Adaptive morphological and bilateral filtering with ensemble convolutional neural network for pose-invariant face recognition

Abstract

Based on behavioural or physical characteristics, humans are recognized by using biometric system. In computer vision and pattern recognition domain, dynamic research is going on in face recognition. Face recognition algorithms are challenged by intra-personal changes in pose, illumination, and expression (PIE). Images are processed and matched with various databases. For face recognition, multi-task learning (MTL) is explored in this work. However, recognition of faces from blur and poor illumination becomes difficult. Recovering face from mixed noise degradation is a challenging and promising theme. This work explores an ensample convolutional neural network (ECNN) for face recognition. Initially a new adaptive morphological bilateral filtering (AMBF) method is proposed. Without introducing undershoot or overshoot, slope of edges is increased for sharpening a blur image. Quality sharpening enhancement is assured by various morphological operations like closing, opening, erosion and dilation with proper size of structure element. In addition to adaptive bilateral filter, mathematical morphology operations are included to enhance the performance. Then a multi-task ECNN is implemented for a main classification task and estimation of pose, blur, illumination, and expression (PBIE) as side tasks. For every side task, loss weights are assigned automatically by developing bat algorithm (BA) based dynamic-weighing method. In multi-task ECNN, balance between various tasks are achieved. Hence, proposed method is effectively demonstrated by the results of experimentation on entire multi-PIE dataset.