A low-resolution image restoration classifier network to identify stored-grain insects from images of sticky boards

Abstract

Sticky board traps are widely used to capture insects in grain bins or warehouses and their surroundings. An Online Insect Trapping Device (OITD), developed by our group, was used to automatically acquire images of insects captured on sticky boards. In order to improve the identification performance of high-noise and low-resolution insect images taken by the OITD, a Low-Resolution Image restoration Classifier Network (LRIRCNet) based on Generative Adversarial Networks (GANs) was developed and evaluated for its performance in this paper. To evaluate effectiveness of the developed LRIRCNet, a dataset of 3327 paired images with high- and low-resolutions was established. The 3327 paired images include pairs of 634, 612, 528, 522 and 1031 single insect images of five species of stored-grain insects, namely Sitophilus oryzae (L.), Lasioderma serricorne (F.), Tribolium castaneum (Herbst), Rhizopertha dominica (F.), and Oryzaephilus surinamensis (L.); respectively. An additional 3500 low-resolution images with a single insect on each image and without paired high resolution images were also used as test data after the 3327 paired images were used. The average recall rate of the proposed classifier network for the low-resolution images was higher than that when an un-adversarial classifier network was used. The maximum difference of the recall rate between the two networks was 19 percentage point.