基 于 特 征 聚 合 与 边 缘 检 测 的 伪 装 目 标 检 测

Camouflaged Object Detection（COD）holds significant research and application value in various fields. The ability of deep learning is pushing the performance of target detection algorithms to new heights. Designing a network that effectively integrates features of different layer sizes and eliminates background noise while preserving detailed information presents the main challenges in this field. We propose Feature Fusion and Edge Detection Net（F2-EDNet），a camouflaged object segmentation model based on feature fusion and edge detection. ConvNeXt is used as the backbone to extract multi-scale contextual features. The extensiveness and diversity of features are then enhanced through two approaches. The first approach involves using the Feature Enhancement Module（FEM） to refine and downsize the multi-scale contextual features. The second approach introduces an auxiliary task to fuse cross-layer features through the Cross-layer Guided Edge prediction Branch（CGEB）. The process extracts edge features and predicts edge information to increase feature diversity. Additionally，the Multiscale Feature Aggregation Module（MFAM）improves feature fusion by capturing and fusing information about interlayer differences between edge features and contextual features through multiscale attention and feature cascading. The model's prediction results are subjected to deep supervision to obtain the final target detection results. To validate the performance of the proposed model，it is compared qualitatively and quantitatively with eight camouflage object models from the past three years on three publicly available datasets. This comparison aims to observe its detection accuracy. Additionally，a model efficiency analysis is conducted by comparing it with five open-source models. Finally，the module's effectiveness is verified through ablation experiments to determine the optimal structure. The results of a quantitative experiment indicate that on the CAMO dataset，the S-measure，Fmeasure，E-measure correlation and mean absolute error metrics for F2-EDNet are optimal. On the COD10K dataset，the structural similarity metric indicates that the proposed algorithm is optimal，while the mean precision and recall，E-measure and MAEmetrics reach sub-optimal levels. On NC4K，all four metrics for the proposed algorithm reach optimization. From the visualized detection results，it can be observed that in the camouflage object detection task，the prediction results of the proposed model are more accurate and refined than those of other methods. Compared with other models，although the number of parameters in the proposed model is higher，the simple structure of the model framework enables it to outperform models specifically designed for lightweight purposes，faster than most other models. In comparison of the number of operations，the arithmetic complexity of the proposed model shows a significant decrease compared to a model that also utilizes multi-task learning. The model presented maintains high accuracy in target detection performance while ensuring a reasonable balance between computing speed and the number of operations. The results of ablation experiments demonstrate that each of the current modules plays the expected role，and the model's performance has been optimized. Experimental results show that the proposed algorithm achieves optimal detection accuracy. Compared to suboptimal models，our model demonstrates an average improvement of 1.41%，1.74%，0.14%，and 0.77% on the S-measure，F-measure，MAE，and E-measure indices across three datasets. Additionally，the model's design achieves a reasonable balance between operation volume and operation rate. During performance testing，the model's test speed was 46 fps，striking a balance between detection accuracy and execution efficiency，demonstrating practical application value. In future work，the algorithms will be lightened to further reduce the amount of computation to improve the speed of model inference；in applications，the model can be helpful in directions such as medical segmentation，defect detection with transparent object segmentation through migration learning.