-&gt;Context. Optical Character Recognition (OCR), having a computer recognize text from an image, is not as intuitive as human recognition. Even small (to human eyes) degradations can thwart the OCR result. The problem is that random unknown degradations are unavoidable in a real-world setting. -&gt;Objectives. The noise tolerance of Tesseract, a state-of-the-art OCR engine, is evaluated in relation to how well it handles salt and pepper noise, a type of image degradation. Noise tolerance is measured as the percentage of aberrant pixels when comparing two images (one with noise and the other without noise). -&gt;Methods. A novel systematic approach for finding the noise tolerance of an OCR engine is presented. A simulated environment is developed, where the test parameters, called test cases (font, font size, text string), can be modified. The simulation program creates a text string image (white background, black text), degrades it iteratively using salt and pepper noise, and lets Tesseract perform OCR on it, in each iteration. The iteration process is stopped when the comparison between the image text string and the OCR result of Tesseract mismatches. -&gt;Results. Simulation results are given as changed pixels percentage (noise tolerance) between the clean text string image and the text string image the degradation iteration before Tesseract OCR failed to recognize all characters in the text string image. The results include 14400 test cases: 4 fonts (Arial, Calibri, Courier and Georgia), 100 font sizes (1-100) and 36 different strings (4*100*36=14400), resulting in about 1.8 million OCR attempts performed by Tesseract. -&gt;Conclusions. The noise tolerance depended on the test parameters. Font sizes smaller than 7 were not recognized at all, even without noise applied. The font size interval 13-22 was the peak performance interval, i.e. the font size interval that had the highest noise tolerance, except for the only monospaced font tested, Courier, which had lower noise tolerance in the peak performance interval. The noise tolerance trend for the font size interval 22-100 was that the noise tolerance decreased for larger font sizes. The noise tolerance of Tesseract as a whole, given the experiment results, was circa 6.21 %, i.e. if 6.21 % of the pixel in the image has changed Tesseract can still recognize all text in the image. / <p>42</p>

This report analyzes the difference between discriminative and generative image classifiers when tested on noise. The generative classifier was a maximum-likelihood based classifier using a normalizing flow as the generative model. In this work, a coupling flow such as RealNVP was used. For the discriminative classifier a convolutional network was implemented. A detailed description of how these classifiers were implemented is given in the report. The report shows how this generative classifier outperforms the discriminative classifier when tested on adversarial noise. However, tests are also conducted on salt and pepper noise and Gaussian noise, here the results show that the generative classifier gets outperformed by the discriminative classifier. Tests were also conducted on Gaussian noise once both classifiers had been trained on Gaussian noise, the results from these tests show that the discriminative classifier performs significantly better once trained on Gaussian noise. However, the generative classifier does only show marginal increases in performance and performs worse on clean data once trained on Gaussian noise. / Den här rapporten analyserar skillnaden mellan diskriminativa och generativa modellklasser för bildigenkänning när de testas på brus. Den generativa modellklassen var en maximum-likelihood baserad generativ klassifikationsmodell. Inom detta arbete användes kopplingsflödet RealNVP. För den diskriminativa bildigenkänningsmodellen så implementerades ett faltningsnätverk. En detaljerad beskrivning för hur dessa bildigenkänningsmodeller genomfördes är given i rapporten. Rapporten visar hur den generativa modellklassen överträffar den diskriminativa modellklassen när de testas på adversarialt brus. Testerna utförs emellertid med salt och peppar brus och Gaussiskt brus, för dessa visar resultaten att den generativa modellklassen överträffas av den diskriminativa modellklassen. Den generativa modellklassen visar emellertid endast marginella ökningar i prestanda, och har en sämre prestanda på ren data efter att den tränats på Gaussiskt brus. / Kandidatexjobb i elektroteknik 2022, KTH, Stockholm