基于多尺度渐进式特征融合的伪装目标检测

Camouflage Object Detection Based on Multi-scale Progressive Feature Fusion

伪装目标检测任务中目标和所处环境具有极大的相似性,只通过简单的特征提取容易使得重要信息丢失,此外,直接聚合不同层特征会引入噪声,导致预测不准确。针对这些问题,本文提出一种基于多尺度渐进式特征融合的伪装目标检测网络。该网络采用金字塔视觉Transformer作为骨干网络提取多尺度特征,利用可变形注意力对提取的多尺度特征增强,以突出伪装目标的边界;然后通过渐进式特征融合模块递增式融合相邻层特征,积累难以分辨但有效的信息,避免非相邻层之间较大语义差距;在融合过程中引入自适应空间融合操作,以减少在同一空间位置发生信息冲突问题;最终输出预测结果,实现伪装目标检测。模型在COD10K和CAMO组成的训练集上进行训练,实验结果表明本文方法与其他方法相比具有较明显优势。

The target and the environment in the camouflage object detection task exhibit great similarity.Important information is easily lost through simple feature extraction.Additionally,the direct aggregation of features from different layers introduces noise and inaccurate predictions.To address these issues,a camouflage object detection network is proposed in this paper based on multi-scale pro-gressive feature fusion.A pyramid vision Transformer is employed as the backbone network to extract multiscale features and the deformable attention is utilized to enhance the extracted multiscale features emphasizing the boundaries of the camouflaged target.Subse-quently,neighboring layers of features are incrementally fused by a progressive feature fusion module to accumulate indistinguishable but effective information and avoid large semantic gaps between non-neighboring layers.An adaptive spatial fusion operation is intro-duced in the fusion process to reduce information conflict at the same spatial location.Finally,the prediction results are output to achieve camouflage object detection.The model is trained on the training set consisting of COD10K and CAMO,and the experimental results demonstrate that the method presented in this paper has more obvious advantages compared with other methods.