A key player within the lumber industry, located in the Antwerp region, reached out to us seeking our assistance in automating their plank counting process. Currently, this task relies on manual labor, with an employee dedicated to tallying the planks for every incoming lumber stack at their warehouse. Complicating matters, certain clients also require precise measurements of plank dimensions, known as 'vallende breedte' in Dutch. Clearly, this procedure is not only laboriously time-consuming but also vulnerable to human errors.

The overarching aim of this project is to improve efficiency through the integration of computer vision and deep learning. By automating the plank quantification process, we aim to eliminate the inefficiencies and inaccuracies tied to manual counting. Our solution seeks to empower the client with accurate, swift, and error-resistant results, marking a significant step forward in their operations.

In the project's initial phase, we undertook the meticulous task of curating an expansive dataset. This dataset comprised a diverse array of images, each meticulously annotated to demarcate individual planks within assortments of lumber stacks. Given the client's involvement with an extensive range of lumber variations, substantial emphasis was placed on the acquisition of a dataset that authentically mirrors the prevailing diversity of lumber types in real-world scenarios. This necessitated the deliberate inclusion of imagery featuring both painted and unpainted planks and planks of different widths against different backgrounds.

Subsequently, we developed a deep learning model to detect and count the number of planks within a given lumber stack. This model was trained on the aforementioned dataset, which was augmented to further enhance the model's robustness. The model was then deployed on a web application, which we developed to facilitate the client's interaction with the model. In the web application, we allow the client to upload an image of a lumber stack, after which the model will detect and count the number of planks within the stack. The client can then save the results to a database, which can be accessed at a later time. Additionally, the client can indicate if we were able to detect all planks within the stack. If not, the client can manually annotate the missing planks, which will be used to further train the model. An example of the web application's interface is shown below.

Want to have a chat about this project? Feel free to contact us here.