Texture recognition method based on deep self-attention network and local feature coding

Abstract

The invention relates to a texture recognition method based on a deep self-attention network and local feature coding. The method comprises the following steps: designing a deep self-attention module with four stages according to characteristics of a texture image, merging local image blocks in the first three stages to increase a receptive field, and limiting a self-attention calculation in local space with a fixed size; in the last stage, canceling local image block merging, globally calculating self-attention, and obtaining a relation between local blocks; therefore, better extracting texture features of a local area, and keeping global features not lost. According to a PET network provided by the invention, texture information in a local area in an image is fully combined, and two-dimensional features output by a backbone network are remodeled into a three-dimensional feature map; block descriptors of multiple scales are densely sampled in a feature map through a moving window, and a group of multi-scale local representations is obtained; finally, local feature coding and fusion are carried out on the multi-scale block features, and a fixed-scale texture representation is generated for final classification.

Description

technical field [0001] The invention belongs to the technical field of texture classification and material classification, and in particular relates to a texture recognition method based on a deep self-attention network and local feature coding. Background technique [0002] In the classic texture recognition method, in the method based on the bag-of-words model, first use manual features (such as GLCM, LBP, LPQ) to extract features, and assign each descriptor to the closest visual word in the codebook, through statistical Visual word frequency or methods for aggregating residuals. With the rapid development of deep learning, convolutional neural network (CNN) is widely used to replace manual feature extraction, and then texture encoding strategy is adopted for final texture classification. [0003] Most existing methods such as FV-CNN(1), DeepTEN(2), DEP-NET(3), LSCTN(4), these methods usually perform texture-based encoding on the overall features extracted by CNN. In the...

AI Technical Summary

Problems solved by technology

[0005] (1) For the classic texture recognition method, its solution often relies on some image preprocessing, manual feature extraction and word bag model. This kind of method is far from meeting the current detection requirements due to its low detection performance.

Second, the method was not optimized using the deep learning framework;

[0006] (2) For similar deep learning methods, first, such methods usually use deep convolutional network (CNN) for deep feature extraction, although CNN has proved its powerful feature capture ability based on images such as targets and objects. , but the ability to extract texture features is limited

Second, in the texture image, the local area has a strong texture recognition ability. The existing methods ignore the texture coding combined with local features, which restricts the model's ability to recognize texture data.

[0015] Comparison of technical points: This article proposes to use a position-aware coding layer for position constraints, in which dictionaries and coding representations are learned simultaneously, but this method does not consider the significant role of local features on texture classification

Images