An enhanced SSD with feature cross-reinforcement for small-object detection

Due to the limited feature information possessed by small objects in images, it is difficult for a single-shot multibox detector (SSD) to quickly notice the important regions of these small image objects. We propose an enhanced SSD based on feature cross-reinforcement (FCR-SSD). For shallow sampling, an improved group shuffling-efficient channel attention (GS-ECA) mechanism is used to make the model focus on the object areas rather than the background. Then, an FCR module allows the multiscale information from the shallow layer to be passed to the subsequent layer and fused to generate an enhanced feature map, which improves the utilization of the context information associated with small objects. We develop an adaptive algorithm for calculating positive and negative candidate box selection thresholds to select positive and negative samples, determine the intersection over union (IOU) thresholds of candidate boxes and ground-truth boxes, and adaptively determine the threshold for each ground-truth box. The proposed FCR-SSD algorithm achieves 79.6% mean average precision (mAP) for the PASCAL VOC 2007 dataset and 30.1% mAP for the MS COCO dataset at 34.2 frames per second (FPS) when run on an RTX 3080Ti GPU. The experimental results show that the FCR-SSD model yields high accuracy and a good detection speed in small-target detection tasks.