In addition to serving as a means of transportation, a highway connects the local economy. But as the pavement ages, several flaws including cracks, potholes, and deformation gradually show up on the surfaces of the road. Crack evaluation is typically done by hand using field surveys performed by humans. Nevertheless, these labor-intensive, time-consuming, manual survey methods are not reproducible or repeatable and place surveyors in dangerous situations. This proposes a novel deep-learning framework for road crack detection based on multi-scale convolutional architecture. To address the challenges of object occlusion, soft non-maximal suppression (NMS) is applied across crack proposals at various feature scales. The experimental results show that the introduced model attains an accuracy of 0.9989 and better detection performance across four distinct object proposal building block (OPBB) architectures. Thus, our model outperforms the benchmark models in every instance of OPBBs. In this way, the problem of large object scale variation in road crack images with different object occlusion challenges has been tackled, and the positive images are divided into a set of semantically significant regions, such as the road surface and cracks.