Early diagnosis of cervical cancer is challenging, especially in developing countries due to limited resource, infrastructure, and experts. Recent improvements in deep learning-based early cervical cancer detection techniques can potentially lower these challenges while increasing diagnosis accuracy. In this study, experiments were conducted on 465 histopathological images collected from patients with adenocarcinoma, squamous cell carcinoma, and squamous intraepithelial lesion. CNN and ResNet-50 models were trained for classification after the application of data augmentation, color normalization, and synthetic minority oversampling technique (SMOTE) on the data. Classification accuracies of 86.43% and 92.14% were found using CNN and ResNet-50 trained on raw data, respectively. ResNet-50 trained on a dataset “without a preprocessing technique,” “affine transform-based data augmentation,” “data augmentation and histogram match-based color normalization,” “SMOTE,” and “color normalization and SMOTE” achieved accuracies of 94.20%, 96.39%, 95.32%, 99.14%, and 90.56%, respectively. In general, the study investigated how image preprocessing, specifically data balancing, influences the effectiveness of deep learning models. Additionally, it examined the possibilities of utilizing computer-aided image processing to detect cervical cancer. The findings show that, with proper data processing, the proposed technique might be used as a cervical cancer screening tool, particularly in low-resource settings.