An Efficient Neural Network for Shape From Focus

Abstract

Three-dimensional (3D) imaging plays an important role in today’s 3D era. As it increases to use 3D technology, the necessity for obtaining precise 3D depth map rises gradually. Especially, shape recovery from focus (SFF) technique is one of the passive optical methods in 3D image area, and it has been actively utilized in many industrial or consumer applications. In SFF technique, a sequence of images of an object is captured by using a single camera with multi-focus levels in fixed finite steps. The shape of the object can be acquired by applying a focus measure operator, which computes focus quality for each pixel with considering its local area, in the image sequence. The focus measured value for each pixel is various with each frame, and a frame having maximum focus measure is regarded as the depth map for each pixel. To refine this initial depth map, additional approximation or machine learning technique can be utilized. Especially, artificial neural network has been successfully used to refine initial depth map for generating more accurate depth map, however, it suffers from high complexity and low convergence rate. Hence, it needs to solve the problems to apply the technique to real industrial or consumer applications. This dissertation consists of three parts which deal with artificial neural network for SFF. The first part presents a neural network model simplified by using dimension reduction. And the second part studies various initial weight setting methods for solving the convergence problem of neural network. Lastly, the third part discusses network parameter setting for obtaining appropriate learning rate parameter. In the first part, we discuss an efficient neural network model for shape from focus. The neural network model is simplified by reducing the input data dimensions and eliminating the redundancies in the conventional model. It helps for decreasing computational complexity without compromising on accuracy. And the representation of the conventional neural network model for SFF is modified into a vectorized expression for readability and usability. The proposed model is evaluated using various image sequences, and experimental results demonstrate that the proposed model is considerably efficient while the accuracy is comparable with the existing systems. In the second part, we discuss various initial weight setting methods in order to increase the convergence rate and efficiency. Weight passing (WP) method selects appropriate initial weights for the first pixel randomly from the neighborhood of the reference depth and it chooses the initial weights for the next pixel by passing the updated weights from the present pixel. This method not only expedites the convergence rate, but also is effective in avoiding the local minimization problem. Moreover, it may also be applied to neural networks with diverse configurations for better depth maps. The weight selection by lookup table (LUT) method is also proposed. A lookup table is compiled in the pre-processing stage and it is utilized in the network learning steps for generating proper initial weights. This method has increased accuracy of 3D restored depth especially when the depth is a high value in depth range. In the third part, we discuss performance pre-checking (PP) method for obtaining an appropriate network learning rate parameter. It selects a proper learning rate by pre-checking the performance between subsets of the reference and the restored depth map from network. The proper learning rate from this method can decrease the training time by reducing the total iteration number while keeping the learning process stable.