Novo modelo de rede neural para detecção de objetos aplicado à inspeção industrial

Abstract

In many industries, assembling specific components to be inserted into a plastic container is a manual procedure. Each kit must be comprised of specific parts following a pre-defined recipe, which can be updated throughout time. Kits assembled inadequately cause rework, reducing production quality and time. Here we propose improvements in an object detection model, capable of performing a quality inspection, increasing the features of Few-Shot Object Detection (FSOD) based on the OS2D model, previously proposed in the literature. The OS2D model has limitations when trying to detect objects of aspect ratio that do not fit the predetermined anchors. In addition, it also has an inference mechanism that restricts it to only one reference image for each class, making it difficult to detect more complex objects, which are different for each angle. Bearing in mind this, the improved OS2D model (OS2D+) was proposed, incorporating layers of distortion and correction and modifying its inference strategy to facilitate the use of multiple reference images per component. In order to be able to evaluate the results of the OS2D+ solution, a image processing based solution (PIMG) was also developed, so the results of both solutions can be compared. The distortion and correction layers of the OS2D+ solution allowing it to detect objects whose aspect ratio does not fit into any OS2D detection anchor. The inference mechanism of the OS2D model has also been modified, seeking to enable the inference of multiple reference images to a kit component. Finally, the performances of both models were compared, in an attempt to evaluate whether the modifications proposed in the OS2D+ model improved the ability of the OS2D model of detecting objects. Finally, the proposed OS2D+solution proved to be more robust than PIMG, detecting fewer false positives and negatives, in addition to presenting a shorter inference time. However, the PIMG solution was able to provide better bounding boxes (BB) estimations due to its process of proposing locations. Despite this, the OS2D+ solution has the potential to have equivalent estimations, requiring a fine adjustment in its parameters. A database composed of 111 photos was also built, describing five different kits and their respective components. This database was annotated and used to measure the results of the two proposed solutions.