Technical Field

The invention relates to a computer-based synthetic vector image generation method for generating a synthetic vector image and/or a synthetic vector image template from an image.

Background

Synthetic data generation, which is among the artificial intelligence applications, is used to generate new data from a data set in case there is not enough data on a subject; it can also be used to generate new, original, and artistic data from the relevant data set.

The existing synthetic image generation methods in the state of the art basically comprise the steps of;

- displaying any image at a reduced size using a pre-trained library;

- statistically learning the dimensionally reduced image data in its own space using applications such as Generative Adversarial Networks (GAN), Auto Encoder or Restricted Boltzmann Machine (MA1RBM),

- this learned space generating new "images" based on random values and normal distribution.

However, in these methods, the points of the image are seen as pixel values (x, y, ; rgb), and therefore, the sharpness of these points decrease after various learning processes. Since the images generated are pixel-based, they cannot be enlarged and used for printing.

Vector images, on the other hand, are a type of graphic that does not depend on resolution, every point of the design consists of numerical data, and no detail is lost despite the change in size. Since the images used as input in the embodiments in the state of the art and outlined above are not displayed as vectors; the resulting output will be a non-vector image. The patent document numbered US10860836B1, which is in the state of the art, discloses the creation of synthetic image data for computer vision models. With the method described in the said document, it is aimed to create a synthetic image, but not a vector image. The invention in the patent application numbered CN113012254A belongs to the field of computerized image processing. In particular, the invention is related to an underwater image synthesis method based on self-controlled training at the pixel level and aims to solve the problem in which the stable, high quality underwater image synthesis cannot be achieved in the prior art. With the said invention, synthetic underwater images are obtained by processing and correcting real underwater images. However, the images obtained are not vector data.

Therefore, there is a need for a method that enables the generation of a synthetic vector image from an image used as an input.

Brief Description and Objects of the Invention

The invention is a synthetic vector image generation method developed to eliminate the above-mentioned disadvantages; with the said method, a standard image input is analyzed, the shapes in the same are determined, and the image is transformed into a vector image by defining it as a composition consisting of these shapes.

Vector synthetic image output, which is defined parametrically in this way, can be modified as desired by changing the said parameters.

The said method consists of three basic stages. In the first stage, the image data is detected, analyzed, and converted into a single vector; in the second stage, training data is created with the obtained vectors; and in the third stage, the neural model is trained, the vector of the desired shapes/objects is learned and generated.

Description of the Figures in the Invention

Figure 1: An exemplary flow chart of the invention.

Figure 2: Flow chart of an exemplary detection method. Detailed Description of the Invention

Since any image desired to be generated may comprise many different parameters such as location, coordinate, size, direction, light, color; available in the prior art, generating a synthetic data regarding an image is not possible with the following methods in the state of the art: averaging of data or images, randomly selecting a subset within the dataset and averaging the subset, modifying a piece of data by moving it up/down at various rates.

Since image data is unstructured data, operations such as averaging/selection cannot be performed. Even if synthetic image data is generated using existing artificial intelligence and deep learning techniques; since the images generated are not vectorial, there are quality problems during printing; standard images that cannot be enlarged to the desired size are created, and this creates problems with sizing and image manipulation.

The present invention provides a computer-based method for generating a synthetic vector image (or image template) from an image input.

The method of the invention, as a solution to the problems involved in the embodiments in the state of the art, comprises analyzing each image received as input, identifying the shapes contained therein, and displaying the image as a composition of shapes. Thus, a vector image/image template output is generated from an image entered into the system.

In this direction, the method of the invention includes the following steps:

- creating at least one shape generation function arranged in accordance with the generation of a vector shape of the desired size (Step 1). Herein, by vector shapes we refer to shapes in vector format, which can include all shapes in any shape library, such as heart shapes, triangles, ovals, arcs, moons and so on, as well as patterns, motifs, decorations and so on.

- randomly calling the said shape generating function(s) and placing the vector shapes to be generated by this/these function(s) on an empty image (Step 2). In this step, not only a vector image is generated, but also a description of the way in which the vector image is generated is established.

- training an artificial intelligence supported object detection program trained using image processing - object detection and recognition methods to detect at least one parameter of said vector shapes (Step 3). Wherein the said parameter preferably comprises at least one of the shape (type of said vector shape, for example triangle, square, zigzag, etc.); location (x,y coordinates); size (width, length etc.); color (RGB color codes); rotation (rotation angle) parameters of the respective vector shape. This step intends to determine the shape, position, size, color, rotation parameters of the said vector shapes in an image consisting of randomly generated vector shapes mentioned in step 2.

- reading an input image to be used as input and detecting at least one parameter (at least one of the parameters mentioned in Step 3) for all shapes included in this input image by the said object detection program (Step 4).

- creating a vector image output (step 5) by defining the said input image as a list of vector shapes containing the parameters determined as a result of the detection (in step 4).

In the said method, Step 1, Step 2, and Step 3 are training steps; wherein the creation of the necessary method/set of rules is provided so that the said object detection program can analyze an image input. In Step 4, an input image is analyzed following the rules learned in Steps 1-3, and through this analysis, the shapes in the said input image are identified and the input image is defined as a composition of shapes. This identification comprises the set of parameters for each shape in the input image. Thus, in Step 5, a reproducible vector image output is obtained from the said input image.

In a preferred embodiment of the invention, the said method can also be used to generate new and original vector images. In the said embodiment, after running Steps 1-5 for multiple input image in the same category or cluster, the following steps are realized;

- matching the generated image printouts and the resulting shapes (Step 6)

- calculating how much each detected vector shape deviates from the said parameters for multiple image outputs obtained as a result of said matching (Step 7). - generating new vector images based on random numbers obtained from a normal distribution, together with the obtained statistical deviations (Step 8).

With this embodiment, a new vector image/vector image template can be generated independently of an input image. With the method of the invention, a vectorized version of the input image is obtained from an input image; this vectorized version is generated as a function; by adding various parameters to the function, the images to be generated can be parametrically changed and thus many different images can be generated. The said embodiments of the invention can preferably be used in industry as pattern generation technology in the field of generation, weaving and printing of textile products.