Lightweight convolutional neural network for counting densely piled steel bars

Abstract

Steel bars are widely used in the construction industry to provide additional strength to modern building structures. Manually counting piled steel bars at construction sites is an important, but tedious, laborious, time-consuming, and error-prone task. A portable device, which can be used to count the bars automatically and accurately, is a potential solution. Deep learning has been overwhelmingly successful in doing this type of tasks. However, most deep learning models are large and compute intensive, therefore, they are not suitable to be used in resource constrained portable devices. This paper proposed an innovative lightweight Convolutional Neural Network (CNN) model, named DPSBC-Net (Densely Piled Steel Bar Counting-Net), which can be used for fast and accurate counting of densely piled steel bars, and can be potentially deployed to a mobile device. We proposed a CBAMDenseCSP block and a new method for relative resolution object scale measurement, which can reduce the number of parameters and computation cost of the proposed model by 76.86% and 44.82%, respectively. To further improve the accuracy of the steel bar counting, we also proposed an object density measurement method, and a detection head to balance positive and negative samples, which can significantly improve the performance of the proposed model. Experimental results show that our proposed model is not only faster, but also more accurate than state-of-the-art methods for counting densely piled steel bars. The proposed model can be used to process up to 66 images per second at a resolution of 576 × 768 and achieve an F1 score of 99.25%. To facilitate the future research on densely piled steel bar counting, our datasets have been made publicly available at https://github.com/LittleBoy1992/Steel-bar-Dataset.