CTFU-Net:CNN-Transformer Fusion U-shaped Network for Moving Object Detection

The technique of extracting changing areas or moving objects from image sequences is known as moving object detection. The two most popular deep learning frameworks available now are convolutional neural network (CNN) and Transformer. CNN’s strong local feature extraction abilities enable it to effectively capture local semantic information, while Transformers excel in modeling long-range semantic information. Consequently, we propose an innovative method for detecting moving objects by leveraging the benefits of both CNN and Transformer: CNN-Transformer fusion U-shaped network (CTFU-Net). The main concept is to achieve information sharing between two different types of branches by hierarchically fusing the local feature information and the global context information through the feature interaction module (FIM). Furthermore, we also present a hybrid loss function to alleviate the foreground and background class imbalance problem. The experimental results demonstrate that CTFU-Net can extract more complete and clearer moving objects with better visual performance compared to the state-of-the-art methods.