Fusion Shuffle Light Detector

The accuracy and speed of the one-stage object detection models based on the deep convolutional neural network in various detection tasks are in an advantageous position in comprehensive performance. However, in order to ensure accuracy, the models require a large number of trainable parameters as support, while the storage space occupied by the weight file obtained by training is too large. This makes it difficult to apply object detection models based on deep convolutional neural networks to embedded devices and mobile devices with limited storage space. To address this problem, we propose the new object detection model – Fusion Shuffle Light Detector(referred to as FSLDet). The model uses an improved ShuffleNet v2 network for feature extraction tasks, which improves the accuracy of the lightweight model. And in order to further improve the accuracy of the model, this paper uses the bidirectional feature pyramid(BiFPN) to improve the feature fusion operation, so that the deep-level feature map containing semantic information and the shallow-level feature map containing detailed information are fully integrated. According to the effectiveness of the improved method adopted by the FSLDet model, this paper conducts experiments on the PASCAL VOC 2007+2012 dataset. Compared with the model before the improvement, FSLDet has achieved a 2.97% improvement in the accuracy of the PASCAL VOC 2007+2012 dataset.