This paper focuses on a decision support system that leverages artificial intelligence to evaluate the editorial quality of textbooks distributed to Brazilian public schools through the Brazilian Textbook Program. The program is crucial for enhancing educational resources for students and educators throughout Brazil. Our study relies on a dataset comprising images presenting a subset of textbook image characteristics to tackle a multiclass problem, categorizing them into three classes: sharp, defocused-blurred, and motion-blurred. We conducted experiments using the Fourier and Haar transforms to evaluate their impact on the performance of our CNN classification model. Using Fourier and 10-fold cross-validation, the mean accuracy of our models varied from 75.44% to 80.25%, the mean precision varied from 78.38% to 83.87%, the mean recall varied from 76.05% to 80.37%, and the mean Area Under the Curve (AUC) varied from 90.27% to 92.14%. Using Haar, the mean accuracy of our models varied from 69.08% to 87.14%, the mean precision varied from 78.38% to 87.80%, the mean recall varied from 76.05% to 87.25%, and the mean AUC varied from 90.27% to 93.95%. These results indicate that employing CNNs with data preprocessed using the Haar transform method can achieve more consistent results for assessing textbook image quality.
Rosa et al. (Thu,) studied this question.