International Conference on Computer and Knowledge Engineering

Home / 14th International Conference on Computer and Knowledge Engineering

Camouflage Object Segmentation with Attention-Guided Pix2Pix and Boundary Awareness

Authors :

Erfan Akbarnezhad Sany¹ Fatemeh Naserizadeh² Parsa Sinichi³ Seyyed Abed Hosseini⁴

1- Faculty of Computer Engineering, Quchan University of Technology, Quchan, Iran. 2- Faculty of Electrical & Computer Engineering, Malek Ashtar University of Technology 3- Faculty of Computer Engineering, Quchan University of Technology, Quchan, Iran. 4- Department of Electrical Engineering, Mashhad Branch, Islamic Azad University, Mashhad, Iran.

Keywords :

Camouflage،Object segmentation،Pix2Pix،Aware –boundary

Abstract :

Segmentation of camouflaged objects presents a significant challenge in computer vision due to their inherent blending with the background environment. This paper proposes an innovative method for precise camouflage segmentation using an advanced Pix2Pix architecture. The proposed method leverages the VGG16 network as a backbone for extracting robust features. Subsequently, convolutional block attention modules are integrated to refine these features by focusing attention on spatial and channel dimensions. The generator's decoding pathway employs upsampling stages and skips connections to reconstruct high-resolution segmentation maps effectively. Simultaneously, a discriminator network differentiates between ground truth segmentation maps and the generated ones. To optimize the model's performance, a combination of adversarial, L1, and Soft Dice losses was used during training. Here, the Dice Score loss function was employed twice. Firstly, it was used to quantify the disparity between the ground truth mask and the mask predicted by the generator. Secondly, it was applied to assess the difference between the boundary of the ground truth mask and the boundary of the predicted region extracted by morphological techniques. Comprehensive evaluations of the CAMO dataset demonstrate the effectiveness of the proposed method. This method achieves a pixel accuracy of 85.4054%, intersection over union of 43.9517, F-measure of 58.5606, and mean absolute error of 0.1307. These results highlight advancements in developing robust computer vision systems capable of identifying and understanding camouflage strategies in real-world scenarios.