Camouflaged object detection method for multiscale and cross-domain generalization

Abstract

Camouflaged object detection is a challenging task in the field of computer vision, whose core lies in identifying targets that are highly similar to the background in color, texture, and morphology. Although deep learning has made some progress in this direction, existing models still face two major bottlenecks: first, camouflaged objects exhibit significant scale differences, making it difficult for single-scale models to handle targets of different sizes; second, models perform well on training data but show a significant decline in generalization performance in cross-scenario and cross-dataset tests. To address these issues, this paper proposes a camouflaged object detection framework that integrates a multi-scale adaptation mechanism and a generalization enhancement strategy. Methodologically, first, an encoder-decoder structure based on convolutional neural networks is constructed, and a scale-adaptive feature pyramid network is introduced. Through dynamic feature alignment and cross-level feature fusion, the model achieves adaptive processing of multiple input sizes. Second, a combined data augmentation scheme is designed, and combined with structured regularization and hierarchical transfer learning strategies, the model’s adaptability to unknown scenarios and the stability of feature representation are effectively improved. Systematic experiments on multiple public datasets show that the proposed method performs stably in multi-scale detection tasks, maintains high detection accuracy and low missed detection rate in cross-domain tests, and its overall performance outperforms current mainstream detection models. This study not only expands the method system of camouflaged object detection theoretically but also provides reliable and scalable technical support for target recognition in complex scenarios in practical applications.