HSF-Net: Multiscale Deep Feature Embedding for Ship Detection in Optical Remote Sensing Imagery

A2RMNet: Adaptively Aspect Ratio Multi-Scale Network for Object Detection in Remote Sensing Images

“…To make networks more robust for geometric variations of objects in remote sensing images, one effective way is to make use of features of middle layers from CNN and then exploit the multi-scale features with multi-level information [25][26][27][28][29][30][31]. Ding et al [25] directly concatenate the multi-scale features of CNN to obtain the fine-grained details for detecting small objects.…”

“…Jiao et al [28,29] introduced the dense feature pyramid network (DFPN) for automatic ship detection in which each feature map was densely connected and merged by concatenation. Furthermore, Li et al [30] proposed a hierarchical selective filtering layer that mapped features of multiple scales to the same scale space for ship detection with various scales. Gao et al [31] designed a tailored pooling pyramid module (TPPM) to take advantage of the contextual information of different subregions with different scales.…”

“…Although the above methods have achieved promising detection results by aggregating multi-scale features, there also exist two common problems. (1) These previous works [25][26][27][28][29][30][31] aggregate multi-scale features of CNN to detect objects of various scales using unified weights in Figure 2a, e.g., a simple summation, which ignore the different importance of different scales. It is easy to interfere with detection due to some noisy features.…”

“…For example, in the right of Figure 2a, the background is incorrectly detected as ship. (2) To extract features of each RoI, these methods [20][21][22]25,[27][28][29][30] compress RoI feature maps of different aspect ratios into the same size in Figure 2c, which is easy to cause excessive shape distortions of objects and then disturb location and recognition of objects. For example, in the right of Figure 2c, the large-vehicle is misclassified as small-vehicle and has an incorrect location.…”

“…However, it is noteworthy that not all features is useful for specific scale object detection when fusing different-scale feature maps. In previous works [25][26][27][28][29][30][31], the different-scale feature maps are fused via a simple summation, which ignores the different importance of different scales. It is easy to produce the false detection results due to some potential noisy features.…”