We present a hybrid architecture called Contextual-Enhanced Transformer Network (CEFormer) that leverages both Convolutional Neural Networks (CNN) and transformer-style networks for computer vision tasks. CNNs are good at modeling local features due to their local nature and weight sharing, while transformers are good at capturing global contextual features due to their self-attention mechanism. Our CEFormer uses a parallel network structure that combines the strengths of both CNNs and transformers for image feature representation. Specifically, we design an enhanced multi-headed attention module and contextual attention module that extracts and enhances globle features and contextual features on two branches for the task of small target detection in an autonomous driving environment. Moreover, we propose a lightweight cross-branch fusion module that reduces the parameters and computational complexity of the feature interaction. Our CEFormer achieves competitive results in target detection with Mask R-CNN and outperforms ResNet and transformer-based models. It also shows significant improvement over other methods on MS COCO, TT100K, and ImageNet datasets.
Small Object Detector Using Contextual Local Features and Global Representations for Autonomous Driving
2023-09-24
369485 byte
Conference paper
Electronic Resource
English
European Patent Office | 2023
|