Efficient Vertex Coordinate Prediction-Based CSP-Hourglass Net for Object OBB Detection in Remote Sensing

Objects in high-resolution remote sensing images are distributed in different directions with multiscale characteristics, resulting in a sharp decline in the detection performance. In this letter, an efficient oriented bounding box (OBB) detection technique is proposed in remote sensing images. By regressing the relative coordinates of OBB relative to its minimum bounding rectangle (MBR), the proposed prediction can provide a more compact representation of the object. Then, on this basis, a new backbone network, Cross Stage Partial (CSP)-Hourglass Net, is proposed for remote sensing to address the problem that multismall object features are easily lost in the deep network. Using a dense up-and-down sampling link structure, the proposed CSP-Hourglass Net exhibits its potential capability in adapting learning the features of small and multiscale objects. To testify the proposed prediction and network, classical high-resolution remote sensing datasets are used. The experimental results show that the proposed prediction and network owns better detection performance, which demonstrates its superiority.