EEN 538: DIGITAL IMAGE PROCESSING

Optical Character Recognition (OCR) using binary image processing with MATLAB
Abstract- Nowadays, Optical Recognition is becoming a very important tool in several fields: medicine, physics, cosmology, traffic (plate numbers), etc. We can also use this to recognize character for example to digitalize a book. We will talk about this last topic in this report: Optical Character Recognition (OCR). I. INTRODUCTION Once we have the b&w image we can start the segmentation process. To do that we can use the function “bwconncomp”. This function returns us a struct from where we can obtain the characters because it gives us all the connected components. Thus, we can use it to get all the character even if they have 2 or 3 objects. This function returns us the pixels of the connected components (characters) but we have to figure out from those, the coordinates of the character in the original matrix (row and columns). To do this, we will obtain the centroid of every connected component and from it and using the first and last pixel detected of the connect component, we can figure out the exact coordinates of the image. The idea is as follows: Firstly, we can to convert the number that the function returns us to a column and a row. We can do this using the total rows of the original image. Once we have the first and last pixel detected of the connect component in (row, column) we can figure out directly the x-coordinates of the character in the image. Then using the centroid and those pixels we can also deduce the x-coordinates of the character in the image. This process will be more difficult to segment characters with 2 or 3 objects. In these cases we will follow the same process as before but now, the centroid will be the mean of the 2 or 3 centroids. We will repeat this process changing the b&w threshold till the segmentation system finds the correct number of objects that we expect. IV. Before starting with the process It is important to say that depending on if the characters are white and the background black or vice versa, we will have to do the not operation to the image matrix because some functions in MATLAB as “bweuler” require these modifications. RECOGNITION

In this report I will talk about a system (implemented with MATLAB) that will have the objective of recognizing several characters of the Russian alphabet from an image that will include more than one character. This image may include noise and different separation between characters. II. IMPLEMENTATION SUMMARY

To do this, I will follow these steps: Segmentation: in this step, we will obtain the different characters from the original image. Recognition: here, we will recognize the characters that we have got in the previous step. III. SEGMENTATION

In this step, first of all, I will convert the original image (.jpg, .png, etc) to a binary image (black and white) because all the functions that we will use in MATLAB to find a solution to the problem work with binary images. To do this conversion, we need a threshold to decide when an original pixel is going to be a black or white one. This threshold either can be obtained with another function or fixed by the programmer.

In this last step we will extract same properties of the characters to compare them with a date base (with properties of several characters previously analyzed) to recognize them. There are a lot of properties that can be used but I have used the following ones: 1. Euler Number: total number of objects in the image minus the total number of holes.

2. 3. 4. 5.

Number of cuts with row=20 Number of cuts with row=25 Number of cuts with column=17 Number of cuts with column=23

Thus, we have 5 properties to distinguish among the characters. It is important to say that I have taken the idea of the “Number of cuts” from [1]. I have done some modification to this idea such as increasing the number of cuts and modify the points of the cuts. Therefore, with these properties we can recognize the characters. V. TEST Figure 4. Final Result Where the black objects are characters that do not belong to the data base. We can observe how the system worked perfectly (accuracy 100%). 2. Test2:

To test the system we will use 4 images in increasing difficulty. 1. Test1:

We will use the following image:

We will use the following image:

Figure 1. Original Image Thus, if we convert it to a b&w image, we obtain:

Figure 5. Original Image Thus, if we convert it to a b&w image, we obtain:

Figure 2. B&W Image Then, after de segmentation process, we get these objects:

Figure 6. B&W Image Then, after de segmentation process, we get these objects:

Figure 3. Connected Components Finally, after the recognition process, we obtain the final result:

Figure 7. Connected Components Finally, after the recognition process, we obtain the final result:

Figure 12. Detected characters Finally, after the recognition process, we obtain the final result: Figure 8. Final Result Where the black objects are characters that do not belong to the data base. We can observe how the system worked perfectly again (accuracy 100%). 3. Test3:

We will use the following image:

Figure 13. Final Result We can observe how the system fails in this case due to the noise of the image (accuracy 20%). Figure 9. Original Image Thus, if we convert it to a b&w image, we obtain: 4. Test4:

We will use the following image:

Figure 14. Original Image Figure 10. B&W Image We erode the image: Thus, if we convert it to a b&w image, we obtain:

Figure 15. B&W Image Figure 11. Dilated Image Then, after de segmentation process, we get these objects: Then, after de segmentation process, we get these objects:

Figure 16. Connected Components

We can observe how the system completely failed because of the high noise level (accuracy 0%). VI. CONCLUSION

We can observe how when the noise is cero or very low, the system works perfectly, but when the noise increases, the system starts to fail till the accurary goes to 0%. We will finish the report I will write the strength and weaknesses of the system: Strength: • There are not input parameters: There are ‘for’ ciclos to get the best result. The system could recognize characters with 1, 2 or 3 objects.

•

Weaknesses: • • The system fails with high noise-level. Similiar charactertics among characters. Two characters may have 4 of 5 identical characteristics.

REFERENCES

[1]http://iris.elf.stuba.sk/JEEEC/data/pdf/5_106 -3.pdf