Sketch recognition method based on double-layer structure

Abstract

The invention discloses a sketch recognition method based on a double-layer substructure, and aims at the current situation that most sketch recognition methods do not consider shape features at present, the method proposes to use the double-layer substructure to simultaneously encode the shape features and visual features for sketch recognition. Comprising the following steps: acquiring sketch samples in two formats, constructing a multi-level shape network based on a hierarchical idea of extracting depth features by a convolutional neural network, and constructing a multi-level visual network by constructing a multi-scale residual block, an inner-layer residual block and an outer-layer residual block. Based on sketch samples, a shape network is trained by using cross entropy loss, and a visual network is trained by using cross entropy loss and weight compression triple center loss. And combining the shape network, the visual network and the multiplication fusion layer to obtain a double-layer structure. And finally, inputting the sketch test sample into the double-layer structure for identification test. According to the method, the sketch sample is not required to contain stroke information, a fine tuning process is not needed, the training process is simple, and the sketch recognition effect has obvious advantages.

Description

technical field [0001] The invention belongs to the field of computer vision, in particular to a sketch recognition method based on a double-level structure. Background technique [0002] Obviously different from images, sketches are highly abstract and usually only contain outline information of objects and some simple details. Therefore, it is more reasonable to express the sketch content with shape features and visual features. However, few previous research efforts have considered these two key features simultaneously. Related research methods mainly include manual feature method and deep learning method. Among them, the manual feature method is mainly based on the key technology of image recognition, designing manual features and combining the aggregation representation method for local features to generate sketch visual features, and finally training the classifier for classification and recognition. The deep learning method usually combines the characteristics or p...

AI Technical Summary

Problems solved by technology

2. Lack of description of shape features

3. The training method and process are more cumbersome

4. It is greatly affected by noise samples

Images