Multi-Branch Cascade Receptive Field Residual Network

Abstract

Deep convolutional neural networks (CNNs) have significantly enhanced image classification in the past decade. This paper proposes Multi-branch Cascade Receptive Field Residual Networks (MCRF-ResNets) based on the original Residual Network (ResNet) architecture for classification and object detection. MCRF-ResNets incorporate multiple branches with different receptive field (RF) sizes to improve image classification performance. Each MCRF residual block contains a $5\times5$ and a $3\times3$ RF-sized convolution block. The $5\times5$ and $3\times3$ RF-sized blocks recognize larger and smaller objects, respectively. Group convolutions are used to reduce redundancy and balance feature extraction and parameter usage in the multiple branches. The proposed model approach demonstrates significant improvement in model performance on CIFAR-10 and CIFAR-100 datasets and a subset of the ImageNet 2012 dataset. The model achieves a 2.8% increase in top-1 accuracy compared with the baseline ResNet-50 model with a similar number of parameters on the subset of the ImageNet 2012 dataset. In addition, the proposed method shows good object detection on Pascal VOC and MS COCO datasets.

Keywords

• Cascade receptive field
• feature fusion
• image classification
• object detection