<p>Existing methods for crack detection in civil and geo-engineering often struggle with high computational costs, limited accuracy, and challenges in handling subtle features within images. These issues hinder the effectiveness of traditional models in real-world applications, where both precision and efficiency are crucial. This paper presents a novel deep learning-based methodology for crack and non-crack classification, utilizing an optimized VGG16 model enhanced with advanced image processing techniques. The model leverages a fusion of CLAHE and corrected image enhancement with Discrete Wavelet Transformations (DWT) to improve the contrast and brightness of both crack and non-crack regions. To address computational complexity, the VGG16 architecture is pruned, leading to a significant reduction in training time (~ 21%) and testing time (~ 17%), while maintaining high performance. Multi-learning techniques using the Convolutional Block Attention Module (CBAM) are incorporated to enhance feature extraction by focusing on the most relevant image regions. Additionally, the Chirplet Transformer Feature Layer (CTFL) is applied after the Flatten layer to enrich feature representation, capturing both optimal feature frequency and spatial information, thereby improving crack classification accuracy. The proposed model was evaluated on two publicly available datasets: the SDNET2018 multiclass dataset, achieving an accuracy of 92.67%, and the Pillow Dam Borehole CCTV image dataset (binary classification), achieving an accuracy of 94.44%. Compared to existing models, the proposed approach demonstrates a significant improvement in both accuracy and efficiency, showcasing its robustness and potential for real-world civil and geo-engineering applications such as infrastructure monitoring and defect detection.</p>