Application of Deep Convolutional Features in Sketch Works Classification and Evaluation

We apply deep convolutional features in the classification and evaluation of sketch works which are collected from the sketch teaching scenario. Firstly, we test the classification performance of sketch works using deep convolutional features, and convert the evaluation task of sketch works to the fine-grained classification task（four classes: best, good, moderate and worst） based on images’ high-level semantic features（e.g. shape, composition, texture, proportion of black, white or gray and so on）; secondly, we propose the bilinear convolutional neural network（CNN） model which is particularly useful for fine-grained categorization classification; finally, we use Tensor Sketch Projection to compact the bilinear CNN features and apply end-to-end training to fine-tune the model. The results demonstrate that: CNN features significantly outperform the traditional features（e.g. Histogram, LBP and SIFT） in sketch works classification task; Compact bilinear CNN features can achieve the similar results under lower dimension in sketch works evaluation task.