Comparison of Light-Weight Multi-Scale CNNs for Texture Regression in Agricultural Context

While texture classification has a long history in image preprocessing tasks, its application in agriculture has only recently gained attention in the context of digital farming. The usage of camera drones allows the inspection of fields with a view from above. This paper proposes a method for the patch-based classification of different and basic ground regions using texture regression with convolutional neural networks. Two shallow multi-scale architectures are compared, which differ in the re-use of feature maps. It can be shown that a light-weight network is able to classify the textures with high accuracy. The performance is checked using the standard data set KTH-TIPS2b. The classification information can be effectively utilised for semantic image segmentation.