ReCUS: Reconvolution and Upsampling Network for Object Detection

Most of the mainstream object detection models such as RetinaNet, SSD, YOLO, and Faster RCNN hardly achieve a good balance between detection accuracy and speed. A major reason is rich deep feature semantic information of images is not fully exploited. To solve this problem, a novel deep convolutional network structure termed as reconvolution and upsampling network (ReCUS) is proposed. In the ReCUS, a modified path aggregation network(mPAN) is added after the backbone, which is beneficial to strengthen the foreground salient feature information and weaken background information. Moreover, two new spatial pyramid pooling (SPP) modules are embedded before output heads for multi-scale fusion of local and global features. The experiments show that the effectiveness of our proposed ReCUS. Furthermore, the better detectability of the ReCUS network is demonstrated for both small scale objects and large scale objects.