Multi-task learning based feature extraction method in signal processing of high resolution remote sensing video images

The accurate acquisition of feature characteristics through satellite remote sensing data is of great significance in guiding engineering research and planning. Aiming at the problem of similar feature misclassification in the signal extraction results of high-resolution remote sensing images, based on the attention mechanism, a multi-task learning mechanism is introduced, and a multi-decoder triple-attention model is constructed, which transforms a multiclassification feature extraction problem into multiple dichotomous feature extraction problems, reduces the parameter competition relationship between different categories, and replaces the model’s multiclassification decoders by multiple dichotomous decoders, with each decoder still consists of three attention modules. The method is compared and analyzed with other cutting-edge semantic segmentation methods respectively, and the experimental results show that the multi-task structure MD TANet adopted in this paper outperforms other methods, with the extraction accuracy further improved by 2.43%~6.38%, while achieving high real-time performance of 90.72 FPS, and optimal performance of the overall mIoU results on different datasets. The feature extraction method based on multi-task learning in this paper achieves higher detection and segmentation accuracy compared to other methods, making it more valuable for engineering applications.