Localization, tracking and segmentation of wire rods in a rolling mill: a computer vision-based approach.

Steel wire rods are a key product of the steel industry with multiple applications such as springs, bearings, bolts and several others. To guarantee high quality standards while maintaining high production throughput and competitive costs, wire rod mills need significant levels of process automation. The rapid development in deep learning, especially applied to computer vision, offers a wide range of tools to design and build solutions to cope with this necessity, following the industry 4.0 revolution. This thesis will address the application of computer vision techniques to localize, track and segment steel wire rods during the cooling phase over the roller conveyor. Different solutions will be tested: methods based on applying object detection in each frame of the video, using a Siamese neural network to process pairs of frames, and solutions built on background removal with the Template Matching algorithm. This last class of approaches will prove to be the most reliable for the task faced. To isolate the foreground object, both a more traditional background subtraction method and a segmentation neural network will be involved.