Multi-Scale Feature Fusion for Improved Image Segmentation

In order to effectively utilize the feature maps of different scales and their contextual relationships, enrich the feature information, and improve the performance of the model, this paper incorporates the multi-scale feature fusion method into the decoder of the proposed segmentation network. The feature pyramid network[42] consists of a bottom-up downsampling path and a top-down feature fusion path, which can take into account both shallow features and abstract semantics, as well as small and large target features in the scene.

Therefore, the decoder in this paper adds the feature pyramid method for multi-scale feature fusion, as illustrated in figure 3. The feature fusion method used by the network decoder proposed in this paper is the concatenate method. Concatenate is a feature fusion method that merges feature images of the same size in the channel direction. This method increases the dimension of image features and retains more original information. The formula for the concatenate method is shown as follows: