Abstract:

Focused on the issue that existing sketch recognition methods based on deep learning still use ordinary convolution as the main method of sketch feature extraction,ignoring the sparsity characteristics of sketch objects,this paper proposes a sketch recognition model based on dilated convolution.This model combines the dilated convolution and ordinary convolution by using the dilated convolution’s characteristics of expanding the receptive field without increasing the number of effective units of the convolution kernel,to realize the preliminary extraction of the structural features of the sketch.Due to the sparsely sampled input signal of the dilated convolution,there is no correlation between the information obtained by the long-distance convolution,which will affect the classification result.Therefore,the model uses the dilated convolution and ordinary convolution to extract the input image features separately,and finally adds the feature output by the two different convolution methods in the channel dimension.This method not only takes advantage of the sparse sampling characteristics of the dilated convolution,but also makes full use of the advantages of multi-scale information from different convolution methods.Experimental results show that this model has achieved a recognition accuracy of 72.6% on the TU-Berlin SKetch dataset,indicating that it has certain advantages over the current mainstream sketch recognition methods.

Key words: dilated convolution, multi-scale fusion, sketch recognition, convolutional neural network, receptive filed