ASABE Technical Library - Abstract

Abstract.

Feature wavebands selection has shown potential for detecting internal and surface defects of cucumbers in hyperspectral imaging (HSI) system. However, the technique heavily relies on prior domain expertise and considerable human labor, making it difficult to generalize to complicated online commercial HSI system. We introduced a self-taught framework that combines a stacked sparse auto-encoder (SSAE) with convolutional neural network (CNN), called CNN-SSAE system, for learning spectral and spatial features in a cucumber defects detection task. CNN was utilized for self-learning of local image features which were, in turn, used for classification by SSAE. SSAE optimized the critical parameters, including sparse constraint, width and depth. This framework was tested experiment on cucumbers with different internal damages (Rolling10, Rolling14), Shrivel, and surface defects (scab/dirt sand and gouge/rot), which obtained from an in-house developed hyperspectral imaging system running at two conveyor speeds of 85 and 165 mm/s. Testing results show that by using spectral features, our system gave the accuracies of 85.6% and 78.3% at the conveyor speeds of 85 and 165 mm/s. By combining spectral and spatial features, we have the improvement of 91.1% and 88.6% at two speeds. The GPU implementation was utilized for speeding up our system, and the average processing rate of CNN-SSAE for one sample achieved 13.5 ms, showing considerable potential for the practical applications.