THESIS
2019
Abstract
An innovative multispectral imaging system in the field of agricultural automation is introduced in this thesis. In the real application of agricultural automation, training dataset is often needed to be specially collected. In our case, mass multispectral images can’t be obtained online. Although we can fetch a lot of unlabeled data using autonomous device, labelling the mass data is a very difficult and time-consuming work. For example, in terms of avocado ripeness predication, it’s difficult for an untrained staff to classify a half-ripe avocado into ripe category or unripe category. So, we can only obtain an accurate small dataset and a large unlabeled dataset. For latest method using other spectral devices, such as spectrometer and hyperspectral camera, only the accurate small data...[
Read more ]
An innovative multispectral imaging system in the field of agricultural automation is introduced in this thesis. In the real application of agricultural automation, training dataset is often needed to be specially collected. In our case, mass multispectral images can’t be obtained online. Although we can fetch a lot of unlabeled data using autonomous device, labelling the mass data is a very difficult and time-consuming work. For example, in terms of avocado ripeness predication, it’s difficult for an untrained staff to classify a half-ripe avocado into ripe category or unripe category. So, we can only obtain an accurate small dataset and a large unlabeled dataset. For latest method using other spectral devices, such as spectrometer and hyperspectral camera, only the accurate small dataset is used along with classification algorithm such as PCA, PLS-DA and LDA methods. In terms of spectrometer the information collected is limited while the hyperspectral camera is much more expensive and take image slowly. Our device successfully utilizes a larger unlabeled dataset to generate a more accurate prediction model with low cost and fast frame rate. A new DNN structure, called MSpecNet, is designed for multispectral images to solve the misalignment problem of multi-camera structure. The MSpecNet model is then converted to an embedded version to use hardware acceleration with the aid of sparsity and quantization. The classification accuracy of on-chip solution achieves 88.1%.
Post a Comment