Highlight Detection and Removal Method Based on Bifurcated-CNN

Abstract

Many visual tasks of intelligent robots as object detection and tracking are very easily interfered by the specular highlights. Existing highlight detection and removal methods often suffer from low sensitivity in dealing with low saturation pixels and large-area highlight, and there are visual artifacts and information distortions in the compensated pixels. In this paper, we propose a novel two-stage convolution neural network (Bifurcated-CNN), to tackle the problem of specular highlight detection and removal on high-reflective materials. 1) The specular highlight features are extracted and removed in two stages from coarse to fine, so as to ensure the generation of diffuse images have no visual artifacts and information distortions. 2) We propose a bifurcated feature selection strategy (BFSS) to filter out the specular highlight features, which enhances the detection capability of our network. 3) Experimental results demonstrate that the proposed network is more effective to remove specular highlight with a 0.06 × 1e⁻² reduction in MSE, a 2.4831 improvement in PSNR, and almost the same in SSIM, compared with the state-of-the-arts. In addition, we build a large-scale transparent solid waste highlight dataset, which will help the solid waste sorting robots to solve the detection failure problems caused by specular highlight, and greatly contribute to accurate visual sorting in the plastic solid waste recycling industry. It is also helpful to evaluate and encourage new deep-learning methods. As we know, this is the first specular highlight dataset for the plastic solid waste recycling industry. The dataset will be made available at https://github.com/aobi12138/Bifurcated-CNN.