EFDet-SPP: efficient anchor-free network for fine vehicle detection

Abstract

Existing vehicle detection methods lack the fine vehicle detection algorithm. In order to improve the accuracy and applicability of anchor-based object detection models, a novel and practical vehicle Fine-grained identification network (EFDet-SPP) based on the EfficientDet is proposed. The improved network adds a Spatial Pyramid Pooling module (SPP) after the feature extraction network for concatenating features to enhance network learning capabilities, and multi-scale extraction of highly semantic features of images. Anchor-based predictions are converted to pixel-based predictions by combining FCOS's head network, eliminating the hyperparameters associated with anchor boxes. And with Mosaic, Copy-Paste data augmentation methods scale small object samples to achieve data sample balance. Experimental results show that the improved network has achieved 94.8% in the actual collected fine vehicle detection dataset, which is greatly improved compared with the EfficientDet network, and does not significantly increase the training parameters and calculation amount of the network.