Traffic sign recognition – 03 – algorithms in object detection (3)

Faster R-CNN (2016)

This algorithm was proposed by Shaoqing Ren and colleagues in 2016 to continue improve speed and detection object in the article https://arxiv.org/abs/1506.01497

This bring a highest quality in both detect object and classify object in ILSVRC-2015 and MS COCO-2015.

This architecture is designed to propose and verify region proposals as a part of training process, call Region Proposal Network (RPN). This part will be used with Fast R-CNN in a single model. This improvement will reduce number of region proposal, also increase training speed with highest quality. Speed is 5fps per CPU

Although is a single model, but it includes 2 modules:

• Region Proposal Network (RPN): CNN Network to propose region and objects in the region
• Fast R-CNN: CNN network to extract features from region proposal and return bounding boxs and labels

Both modules working on output of deep learning network. RPN network works as a attention for Fast R-CNN, notice where it should look

RPN get a pretrained deep CNN, (VGG-16) then transfer feature map into a small network and return region proposals and labels. Region proposals are bounding boxes, base on Anchor boxes or shape to increase speed. Predict label will return region proposal exists in object or not.

You need to login in order to like this post: click here