Improving Camouflage Object Detection Using U-NET and VGG16 Deep Neural Networks and CBAM Attention Mechanism

Camouflaged object detection is an essential concern in computer vision and is commonly applied in military, security, biological, and various other fields. In machine vision, camouflage pertains to the situation where the desired object in an image is often challenging to discern or nearly impossible to see due to patterns or visual characteristics that closely resemble the background or the surrounding environment. Essentially, camouflage involves an attempt to blend the texture of a foreground object with that of the background. Concealing objects can be accomplished through methods like color matching, pattern blending, or structural adaptation. Detecting camouflaged objects presents a significant challenge because of the inherent resemblance of camouflaged objects to their background and the indistinct boundaries between them.In this article, a deep neural network-based approach is proposed to address this issue. By utilizing the U-NET and VGG16 algorithms in combination with the CBAM attention mechanism, a solution for detecting camouflaged objects in the CAMO and CAMO-COCO datasets is presented. Also, a data augmentation process is employed to enhance the training of the model. The proposed method demonstrates remarkable performance in detecting camouflaged objects in both the CAMO and CAMO-COCO datasets. On the CAMO dataset, it outperforms the comparison methods with an F-measure value of 0.5867, along with an IOU value of 0.4168 and an MAE value of 0.1186. Similarly, on the CAMO-COCO dataset, the proposed method achieves superior results with an IOU value of 0.7101 and an MAE value of 0.0813 and an F-measure value of 0.7703. These findings highlight the high accuracy and minimal error of the proposed method in detecting camouflaged object boundaries, reinforcing its superiority over the compared methods on both datasets.