Sketch recognition based on attention mechanism and improved residual network

Sketches are usually composed of simple strokes. Compared with the natural image, they lack the information of texture and color. Most of the existing works do not reduce the impact of blank region in sketches very well. This paper proposes a new deep convolutional neural network named Sketch Fusion Net (SFN) model to minimize the effect of the blank region and focus on information region. This model is mainly composed of Sketch Block Convolutional (SBConv) modules, the SBCnov module integrates an attention mechanism and an improved residual network. The first one can effectively extract the effective part of the sketch information. The other one uses multi-level feature combination strategy to extract richer semantic information, which can effectively relieve the problem of model degradation. Finally, two public datasets, namely the TU-Berlin and Sketchy, are used for sketch recognition experiments. The results demonstrate that this proposed method improves the recognition accuracy by 2.1% and 7.2% over several state-of-the-art methods and yields promising results.