D dNet-65 R-CNN: Object Detection Model Fusing Deep Dilated Convolutions and Light-Weight Networks

摘要

In recent years, object detection has become a popular direction of computer vision and digital image processing. All the research work in this paper is a two-stage object detection algorithm based on deep learning. First, this paper proposes the Deep_Dilated Convolution Network (D_dNet). That is, by adding the operation of dilated convolution into the backbone network, in this way, not only the number of training parameters can be further reduced, but also the resolution of feature map and the size of receptive field can be improved. Second, the Fully Convolutional Layer (FC) is usually involved in the re-identification process of region proposal in the traditional object detection. This too "thick" network structure will easily lead to reduced detection speed and excessive computation. Therefore, the feature map before training is compressed in this paper to establish a light-weight network. Then, transfer learning method is introduced in training network to optimize the model. The whole experiment is evaluated based on MSCOCO dataset. Experiments show that the accuracy of the proposed model is improved by 1.3 to 2.2% points.