Optimizing Non-Differentiable Metrics for Hashing

Abstract

Image hashing embeds the image to binary codes which can boost the efficiency of approximately nearest neighbors search. F-measure is a widely-used metric for evaluating the performance of hashing methods. However, it is non-differentiable and hence it has not been used as an object function for hashing. Heuristic algorithms, e.g. evolutionary computation and particle swarm optimization (PSO), are good at optimizing non-differentiable objectives, while they are inefficient in very high-dimensional variables which are commonly used in hashing models. To address this contradict, we propose a scheme to bridge hashing methods and F-measure objective using PSO. The hashing methods are used to generate real-valued codes for images and then the parameters of quantization procedure are optimized by PSO. Our scheme can incorporate a wide range of hashing methods, heuristic optimization algorithms and non-differentiable metrics. Experimental results demonstrate that our scheme can be used to further improve the performance of existing hashing methods.