<p>Detecting cracks in building structures is a critical part of ensuring their security and long-term performance. For buildings to remain safe and durable, regular inspections by qualified professionals are necessary. Historically, these inspections have been carried out by human inspectors through visual observation. However, with the advancement of artificial intelligence, particularly in computer vision, automated methods have become increasingly successful in identifying cracks and other structural issues. This work proposes a novel method for surface crack detection in concrete structures using an improved deep learning framework. By integrating preprocessing, segmentation, feature extraction, and detection, the proposed method ensures the accurate identification and classification of cracks in the images. Initially, RGB images are converted to grayscale and resized, followed by semantic segmentation using an improved U-Net model for more efficient segmentation of concrete surface cracks. Then, significant features including improved SLBT, raw features, statistical features and GLCM-based features are extracted, which capture texture, intensity, and statistical patterns from the segmented image. Finally, the combined feature set is input into an Improved Parallel CNN (IP-CNN) model, which combines the strengths of 1D and 2D convolutional networks to detect cracks. The IP-CNN model fuses information across different domains and utilizes attention mechanisms and advanced activation functions to minimize redundancy and prevent overfitting. Experimental results demonstrate that the suggested method outperforms conventional models with respect to accuracy, robustness, and efficiency, offering a reliable tool for automated crack detection in structural health monitoring applications. The IP-CNN strategy attained higher values with accuracy at 0.909, precision at 0.917 and f-measure at 0.912, respectively.</p>