Mango double-sided maturity online detection and classification system

Abstract: Mangoes are ripe fruits. The maturity classification of mangoes can provide convenience for mango ripening, processing, packaging and transportation. Compared with spherical fruit, mango is special shape and soft texture, and its multi-side images are difficult to acquire. Aiming at the irregular shape of mango and the difficulty of detecting the entire appearance, a mango double-sided maturity online detection and classification system based on transfer learning was proposed. The system is mainly composed of conveying device, flexible turning-over mechanism, image acquisition device, grading device and control system. According to the maturity grading standard, the mango is divided into three levels: green ripe, mature and overripe. Then, the mangoes are transported to the image acquisition area through the upper and lower conveyor belts, and image acquisition device obtain images from opposite sides. The control system combines the two-sided image of mango for maturity analysis, and the control grading device classifies mango according to the grading results. A flexible turning-over mechanism is arranged between the two conveyor belts, and the mango is clamped by two flexible belts with the same moving speed and opposite direction, which can ensure that a mango is only turned over once. The mechanical parameters of mango were measured by texture analyzer, and the maximum extrusion force of mango was 1.47-15N through ANSYS analysis. Then, the spring is adaptively adjusted to maintain the proper clamping force for the mango. The whole process of the device is carried out by the motor to drive the conveyor belt to achieve mango transportation. Conveying speed can be adjusted from 0 to 0.5m/s. The upper conveyor belt camera and the lower conveyor belt camera of the device respectively acquire the front and back images of each mango into one image, and the combined image size is 640(960 pixels. The image is preprocessed, and the size of the image is scaled to 224(224 pixels. At the same time, the gray value of each channel is determined in the range of (127-128. In order to acquire the mango classification results, we use the pretrained_dl_classifier_ compact.hdl model pre-trained by hdevelop. 1100 mango samples were used to test. The data were divided into training set, validation set and test set in the ratio of 8:1.5:1.5. SGD method was used to train and validate convolutional neural network. Making tuning of hyper-parameters and setting the appropriate learning rate to 0.001, batch processing to 64, momentum to 0.5. TOP1 error was used as the evaluation standard. When the number of iterations is 460, the Top1 error value of the training set is 3.6%, and the Top1 error value of the validation set is 2.0%. The training model is ideal. The trained convolution neural network model was used to classify 150 mango test sets. The average time consumed for systematic determination of single sample was 0.16 s. The average accuracy of mango maturity classification using convolution neural network model was 96.72%, Among them, the grading effect of maturity of green ripe mango was better, and its accuracy was 97.67%. The accuracy rates of mature and over-ripening were 96.00% and 96.49% respectively. This paper presents an on-line double-sided mango maturity detection and classification device based on transfer learning, and trains the model, which can provide reference for mango maturity automatic classification.