Soldered Dots Detection of Automobile Door Panels based on Faster R-CNN Model

Abstract

This paper mainly introduces a Faster R-CNN algorithm to identify the soldered dots of automobile door panels. It compares the effect of state-of-the-art algorithm on identification of soldered dots and proposes a better calculating method for the project. The dataset in this research was a collection of more than 800 unwelded inner door panels collected by cameras, including 500 photos of entire door panels and 300 partials. Detection of soldered dots is based on Faster R-CNN, extracting the feature maps from image via the VGG16 convolution neural network. Region proposal networks generate region proposals for feature maps. ROI pooling extracts proposal feature maps from the input feature maps and proposals. The fully connected layer utilizes the proposal feature maps to calculate the classes of proposals, while the bounding box regressor obtains the exact location of predicted boxes. The experimental results show that both Faster R-CNN and YOLOv3 can be applied to small soldered dots detection. YOLOv3 has a faster speed and it is suitable for real-time detection. Faster R-CNN has high detection accuracy and it is suitable for non-real-time high-precision detection. In addition, Faster R-CNN is more generalized which should be given priority in complex scenarios.