QATM-KCNN: IMPROVING TEMPLATE MATCHING PERFORMANCE BASED ON INTEGRATION OF CONVOLUTIONAL NEURAL NETWORKS AND KALMAN FILTERING

Abstract The atmospheric effects and variability in data collection scenarios due to sensor geometries and complex backgrounds within the images make template matching a very difficult task. It may also confine aircraft detection from satellite images. This dataset was used for applied example testing the model being performed. This, however, does not confine the model we are presenting, which can be applied to template matching in various other fields and applications in general. The aircraft dataset selection is just an applied step to show the effectiveness of the model and nothing more, not limited to this dataset. Here, it’s introduced QATM-KCNN, a new method for improving the accuracy of template matching based on combining Kalman filtering, QATM method and Convolutional Neural Networks (CNNs). Using CNN (Here, VGG19) allows effective extraction of initial item locations on images. However, this technique could be affected by errors made during measurements or noise present leading to wrong results; hence the Kalman filter is used to enhance these outcomes. In this work we use QATM to extract the initial coordinates of an object then delivered to a Kalman filter for further refinement. The Experimental results based on the evaluation measure such as the Intersection over Union (IoU) index, show that the combination of the QATM algorithm with Convolutional Neural Network and Kalman filter can lead to significant improvements in object recognition accuracy in Template Matching tasks.