A Novel Learning Dictionary for Sparse Coding-Based Key Point Detection

Recently, a sparse coding-based key point detector (SCK) was proposed. An SCK shows very impressive performance compared with state-of-the-art key point detection methods on different challenging conditions, such as variations in scale, rotation, context, and nonuniform lighting. The rotational-invariant dictionary in the SCK is, however, manually generated using a time-consuming process of selecting a good seed dictionary and combining multiple versions of its rotated atoms. In this work, the process is automated using a novel duplet autoencoder structure, in which the weights between the input and the hidden layers are designed to embed a rotational-invariant dictionary. A set of loss functions is also proposed to enforce the learning process. A novel retinal image registration pipeline that best uses the new detector is also designed with thorough analysis for selection of different technologies. Extensive experiments on four challenging datasets have confirmed that SCK with the learned dictionary achieves state-of-the-art key point detection performance.