Steel Sheet Counting From an Image With a Two-Stream Network

Steel sheets play a pivotal role in a wide range of industrial processes, including the production of ships and vehicles, as well as the construction of buildings and bridges. Meanwhile, counting steel sheets accurately is essential for effective production management in factories. However, manual counting of large numbers of stacked steel sheets can lead to visual vertigo, resulting in inaccurate counts. Moreover, physical methods such as weighing are also labor-intensive and inconvenient. Fortunately, advancements in computer vision technology have opened up new possibilities for efficient steel sheet counting. Nevertheless, implementing an automatic counting method encounters challenges due to the limited texture features present in steel sheets. In this article, we present a novel approach to count steel sheets from a captured image. To the best of our knowledge, this is the pioneering work that addresses this problem using a computational approach. We make the following contributions. First, we construct a comprehensive steel sheet dataset that contains steel sheet images with corresponding manually annotated dots. Second, we propose a novel network, called TSNet, which effectively extracts features from both the RGB image and its gradient map for precise steel sheet counting. Third, we conduct extensive experiments to evaluate the effectiveness of our proposed method and demonstrate its superiority over carefully chosen baselines from state-of-the-art counting methods.