Sketch recognition with self-supervised deep learning

Sketches play an important role in people's ability to record and convey their thoughts. However, even the simplest object drawings are not the same due to people's differences in perception, different experiences, the tools they draw with and many other factors. There are many studies in the literature in the field of sketch recognition with artificial intelligence. With the development of deep learning models, studies in these areas have differentiated and big datasets have been constructed and developed. Within the scope of this thesis, a state-of-the-art method that uses a fusion of Convolutional Neural Networks (CNN) and Textual Convolutional Networks (TCN) with having Self-Supervised structure on Quick, Draw! dataset, has been examined. To increase the Self-Supervised sketch recognition performance, studies have been carried out as using additional features on the data and using different fusion methods. During the entire study, pretext tasks that are designed to help with the difficulty of obtaining labeled samples and with speeding up the training processes were used. In the thesis, (1) adding the "distance between points" and "angle between points" as additional features to the image and stroke data to increase the sketch recognition performance; (2) expressing the sketches with Bézier curves to reduce noise, and (3) weighting with convex and concave functions instead of the existing linear weighted fusion have been proposed, respectively. The proposed applications were first carried out as a preliminary study on a small subset of the Quick, Draw! dataset with 414 thousand samples, and the results and inferences are given. These recommendations were subsequently reflected in a larger set of 3.8 million people. In comparative analysis, it has been observed that adding distance features and transforming the lines into Bézier curves significantly improves the predictions for certain classes. In the proposed fusion strategies, it has been seen that concave functions may be an area that can be studied to improve drawing recognition performance.