Abstract

Background: Detecting occluded objects in images remains a challenging problem since the occluded objects often occlude each other or are occluded by other objects. It is hard to identify the occluded objects in an image, especially when the occlusion is significant.

Methods: In this work, a two-stage object detection method has been proposed. The proposed method is based on the Faster RCNN model and uses ResNet50 as the backbone network. In addition, the method uses the feature pyramid network to reuse the higher-resolution maps of the feature hierarchy. The dilated convolution in the architecture of the proposed network has been added to expand the receptive field of the feature maps and the loss function and gradient learning rate are optimized.

Results: The proposed detector is trained in an end-to-end fashion, which achieves state-of-theart results on two datasets, i.e., MAFA and WIDER FACE, particularly for WIDER FACE (with the highest mAP, 66.4%).

Conclusion: In conclusion, by adding dilated convolution and optimizing gradient learning rate to the object detection model, the precision of the occlusion object detection can be improved effectively.

Keywords: Convolutional neural network, dilated convolution, obscured object detection, object detection, deep learning, machine learning.

Graphical Abstract