Application of Convolutional Neural Network in the Segmentation and Classification of High-Resolution Remote Sensing Images

Abstract

Numerous convolution neural networks increase accuracy of classification for remote sensing scene images at the expense of the models' space and time sophistication. This causes the model to run slowly and prevents the realization of a trade-off among model accuracy and running time. The loss of deep characteristics as the network gets deeper makes it impossible to retrieve the key aspects with a sample double branching structure, which is bad for classifying remote sensing scene photos. We suggest a dual branch inter feature dense fusion-based lightweight convolutional neural network to address this issue (BMDF-LCNN). In order to prevent the loss of shallow data due to network development, the network model can fully extricate the data from the current layer through 3 x 3 depthwise separable method is structured and 1 x 1 standard pooling layers, identity sections, and fusion with the extracted features out from preceding stage through 1 x 1 standard pooling layer.