Enhancement of receptive field using dilated convolution for camouflaged human segmentation in RGB images

Detection of camouflaged human intruder in RGB images is a challenging problem and is of much interest to defence forces and security agencies. Limited research has been undertaken in this area due to various reasons including non-availability of a suitable dataset. In this context, our work has the following contributions. First, the construction of a manually annotated dataset called Camouflaged human dataset (CHD3.5K) consisting of images of humans camouflaged in background. Second, using dilated convolution on image features extracted using deep neural network for the segmentation of camouflaged human intruder in an image using our own CHD3.5K dataset. We experimented using different kinds of loss functions and deduced that the use of hybrid loss function yielded superior results compared to other popular loss functions. The results obtained on CHD3.5K dataset have been bench-marked against twelve different state of the art baseline methods using three popular evaluation metrics. Qualitative and quantitative performance analysis on the CHD3.5K dataset has shown encouraging results. We used ablation analysis to highlight the importance of key elements of detection network.