Real time confidence estimation and multi-scale camouflage object detection

Due to the fact that objects with camouflage blend seamlessly into their surroundings, exhibiting a exceptional resemblance in terms of texture, color, shape, etc., and possessing boundaries that are indistinguishable from their environment, the detection of such camouflaged objects becomes significantly more challenging. In this research, we introduce a novel network called CRNet, which incorporates a detection network for identifying camouflage objects with a real-time network for estimating confidence. Unlike other networks, in the camouflage object detection network, we design MSF
²
M by capturing maps at different scales and using the relationship between their maps. In the real-time confidence estimation network, we use spatial adaptive normalization for the edge region to combine the semantic information from high latitude images embedded in downstream low dimensional features. The network’s outstanding performance, as described in the paper, is demonstrated through tests conducted on three standard datasets.