The Research Sketch-Based Object Detection And Retrieval

With the popularity of touch screens in recent years,touch screen technology has been applied to mobile phones and laptops,and even some self-service devices in banks and communication service halls.The popularity of touch screens has made the interaction between people and computers more convenient and more diversified.If the computer can understand the sketches drawn by people,the interaction between human and computer will be more convenient and efficient.Because the drawn graph contains many details and geometric information,the drawn graph usually carries more information than the written text in the same size area.Currently,due to the lack of sketch data sets(the largest sketch data set is the TUBerlin data set),and most of the sketches lack of color information and texture information relative to real pictures,many researches on sketch using deep learning methods focus on the classification of sketches.This paper studies the object detection and sketch-based retrieval,and our mainly work completes the following tasks: 1.collected a sketch data set containing over 9900 sketches of 20 categories for recognition based on sketches,and manually labeled each sketch in xml format.The annotation content includes the category and the location of the objects.2.Modified the input layer of a deep convolutional neural network based on R-CNN,optimized the initialization parameters,adapted it to the single-channel sketch map data set,and trained and tested the performance of object detection on the data set.3.Proposed a twin network structure based on R-CNN.The training set is encoded into a set of triples,the pre-trained twinning network is used to extract the high-dimensional semantic features of the triples,and finally achieved an End-to-End task of sketch-based(PASCAL VOC2007 data set)cross-domain retrieval.