Civil engineering fields and manufacturing plants need surface crack assessment to prevent structural failures. The study conducts research to create a computational system which integrates stacking ensemble models with deep learning models to enhance detection accuracy. The studied data contains labeled images for cracks and non-defects which underwent extensive preprocessing before applying augmentation techniques. The research applies ResNet50 in combination with VGG16 and InceptionV3 CNN training models to generate their outputs through logistic regression meta-learners. A stacking ensemble model achieves better performance than single parasitic structures with 99.90% recall and 99.81% accuracy and 99.81% F1-score and 99.73% precision as well as. InceptionV3 secured the most notable position among original models due to its achievement of 99.66% accuracy coupled with 99.66% F1-score and VGG16 occupied the second position with matching 99.56% accuracy and F1-score but ResNet50 maintained third place with 95.65% accuracy and 95.71% F1-score. The third position belonged to ResNet50 that delivered 95.65% accuracy and 95.71% average F1-score. Applications in real-world settings stand to gain substantially from the identification methods outlined which produce excellent performance results in crack detection tasks.