Enhanced Multiscale Vision Transformer with Cascaded Feature Fusion for Efficient Object Detection in Remote Sensing Images

Using the picture data acquired by remote sensing technology, remote sensing image processing detection classifies ground objects. It identifies changes to help determine the state of natural and human activity on the earth’s surface. Owing to various obstacles in various application scenarios and the management of large amounts of data, most remote sensing picture analyses today use techniques based on Deep Learning (DL) and Machine Learning (ML). By using vast quantities of sample training data, these techniques enable the automatic gathering of information about ground objects and minimize the need for operator involvement. Transformer-based models are not often used in feature extraction techniques due to the computational complexity that increases with the input’s length square and their incapability to extract Multi-scale Features (MSF). It is mainly for the more current transformer-based detectors, which are pretty effective at detecting objects but frequently come with significant and prohibitive additional computing costs. This paper designs an enhanced Multi-scale Vision Transformer (MSVT) to extract the MSF from the input. Cascaded Multi-scale Feature Fusion (CMSFF) is proposed to utilize the extracted MSF from the transformer-assisted object detectors to improve the feature dimensionality and detection accuracy. From the critical locations, the sparse scale features are extracted using MSVT and CMSFF and the transformer encoder and decoder are updated based on the Generative Adversarial Network (GAN) prediction. CMSFF obtained the refined detection of the locations from the prior detection using GAN. Experimental analysis demonstrates that the proposed system is efficient and beneficial for object detection with reduced computational overhead.