compressed color signal 1o is transmitted to a code signal input terminal 13 of the encoder 11.
The data is given to the code and converted into a code to become the final data 14.

After the final data 14 is output from the encoder 11, it is mixed with the reference color signal 9 by the output controller 15 to become the output signal 16. Further, the encoder 11 outputs a timing signal 17 for controlling the operations of the first scanner 5 and the second scanner 7.
of the scanner 7.

Next, each of the above components will be explained in more detail.

As shown in FIG. 2, the image cutting unit 4 includes an image memory 20 for storing image data of one screen of a document, and a signal of a block selected from the image memory 20 by a first address signal 6. The block memory 21 stores the data of the pixel selected by the second address signal 8, and outputs the reference color signal 9 and the compressed color signal IO.

The internal configuration of the encoder 11 shown in FIG. 1 is as shown in FIG. 3, and the reference color signal 9 applied to the reference signal input terminal 12 has a neural network power of 300 Å. Neural network 30, first output end 3
1 and a second output end 32. First output end 3
1] output 33 is sent to the first subtractor 36 and the second
The second output 34 from the output 32 of the second subtractor 3
6 and 0 is added as the subtractable input. The subtraction input 37 of the first subtractor 350 receives data of one color of the compressed color signal 10, and the subtraction input 38 of the second subtractor 36 receives data of another color of the compressed color signal 10. data is given. The output of the first subtractor 35 is the first teacher signal 39
, and the output of the second subtractor 36 is given to the learning control section 41 as the second teacher signal 40. The learning control unit 41 controls the neural network 30 in correspondence to each of the two color data of the compressed color signal 10.
Two teacher signals 39 and 40 are given by the difference between the output signals 33 and 34 output from the compressed color signal lO.
Accordingly, the internal state signal 4 of the neural network 30
3, it outputs a control signal 42 for updating. At this time, a control signal 42 for changing the internal state
What kind of noise is given is determined by a well-known neural network learning algorithm. For example, the error backpropagation learning method by Rumerno et al.
E, Rnmelhart et al., Parallel Distributed Processing (D, E, Rnmelhart et al.,
PARALLEL DISTRIBUTED P C)
CESSING Chapter 8)] Ahe. Learning is repeated until the outputs 33 and 34 of the neural network 30 converge on each component of the compressed color signal 10 to a sufficiently small error. For this purpose, the learning control unit 41 outputs the timing signal 17 and The second scanner 7 shown in FIG. 1 is controlled to repeatedly read out pixels within the block. After the iterative learning is completed, the output 33 of the neural network 30
and 34 give a sufficiently good approximation to the compressed color signal 10, in other words, when the absolute directivity of the teacher signals 39 and 40 becomes sufficiently small, the internal state 43 of the neural network 30 is read out by the readout control unit 44. served,
Output all 14 final data.

FIG. 4 shows the configuration of the neural network 30 in more detail. The signal inputted to the reference signal input terminal 12 has a weighting coefficient Wh.

is multiplied by , and added to the neuron unit 50 of the input layer 51. Fixed human power 52 is also added to the neuron unit 50 after being multiplied by a 7Io weighting coefficient Wis. The output 53 of the neuron unit 50 is connected to the neuron units 55a-55n in the intermediate layer 54 via weighting factors Wm--Wmnf. middle layer 54
A fixed input 58 is applied to the neuron units 55a to 55n, respectively, via a weighting factor of W ca % - W c n. Further, the outputs 56a to 56n of the neuron units 55a to 55n are the same as those of the two neuron units 57a and 57b existing in the output layer.
, the weighting coefficients W o a %−W o n and W o
Connected via a-W on f.

Neuron units 57a, 57bu, [
A constant input 59 is applied via on-multiplying coefficients Wco and Wct. Neuron unit 57a, 57
b) Output ports are connected to the outside as a first output end 31 and a second output end 32 of the neural network 30, respectively.

The internal state signal 43 is for transmitting all the values of these weighting coefficients to the outside, and is output while sequentially scanning the weighting coefficients, for example. Control signal 42 controls each of the weighting factors l! It is a signal given from the outside to change the weighting coefficient, and is transmitted while sequentially scanning the amount of change in each weighting coefficient, for example.

Each of the neuron units 50.55a to 55n, 573, and 57b has a characteristic of providing a nonlinear output such as a sigmoid function with respect to the sum of input signals.

According to the image encoding device having the configuration and learning process described above, a color input signal can be easily reproduced from only color data within an image block. Further, although not described here, the amount of codes for the reference color signal can be further reduced by using an encoding method that reduces the large-scale loss of the reference color signal.

Effects of the Invention As described above, the present invention uses a neural network to realize a transfer system that accurately approximates the characteristics between color components that have a complex correlation with a simple configuration.
Furthermore, by using the internal state of the neural network as the image code instead of the color signal, the amount of information can be compressed to a certain degree.

[Brief explanation of the drawing]

Figure 1 is a block diagram of an image encoding device according to an embodiment of the present invention, Figures 2 to 4 are block diagrams of each main part of the same device, and Figure 5 is a block diagram of a conventional image encoding device. It is a wiring diagram. DESCRIPTION OF SYMBOLS 1... Color document, 2... Image input unit, 4... Image cutting unit, 5... First scanner, 7... Second scanner, 11... Encoder, 15., Output controller.

Claims (2)

[Claims] (1) An image cutting means for cutting out a block of n×m pixels for a characteristic reference color from a target color image, and inputting a representative color value representing each pixel in the cut out block to a neural network model means for estimating and outputting the color of pixels at the same position; a first scanning means for scanning the n×m pixel block over the entire screen; A second scanning means for scanning, and a parameter reading means for reading out internal state parameters of the neural network model means, and before encoding in each block, the second scanning means scans the inside of the block. While scanning, the true color signal at that time is simultaneously given to the neural network model means to adaptively learn and optimize its internal parameters, and then the values of the learned internal parameters of the neural network model means are calculated. and the reference color data for each block. (2) The image encoding device according to claim 1, characterized in that a signal of a certain color among the color components is always used as the reference color.