Camouflaged Object Segmentation Using a U-Net Based Deep Learning Approach

Abstract

The paper focuses on the computer vision task of camouflaged object detection that has various applications, including, but not limited to, military, security, biology, and industry. From the perspective of machine vision, camouflage is a state in which the target’s texture, colour, and pattern are indistinguishable from the surrounding background, making it hard, if not impossible, to isolate those objects from the background. It implies that the COD problem statement extends far beyond the scope of ordinary object detection. To tackle the COD task, the authors used the 2D U-Net model that demonstrated its ability to address image segmentation problems through employing a contracting encoder and expansive decoder connected with skip links. In particular, the U-Net architecture was utilised due to its ability to maintain a high level of detail in an image, which is critical during camouflage detection, in contrast to typical bounding box estimation used in object detection tasks. The model was trained on MCSIK training sets and evaluated on MCSIK test sets, with an average accuracy of 70.14%. Overall, the U-Net architecture is promising for addressing camouflage object detection; however, there is a noticeable accuracy drop on real-world sensor data, which the paper tries to address further.