Template-Based Supervised Deep Homography Estimation Based on SqueezeNet

Homography estimation is a common method used to solve many computer vision problems. Homography can be used to implement measurement on instruments based on image processing, replacing the common sensor required for a direct approach with the measured objects. Using the deep neural network in estimating homography is already quite popular among researchers and computer vision-related developers. Many networks are being implemented to solve homography based on classification network to find the transformation matrix. However, implementation of homography from previous studies with complex network requires high computing and is difficult to apply to embedded systems with limited capacity. Our research demonstrates the deep homography estimation for image transformation calibration based on template image with simpler network and requires less computational capability. We use the renowned SqueezeNet network model for the main network backbone with modification on the network head to solve the corner displacement of the homograph. We create the dataset from template image of symbols and train the network from the beginning. We change the output of the SqueezeNet classifier to regression output to acquire the displacement of the corner of perturbed image. To validate the network results, we compare the perturbed input image with the result of homography matrix prediction against the main template image to get the confidence level of the predicted homographic image. The results of our study are close to those of similar studies using more complex networks, but our network requires significantly less computing and power consumption and shorter training time.