Fine-grained Retrieval Method Of Textile Image

There are a large category of textile images that have the characteristics high local feature repetition rate and complex background information. These types significant intra-class differences small inter-class differences, making it impossible to perform classification training. Making difficult for existing methods accurately retrieve images. To improve retrieval accuracy images, this paper defines multiple repeated fine-grained features in as image "feature components", extracts components" from image, fuses generate definition "fingerprints".We propose an method based on pre-trained Mask R-CNN model extract then depth again through convolution neural network, fuse extracted obtain "fingerprint" textile. The obtained can effectively eliminate interference area number efficiency retrieval. A series comparative experiments carried out data sets with features. experimental results show proposed is generally effective.