An Improved Automatic Classification Method for Artistic Design Image Styles Using Convolutional Neural Networks

Abstract

Due to the lack of effective decoupling between style features and content features, the fine-grained style recognition accuracy of general-purpose CNN (Convolutional Neural Network) in art and design image style classification is limited. In this paper, it proposes an improved CNN method with multi-scale dilated convolutions and Squeeze-and-Excitation (SE). The ResNet-50 backbone network extracts the features of input images, which are taken as the input of the multi-scale dilated convolution parallel module. The parallel module has four 3×3 dilated convolution branches with dilation rate 1, 2, 3, and 4. The feature maps of four branches are concatenated in channel dimension respectively, then extract the brushstroke textures and compositional patterns of different scales. Finally, the concatenation feature enters the SE module, global average pooling compresses each channel into a scalar, and the two fully connected layers produce channel recalibrated weights, and each weight is multiplied by the original feature map, so that the channel with strong style discrimination is enhanced. At last, the global average pooling reduces the dimensionality of weighted feature, and output style probability distribution through fully connected layer and Softmax. Experimental results show that compared with standard ResNet-50, the accuracy of classification of art and design image style achieves 92.56%, and macro-average F1 value is 91.78%, which are 5.11 and 5.66 percentage points, respectively. This provides an effective deep network improvement scheme for automatic classification of art and design image styles.