Research On Computer Vision System Of Agaricus Bisporus Harvesting Robot Based On Deep Learning

Agaricus bisporus is one kind of the most common mushrooms.The cultivation mode is gradually changing to the intelligent factory.Most of the production processes have been mechanized and automated in the production factory.However,the harvest and quality grading still rely on human.It is tedious,laborious and expensive to harvest and grade Agaricus bisporus manually.Picking and grading Agaricus bisporus automatically become urgent demands in the industry.The computer vision system is an significant component of the Agaricus bisporus harvesting robot.The research on the vision system of the Agaricus bisporus harvesting robot contribute to the realization of the automatic harvest and quality grading.This paper focuses on the industrial demand of picking and grading Agaricus bisporus automatically.The multi-modal object detection algorithm and lightweight quality grading algorithm are mainly studied.The harvesting robot vision system of Agaricus bisporus is constructed,which provides strong technical support for harvesting and grading Agaricus bisporus automatically.In general,the main research achievements of this paper are as follows:(1)The image datasets of Agaricus bisporus is constructed.To improve the performance of object detection,the RGB-D image dataset of Agaricus bisporus is constructed.In order to identify the quality of Agaricus bisporus automatically,the quality grading dataset is created.The datasets contribute to the research of automatic picking and grading Agaricus bisporus.(2)The multi-modal attention fusion network is presented to detect Agaricus bisporus.The presented network fuses RGB and depth features to extract more complete features.The multi-scale network architecture is adopted to increase the receptive field.Moreover,we adopt attention mechanism to recalibrate channel features to enhance the representation ability of network.Through the method of complementing information between multi-modal,the accuracy and robustness of the detector are improved.(3)A lightweight neural network is proposed to grade Agaricus bisporus automatically.To elevating the efficiency of quality grading,a compressed block is proposed.The principal idea is converting serial pooling and standard convolution to parallel pattern for reducing the computation complexity.A novel neural network,called Light Net has been presented base on the compressed block.By extracting image features,the network divides Agaricus bisporus into two classes,high quality and inferior quality.The categories of inferior quality include irregular shape and rusty spot.The proposed network has less parameters and computation.Meanwhile,the proposed network can obtain comparable accuracy.The generalization ability of the algorithm is verified on the Zizania quality grading dataset.The experimental results show that the network can be expanded to other tasks related to agricultural product grading.(4)The vision system of Agaricus bisporus harvesting robot is developed.The methods of calculating Agaricus bisporus three-dimensional coordinates and picking sequence planning are designed and implemented.The vision system of Agaricus bisporus harvesting robot is built.The vision system is integrated with the hardware device of Agaricus bisporus harvesting robot.The vision system is tested by performing multiple sets of experiment.The experimental results show that the vision system can detect and locate Agaricus bisporus accurately.Through the above work,harvesting and grading Agaricus bisporus automatically are studied in this paper.The results show that the proposed multi-modal attention fusion network improves the detection performance of Agaricus bisporus.The proposed lightweight neural network has lower time complexity and space complexity and maintains comparable grading accuracy.The vision system of Agaricus bisporus harvesting robot is developed.Moreover,the vision system is integrated with the Agaricus bisporus harvesting robot.The extensive experimental results show that the vision system can detect and locate Agaricus bisporus accurately,which provides strong technical support for harvesting and grading Agaricus bisporus automatically.