wherein omega is a face region based on the initial position information(s)_l，s_r，s_t，s_b) Cutting the mask image E to obtain a mask image M with the same size as the original image I;

when a plurality of faces exist in the input initial face image I, respectively generating corresponding mask images M for the faces in the initial face image I according to the mode of the step one_KThat is, the mask image M corresponding to the initial face image I is all the mask images M_KIs formulated as follows:

M(x，y)＝∑_kM_k(x，y)

then, normalization is carried out according to a formula to obtain a final mask image M, wherein the formula is as follows:

step two, generating and processing lines of the face image

Performing secondary processing on the initial face image I input in the step one, in order to obtain a line image B, firstly calculating a global image texture mutation signal of the initial face image I, and obtaining a mutation signal image S, wherein the formula is as follows:

c represents RGB three channels of the initial face image I;

and then, multiplying the mutation signal image S and the mask image M obtained in the step one pixel by pixel to obtain a mutation signal image U with a prominent human face, so that the algorithm focuses more on the image characteristics of the human face region, and the formula is expressed as follows:

U(x，y)＝S(x，y)·M(x，y)

then, a histogram of the abrupt signal image U is further calculated, and the formula is expressed as:

H(k)＝∑I(U(x，y)，k)

in the formula, the indication function I (means) indicates that when the value of U (x, y) is k, the result is 1, otherwise, the result is 0;

the cumulative histogram values are calculated on the basis of the histogram, and the formula is expressed as:

A(k)＝A(k-1)+H(k)

and then, calculating the selected data quantity M of the cumulative histogram according to the numerical value of the mask image M in the first step so that the generation parameters of the line image B can be self-adapted to the characteristics of the face region of the image, and the formula expression is as follows:

m＝0.05×∑I(M(x，y)，1)

the indication function I () in the formula indicates that when the value of M (x, y) is 1, the result is 1, otherwise, the result is 0;

searching the cumulative histogram according to the data volume example to obtain a line generation threshold t, wherein the formula is expressed as:

t^*＝arg min_tA(t)＞m

carrying out binarization on the mutation signal image U by adopting a threshold value t to generate a line image B, wherein a formula is expressed as;

step three, face image information coding processing

Processing the line image B obtained in the step two, coding by recording the position of the non-0 coordinate of the line image B, scanning and coding the image B from top to bottom by a row unit, recording the ith row data of the image B if the number of the non-0 pixel points is more than 0, otherwise discarding the ith row data, and recording all the non-0 pixel point coordinates of the row by using a method of (row number, column number 1, column number 2, column number 3, …)', wherein the column number 1 represents the row coordinate value of the 1 st non-0 pixel point coordinate of the row, the column number 2 represents the row coordinate value of the 2 nd non-0 pixel point coordinate, and so on to generate final line image coding information br;

and secondly, performing sparse color coding processing on the line image obtained in the second step, namely, equally dividing the line image B into 32 multiplied by 32 image blocks, counting the number of non-0 pixel points in each image block, and recording the number as n_i,jMeanwhile, the original face image I is equally divided into 32 × 32 image blocks in the same manner, and the color mean c of each image block is calculated_i,jThen, an image C with the same size as the original face image I is generated, all pixels are initialized to (255 ), color thinning is carried out from top to bottom and from left to left, and when n corresponding to the image block is in the image_i,jIf the color mean value is more than 0, the color mean value c corresponding to the image block is used_i,jReplacing the values of all the pixel points of the block, judging the image block of the coded image C from top to bottom and from left to left during coding, and judging the color C of the image block_i,jWhen not equal to (255,255,255), the term "(block row number, block column number, c)_i,j) The format record of' is recorded, otherwise, abandon, until getting the coded information cr of the color picture, then form the coded information E of the original facial image I by coded information cr of the color picture and coded information br of the line picture and size (w, h) of the original facial image I;

step four, image decoding processing of face recognition model constraint

Generating a binary image according to the coding information E obtained in the third step, namely, using the size data (w, h) of the coding information E to initialize all pixels of the binary image to 0, reading the line image coding information data br, deconstructing the data into line data according to the "()" mark in the data, then assigning values to the binary image according to the line number and the column number information in the line data, and then obtaining the restored line image

Then, a color image of the same size is generated, all pixels are initialized to (255 ), and restored according to the color image coding information crIn the process, the data of the color image coding information cr is read firstly, the data is deconstructed into block row-column records according to the "()" mark in the data, and the pixel in the area with the row (i-1) multiplied by 32 more than ii < i multiplied by 32 and the column (j-1) multiplied by 32 more than jj < j multiplied by 32 in the image is assigned with the value c according to the block row number i and the block column number j in the row data_i,j，Obtaining a restored color image

Line image to be restored later

And three channel color image

The image F superposed into four channels is input into an image reconstruction network N as input data to generate a reconstructed face image, wherein the image reconstruction network N is composed of an image recovery network R and a discrimination network D, the image recovery network R is composed of ten layers of full convolution networks, and the discrimination network D is composed of five layers of full convolution networks.

Example two

According to fig. 2, this embodiment provides an image compression system with maintained face recognition accuracy, which includes an image encoding module and an image decoding module, wherein an output end of the image encoding module is connected to an input end of the image decoding module, the image encoding module includes a highlighting processing module, a line generating module and an image information encoding module, the image decoding module includes a line color decoding module and a reconstruction network processing module, wherein an image is encoded as an input end, an initial face image I enters the image encoding module, is processed by the highlighting processing module, the line generating module and the image information encoding module, then enters the image decoding module, and finally a restored face image is output, a face recognition similarity penalty item is preset in the reconstruction network processing module, and by embedding the face recognition similarity penalty item, the influence of image compression on the face recognition precision can be effectively reduced.

The training loss function of the image reconstruction network N is as follows:

in the formula, R (right) represents an image F of four channels input by an image recovery network R, and a reconstructed face image is output

Phi () is an evaluation function of the human eye visual similarity of the image, and is used for judging the initial human face image I and the reconstructed human face image

The visual similarity is realized by a mean square error function, and psi () is an image face recognition similarity evaluation function used for evaluating an initial face image I and a reconstructed face image I

The degree of similarity on the face recognition model.

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. The scope of the invention is defined by the appended claims and equivalents thereof.

Claims

1. An image compression method for maintaining face recognition accuracy is characterized in that: the method comprises the following steps:

step one, highlighting image face area

step two, generating and processing lines of the face image

step three, face image information coding processing

step four, image decoding processing of face recognition model constraint

2. The image compression method with maintained face recognition accuracy according to claim 1, wherein: in the first step, when a plurality of faces exist in the input initial face image, corresponding mask images are generated for the faces in the initial face image according to the first step, that is, the mask image corresponding to the initial face image is the pixel-by-pixel sum of all the mask images, and then normalization is performed to obtain the final mask image.

3. The image compression method with maintained face recognition accuracy according to claim 1, wherein: and in the third step, scanning and coding the line image from top to bottom in a line unit, and recording the data of the line image if the number of non-0 pixel points is greater than 0, otherwise, discarding the data.

4. The image compression method with maintained face recognition accuracy according to claim 1, wherein: in the fourth step, in the process of restoring the color image, the data of the color image coding information is read, the data is interpreted into line data according to the "()" mark in the data, and then the pixels in the image area are assigned with values according to the line number and the column number information in the line data.

5. The image compression method with maintained face recognition accuracy according to claim 1, wherein: in the fourth step, the image reconstruction network is composed of an image recovery network and a discrimination network, wherein the image recovery network is composed of ten layers of full convolution networks, and the discrimination network is composed of five layers of full convolution networks.

6. The image compression method with maintained face recognition accuracy according to claim 1, wherein: in the fourth step, the pixel is initialized to (255 ).

7. An image compression system with maintained face recognition accuracy, applied to claim 1, characterized in that: the image decoding system comprises an image coding module and an image decoding module, wherein the output end of the image coding module is connected with the input end of the image decoding module, the image coding module comprises a highlighting processing module, a line generating module and an image information coding module, and the image decoding module comprises a line color decoding module and a reconstruction network processing module.

8. The image compression system with maintained face recognition accuracy as claimed in claim 7, wherein: and a face recognition similarity penalty item is also preset in the reconstruction network processing module.