Deep learning computer vision for robotic disassembly and servicing applications

Fastener detection is a necessary step for computer vision (CV) based robotic disassembly and servicing applications. Deep learning (DL) provides robust approach creating CV models capable of generalizing to diverse visual environments. Such DL systems rely on tuning input resolution mini-batch size parameters fit the needs application. This paper method determining optimal compromise between determine highest performance cross-recessed screw (CRS) while utilizing maximum graphics processing unit resources. The Tiny-You Only Look Once v2 (Tiny-YOLO v2) object system was chosen evaluate this method. Tiny-YOLO employed solve specialized task detecting CRS which are highly common in electronic devices. used meant lay ground-work multi-class fastener detection, as not dependent type or number classes. An original dataset 900 images 12.3 MPx manually collected annotated training. Three additional distinct datasets 90 each were testing. It found an 1664 x pixels paired with 16 yielded average precision (AP) among seven tested all three testing datasets. model scored AP 92.60% first dataset, 99.20% second 98.39% third dataset.