A Human Target Infrared Image Segmentation Approach Based on Convolution Neural Network

Abstract

In order to effectively segment the human target under complex background constraints, we present an infrared target segmentation method based on deep convolution neural network, and proposes the loss function based on the intersection-over-union for network optimization. Firstly, we design the network architecture which consists of a contracting path to capture the feature content and a symmetric expanding path that enables precise localization. And then rely on the powerful data amplification technology to effectively train the available sample data. The experimental results show that the network can make full use of the prior information of the data to study the characteristics of the human target, which can use less training data for end-to-end training in the human body target infrared image segmentation. And segmentation effect is superior to the traditional image segmentation algorithm. In addition, the network segmentation speed is very fast, 320 ×256 size image segmentation takes less than 0.2 seconds, to meet the human body target image segmentation of the effectiveness and real-time needs.