Abstract: Since the camouflaged target is highly similar to the background, it is easily confused by background features, making it difficult to distinguish boundary information and extract target features. The current mainstream Camouflaged Object Detection (COD) algorithm mainly studies the camouflage target itself and its boundaries, ignoring the relationship between the image background and the target, and the detection results are not ideal in complex scenes. To this end, in order to explore the potential connection between background and target, an algorithm for camouflaged target detection by mining boundaries and background was proposed. The algorithm consisted of five parts: the encoder processes the initial raw data; the boundary guidance framework extracts more refined boundary features through feature processing and feature mining; the background guidance framework explores more salient features through multi-scale convolution while being based on attention The hybrid attention module is designed to enhance the selection of background features; the information compensation module makes up for the detailed information lost during feature processing; the multi-task collaborative segmentation decoder efficiently fuses the features extracted from different tasks and modules and outputs the final prediction results. Experimental results on three widely used data sets are better than the other 15 state-of-the-art models, especially on CAMO data set, the average absolute error index drops to 0.042.

Key words: camouflaged object detection, reverse guidance, multi-scale convolution, attention mechanism, feature aggregation

摘要： 由于伪装目标与背景具有高度的相似性，极易受背景特征混淆，导致边界信息难以分辨和提取目标特征困难。目前主流的伪装目标检测（COD）算法主要针对伪装目标本身及其边界进行研究，忽略了图像背景与目标的相互关系，在复杂场景下检测结果不理想。为此，为了探索背景和目标的潜在联系，提出一种通过挖掘边界和背景进行伪装目标检测的算法。该算法由5个部分组成：编码器处理初始原始数据；边界指导框架通过特征处理和特征挖掘提取更多精细的边界特征；背景引导框架通过多尺度卷积探索更多的显著特征，同时基于注意力设计了混合注意力模块增强对背景特征的强化选择；信息补偿模块弥补了在特征处理过程中损失的细节信息；多任务协同分割解码器则对不同任务以及模块所提取的特征进行高效融合并输出最终的预测结果。在广泛使用的3个数据集上的实验结果优于其他15个较先进的模型，尤其在CAMO数据集上的平均绝对误差指标降到了0.042。

关键词: 伪装目标检测, 反向引导, 多尺度卷积, 注意力机制, 特征聚合