Feature Aggregation With Attention for Aerial Image Segmentation

Deep Learning based algorithms particularly Convolutional Neural Networks (CNN) have shown better results in the challenging task of image segmentation. The prerequisites of high-end hardware, the large amount of ground-truth labeled data, and high computational complexity are undesirable. Similarly, environmental and object size variations are additional challenges to image segmentation. This work proposes a technique to modify the VGG-16 network with the specialties of several best performing CNN networks such as ResNet, DenseNet, and Squeeze Net, etc. The high-level features in the feature extractor are learned using dilated convolution kernels. The varied dilated rates are applied to form a pyramid of features. The preceding layers are added as channel-wise attention to the next layer. In this way, the lost features are retrieved and the global feature is learned. The learned features are then upsampled as bilinear interpolation followed by 3 x 3 convolution. The features from the mid-level and low-level of feature extractor are also added to corresponding layers of the upsampling network to recover foreground object. The proposed algorithm is tested on two publicly available data-sets with top-ranked image segmentation algorithms.

Automated summary

This paper introduces a deep learning-based algorithm that modifies the VGG-16 network to enhance image segmentation results by combining features of various CNN networks. The use of dilated convolution kernels, feature pyramids, and channel-wise attention aids in recovering lost features and learning global features for foreground object recovery. The algorithm is tested on two public datasets against top-ranked image segmentation methods.