| Camouflage is an attempt to hide foreground objects in background texture,and many animals in nature use camouflage to avoid attracting the attention of predators.This mechanism is also widely used in real life,such as polyp segmentation,lung infection segmentation,etc.Recently,detecting and segmenting camouflaged object from image,namely Camouflaged Object Detection(COD),has attracted wide attention in the field of computer vision.Since the camouflage object is highly similar to the background and even blends into one,it is very challenging to detect and segment the camouflage object.The main research work of this thesis is as follows:(1)Aiming at the problem that the camouflage object is highly similar to the background and the background information of the camouflage object is not fully utilized,it is difficult to detect the camouflage object,this thesis proposes a camouflage object detection algorithm based on a two-stage generative enhanced neural network.By alternately updating the training segmentor and the reconstructor,the performance of the segmentor is continuously improved,and the training results of the reconstructor promote the training of the segmentor and the improvement of segmentation performance.Experiments prove that the algorithm model has better detection performance than other camouflage object detection models.(2)Aiming at the problem that the difference between camouflage object and background is reduced due to image frequency information loss and insufficient camouflage object and prior information,this thesis proposes a camouflage object detection algorithm based on frequency domain and edge enhancement neural network,which can not only effectively extract camouflage object frequency feature information in the image,but also use the camouflage object edge prior information to assist in restoring the camouflage object edge contour,thereby improving the performance of camouflage object detection algorithm.Experiments show that the algorithm model has better detection performance than other camouflage object detection models. |
