Architecture of a configurable 2-D adaptive filter used for small object detection and digital image processing

Abstract

A configurable 2-D adaptive digital filter architecture based on 2-D normalized least mean square (TDNLMS) algorithms is proposed. This architecture can be used as a prewhitening filter to enhance the detectability of small objects in digital images. Every tap in this architecture can be configured as one of three types: input, reference, and disable taps. By setting the type of each tap, this architecture can be utilized to implement a TDNLMS adaptive prediction error filter in any type of support region on a 5x5 tap array. This feature allows the adaptive prediction filter to detect small objects whose area changes gradually during the detection process, or to process images with different stochastic characteristics. This architecture can also be configured as an adaptive prediction filter (APF) used in image noise reduction, or a 2-D convolver used in many image processing algorithms. In the calculation of coefficient updating, the power of the input vector is approximate to the power of two, hence the critical path delay is reduced significantly. A local parallel global serial mode is adopted in this architecture to increase the coefficients' word length, and to reduce the finite-precision effect. Simulation results show that TDNLMS APEF based on this architecture can provide satisfactory small object detection performance in various environments.