Accurate flower recognition is a challenging task in computer vision due to high intra-class variation, complex background textures, and illumination inconsistencies. This study proposes an enhanced image classification framework integrating LAB color space transformation and Contrast Limited Adaptive Histogram Equalization (CLAHE) with the EfficientNet architecture. The proposed approach aims to improve visual feature separability by enhancing color stability and local contrast prior to network training. Experiments were conducted using a 17-class flower dataset, and the model achieved an overall accuracy of 98.53%, a macro-averaged F1-score of 0.9704, and AUC values close to 1.00 for most species. Visual analysis through the confusion matrix and ROC curves confirmed the model’s robustness, with only minor misclassifications observed between morphologically similar classes such as Iris–Crocus and Daffodil–Tulip. These findings demonstrate that combining LAB and CLAHE preprocessing with EfficientNet significantly enhances model generalization and visual discriminability. The method provides a lightweight yet effective solution for applications in biodiversity monitoring, precision agriculture, and automated plant taxonomy.