Research On Strategies For Sketch Recognition

Sketch has a long history, it is a universal form to communicate. Humans have used sketch to describe what they saw since prehistoric times. Even today, sketching is the technique mastered by all humans to record information. In past decades, Scholars have done lots of researches in the field of graphics. However, there has never been a formal study of how humans sketch objects and how well such sketches can be recognized by computers. The study group leaded by Professor Eitz of Berlin Institute of Technology had the first large scale exploration of human sketches and proposed an algorithm of sketch recognition in 2012.They collected the first dataset and used it to evaluate the algorithm. This dataset contains the sketches of everyday objects evenly distributed over 250 object categories such as’car’or’teapot’. Many new applications could benefit from the new work such as human-computer interaction、sketch-based synthesis、sketch beautification and sketch-based retrieval. Study on this field has just started and substantial results are rare. To overcome the weakness in Eitz algorithm, we propose several improved strategies. Specific work is as follows:(1)We carry out specific discussions on the techniques related to sketch recognition. In addition, we make a detailed analysis of the Eitz algorithm and expound the advantages and disadvantages of the algorithm.(2) We propose a new algorithm of sketch recognition using location information. In Eitz algorithm, neither spatial location information nor effective selection of kernel functions for classification is considered. To overcome these drawbacks, we propose a hog-like feature extraction algorithm using location information. In classification process, we employ SVM classifiers with histogram intersection function as the kernel function. The experiments show the proposed algorithm reduces the computational complexity by extracting fewer local features with lower dimensions, and improves classification accuracy in most cases.(3)We propose a new algorithm of sketch recognition using context information. In feature extraction process, Eitz algorithm does not encode the context relations among the local features. To overcome this defect, we add the multi-direction context information to the features extracted above. The proposed algorithm aims to make the features be more semantic. In the process of feature representation, as the limitation of the traditional bag-of-features representation, we introduce a coding scheme called Locality-constrained Linear Coding as a representation of the sketch features. In classification process, we employ the linear SVM classifier. The experiments show the proposed algorithm performs better than the Eitz algorithm, achieving higher recognition rate while using less time.