Detection Of Blueberry Bruising Using Spectroscopic Imaging

Internal bruise damage in blueberry that occurs during harvesting and processing,could affect the quality,appearance,and price paid for blueberry product.Current sensing methods can detect blueberry internal bruising,but they are not able to efficiently and accurately detect and quantify it.The overall goal of this study was to improve the detection of blueberry internal bruises using two hyperspectral imaging systems with different sensing ranges and detectors.Optical properties of blueberry bruised and non-bruised tissue were calculated using a single integrating sphere(IS)-based spectroscopic system.Monte-Carlo multi-layered simulation was used to understand light propagation in blueberries for better employing spectroscopy or imaging techniques.Fully convolutional neural network(FCN)was used to segment and quantify blueberry bruising to help assess the quality.The main results were described as follows:(1)The total reflectance,total transmittance,and collimated transmittance of blueberry flesh and skin with three treatments(non-bruised,30-min bruised,and 24-h bruised)were collected using the SI spectroscopic system in the spectral regions of 500–800 nm and 930–1400 nm.Using the collected spectra,the inverse adding-doubling(IAD)method was applied to calculate the absorption coefficient(μ_a),reduced scattering coefficient(μ_s′),and scattering anisotropy(g).The light propagation model of blueberries was investigated using Monte Carlo multi-layered(MCML)simulation.Results indicated that the differences between bruised(30min and 24 h)and non-bruised flesh samples for bothμs′and g were significant from 930 nm to 1400 nm.Microscope images revealed that the differences were caused by the damaged and ruptured cellular structure of bruised flesh.(2)The two HSI systems were integrated to acquire reflectance images of bruised and non-bruised blueberries.The hypercubes from the two systems were aligned and concatenated into one hypercube of each blueberry.The fused hypercube(700–1650 nm)was fed into a convolutional neural network(CNN)model.The CNN was trained to classified bruised and non-bruised blueberries and the average accuracy was 91.48%,which was better than the results of previous works.(3)The near-infrared HSI system was used to acquire transmittance images of bruised and non-bruised blueberries from 970 to 1400 nm.SVM classifier was trained and tested to classify pixels of bruised and non-bruised tissue.Classification maps were produced,and the bruise ratio was calculated to identify bruised blueberries(bruise ratio>25%).The average accuracy of blueberry identification was 94.5%with the stem-up orientation.The results indicate that detecting bruised blueberries as soon as 30 min after mechanical damage is feasible using hyperspectral transmittance imaging.(4)FCN was used to accurately detect blueberry internal bruising using hyperspectral transmittance images.A total of three classes,including bruised tissue,non-bruised tissue,and the calyx end of blueberries,were treated as segmentation targets.A total of 1,200 blueberry hyperspectral images(HSIs)were randomly divided to form training,validation and testing sets(720:240:240),including 72 images of early bruising in the test set.The new model using full-wavelength input HSIs obtained the highest average intersection over union(IoU)accuracy as 81.2%at the whole test set and 81.1%at early bruising test set.An extra 360 HSIs were used to observe bruise development using the new model.These findings demonstrate the potential for quantifying blueberry bruising more accurately in future work.