A convolutional Riemannian texture model with differential entropic active contours for unsupervised pest detection

Pest camouflages in grains or natural environment cause significant difficulties in pest detection using imaging technologies. This paper proposes a convolutional Riemannian texture with differential entropic active contours to distinguish the background regions and expose pest regions. An image texture model is firstly introduced on the Riemannian manifold. A convolutional Riemannian texture structure is then explored to reduce the environmental background textures and highlight potential pest textures. Subsequently, a differential entropic active contour model is developed to estimate the foreground and background distributions. Finally, the estimated foreground and background distributions are used to distinguish pest textures and environmental textures. The final detected regions are obtained by maximizing pixel-wise posterior probabilities on the estimated distributions. Experimental results show that effective detections can be achieved by the proposed method on forestry pests imaging datasets.