Medical personal protective equipment detection based on attention mechanism and multi-scale fusion

Deep neural networks (DNNs) have shown excellent effectiveness in object detection and greatly benefit people in various physical scenes. In this paper, we focus on a meaningful physical scene, medical personal protective equipment detection, where the performance degrades for two reasons: background information interference and different detection target scales. To solve the problems above, we propose two novel modules, a deformable and attention residual with 50 layers (DAR50) feature extraction module and a criss-cross feature pyramid network (CCFPN) feature fusion module. Concretely, the DAR50 is target morphology-aware and can enhance the feature information. The CCFPN raises the multi-scale detection performance by fusing the pixel information of the feature maps and then fusing the features of different stages. Combining the two modules, we construct a novel object detection network called attention and multi-scale fusion-based regions with convolution neural network (AMS R-CNN) features. Empirically, we prove the superiority of AMS R-CNN on a medical personal protective equipment detection dataset CPPE-5 (medical personal protective equipment) and The Visual Object Classes Challenge 2007 (VOC 2007) dataset compared with several state-of-the-art methods.