Enhanced Mask R-CNN blur instance segmentation based on generated adversarial network

Abstract

Aiming at the detection of blur SEM nanoparticle images, a blur instance segmentation network (BL-Mask R-CNN) based on generative adversarial network deblurring convolution block (Deblur) and Mask R-CNN instance segmentation algorithm is proposed. The network uses the enhanced method of the blind deblurring algorithm (DeblurGAN-v2) of the generative confrontation network to preprocess the image, restore the grainy texture details in the image and generate a clear image that is conducive to instance segmentation and detection. The experimental results show that this method effectively improves the detection accuracy of blurred SEM nanoparticle images. Tested on the blur images of the NFFA-EUROPE dataset, the improved BL-Mask R-CNN has compared the detection accuracy of the original Mask R-CNN. The accuracy rate AP has increased from 0.8339 to 0.9613, and achieved good results in the test.