One of the most important sources of grain losses happens during harvesting. A machine vision system is capable to send online information of the grain condition to the adjustment units in order to reduce the grain losses. However, the continuous vibration of combine during harvesting prevents to acquire desirable images for grain monitoring. To solve this issue, several image deblurring algorithms were investigated and the best one was introduced. The x-y-z accelerometer sensors were installed on the combine tank cap and then the acceleration data was extracted. These data were then used for fabrication of a small-scale combine tank simulator for preparation of the image dataset. Five algorithms were used for deblurring, including Fast Image Deconvolution Hyper-Laplacian Priors, Wiener, Lucy-Richardson, Maximum likelihood deconvolution, and regularized iterative image restoration algorithm. Also, three filters of Motion, Average, and Disk were used. Blind/Reference-less Image Spatial QUality Evaluator (BRISQUE) and Natural Image Quality Evaluator (NIQE) were used as the image quality indices. The results showed that Lucy-Richardson had the best performance in image deblurring (NIQE=6.0539 and BRISQUE=43.0489). Maximum Likelihood showed similar performance (NIQE=6.1974 and BRISQUE=43.4288). Analyzing the execution time of the algorithms showed that Lucy-Richardson was the fastest one (0.8538 s) and therefore, this is preferred to Maximum Likelihood in online applications. The results of the applied filters based on the subjective criterion showed that motion filter has produced sharper images while, disk filter performs better in de-noising. Comparing two objective indices showed that both indices performed well in evaluating the quality of deblurred images. However, the results of original images (blurred images) revealed that NIQE showed more sensitivity to image contrast and noise, compared to BRISQUE.
