Improved Yolo-v3 Model with Enhanced Feature Learning for Remote Sensing Image Analysis

Abstract

Remote sensing technique has played important roles in various fields like urban planning and military reconnaissance, however, due to remote sensing images (RSI) have the unique characteristics of complicated background, densely distribution of targets with varying scales, etc., it remains a challenging work to apply popular object detection algorithms for RSI analysis. In this paper, an improved Yolo-v3 (Im-Yolo) model is developed with enhanced feature learning ability, which can better adapt to handling RSI. In particular, residual convolution and path aggregation are employed so as to effectively enhance the multi-scale feature extraction and semantic-detail information fusion ability of Im-Yolo. Experiments on two challenging remote sensing detection databases have sufficiently demonstrated the reliability and superiority of proposed Im-Yolo on both detection accuracy and inference speed in comparison to the baseline model Yolo-v3. Im-Yolo is proven a competent method for handling RSI with satisfactory performances even in complicated scenarios, which can provide experiences to design RSI-oriented object detection algorithms.