It has long been a great concern in deep learning that we lack massive data for high-precision training sets, especially in the agriculture field. Plants in images captured in greenhouses, from a distance or up close, not only have various morphological structures but also can have a busy background, leading to huge challenges in labeling and segmentation.
A new article proposes an unsupervised statistical algorithm SAI-LDA (self-adaptive iterative latent Dirichlet allocation) to segment greenhouse tomato images from a field surveillance camera automatically, borrowing the language model LDA.
Hierarchical wavelet features with an overlapping grid word document design and a modified density-based method quick-shift are adopted, respectively, according to different kinds of images, which are classified by specific proportions between fruits, leaves, and the background. The researchers also utilize the feature correlation between several layers of the image to make further optimization through three rounds of iteration of LDA, with updated documents to achieve finer segmentation.
Experiment results show that this method can automatically label the organs of the greenhouse plant under complex circumstances, fast and precisely, overcoming the difficulty of inferior real-time image quality caused by a surveillance camera, and thus obtain large amounts of valuable training sets.