Depth-of-Field Region Detection and Recognition From a Single Image Using Adaptively Sampled Learning Representation

This study describes a network and its application methods for efficient detection and recognition of the depth-of-field(DoF) region blurred in the image by focusing and defocusing the camera. This approach uses a cross-correlation filter based on RGB color channels to efficiently extract DoF regions in images and construct a dataset for training in the convolutional neural network. A data pair corresponding to the image-DoF weight map is set using the data. The training data are from a DoF weight map extracted based on an image and cross-correlation filter. The loss function is modeled using the result of applying Gaussian derivatives of the image to improve the convergence rate efficiently in the network training phase. The DoF weight map obtained as a test result and proposed in this paper reliably extracted the DoF region in the input image. In addition, this study experimentally demonstrates that the proposed method can be used in various applications, such as non-photorealistic rendering, viewpoint tracking, object detection and recognition, optical character recognition, and adaptive sampling, that employ the user regions of interest.