This study presents the development and evaluation of an intelligent flower classification system using convolutional neural networks (CNNs) for three species: Bellis perennis (Daisy), Lavandula angustifolia (Lavender), and Helianthus annuus (Sunflower). Given the morphological similarity among floral species and the need for non-invasive and precise methods, four widely documented CNN architectures were employed: VGG16, VGG19, InceptionV3, and Xception. The training was structured through three strategies: Data Augmentation (DA), Feature Extraction (FE), and Fine Tuning (FT). Each model was evaluated with different optimizers (Adam, RMSprop, and SGD), selected based on their performance reported in previous studies on flower classification. A customized dataset, balanced across classes, totaling 3,000 images was used. The results demonstrate that CNN-based models are highly effective, exceeding 94% accuracy in the test set in several cases. InceptionV3 and VGG19 optimized with Adam showed the best overall performance, highlighting the consistency and robustness of this optimizer. The Xception model also presented competitive results, achieving accuracies above 93% in testing, with stable behavior across different training strategies and optimizers evaluated. In contrast, the SGD optimizer showed good performance only with InceptionV3. The Feature Extraction technique achieved the highest validation accuracy (97.50%), indicating that fine tuning is not always essential. Overall, it is concluded that the combination of architecture and optimizer has a more significant impact on performance than the applied training strategy.