Picture processing has a wide range of applications in daily life, image recognization is one of them. From verification code recognization to face recognization, there is much knowledge related to image processing. Now, I would like to roughly introduce my understanding about verification code recognization.
To indentity the verifization code recognization, we need to select enough material. And most of the material comes from the browser’s grab packages. Based on my experence, I strongly recommend using Google Chrome,. After grabbing packages, we can start requesting the interface to get the image and save it to the local disk, it is important to divided these pictures into training sets and test sets, which make a difference in the following steps. The next step is image preprocessing, the pictures obatined should be binarized, in other word, we need to convert the RGB graph (RGB is obtained by changing the red, green and blue color channels and superimposing them on each other) into grayscale graph（grayscale graph has a logarithmically graded image between white and black） and change it to the binary graph（Binary image refers to the image of each pixel only two possible values or gray level state）,according to a certain threshold value (a kind of special value in picture processing) , which is beneficial to the recognization. It is known that Internal system or external environment will produce noise, therefore, in order to ensure the accuracy of recognition, we have to remove these noises reasonably. How to denoise is a deep topic, there are many methods such as gaussian low-pass filtering and non-local mean algorithm, but I have not understood them completely. Hence, what I am going to show you here is a slightly simpler algorithm that I learned online, which is called outlier testing method, to be specific, we need to test the number of black spots in the grid of the Sudoku. In general, the number of black points in a noise-filled Sudoku is small, so we can then go through the process of testing each of the nine grids to remove these outliers. In the next stage, all the resulting binary graphs should be divided into single characters, and we need to judge these characters divided by our sense of classification, which is helpful for machine learning. Then it is siginficant to unify the pixel size of each character that we have divided, which is convenient for picture processing. This step might involve scaling the image, and I will not introuduce the details here.
Then there is a crucial step, looking for eigenvectors. It is obviously not advisable to find eigenvector directly from the values of all the pixels because the data would be too big, which is hard for us to perform. To simplify the data, we can obtain the eigenvectors by calculating the black points per row and per column, and it can greatly reduce the complexity of the calculation. In other image recognition, we should reasonably reduce the dimension of eigenvectors without reducing the accuracy, which facilitates the operation of the program, and this is one of the most important step in image processing. After selecting the eigenvectors, we can start the simulation training, and SVM training model is used to train files with tagged features. The accuracy rate will gradually stabilize after continuous training. Of course, we can not just check the training set, because that is the source that we take an advantage of to do machine learning, that is to say when we test, we are going to use the test set that we have divided up before.
The whole process we can use MATLAB, python, LIBSVM and other programming approachs to achieve our purpose. In the later stage, better efficiency can be achieved through optimization including software hierarchy optimization and hardware hierarchy optimization.
In conclusion, what I have introduced is the most basic verification code recognition with regular arrangement of characters and other limiations, the more complex cases require better algorithms instead of what I have talked about.