A Recognition And Grading Method For Machine Picking Of Elevated Strawberries

In recent years,with the rapid promotion of the elevated strawberry cultivation model,the mechanical picking of strawberries has been put on the agenda.For the trend of mechanized picking,research on the efficient recognition and grading techniques for strawberries has become a priority.However,there are problems such as strawberries overlap with each other or may be sheltered by leaves and individual maturity,which poses a great challenge for accurate and efficient mechanized strawberry picking.Simultaneously,the quality classification of mature strawberries is limited to the secondary picking,which in turn leads to additional fruit damage.If the quality grading can be completed at the same time as mechanized picking,the efficiency of automatic picking will be further improved.This paper studies the methods for rapid recognition and quality classification of elevated mature strawberries while picking.The specific work and achievements are as follows:Firstly,the pre-processing methods for image samples of elevated strawberries under picking environment were proposed,including: separation of targets and background using the bilateral filtering and Ostu segmentation algorithm,the experimental samples of mature strawberries recognition and quality classification methods were obtained using the minimum external rectangular marking method.Secondly,a multi-scale sliding window target detection method for strawberries is proposed,and a detection algorithm based on HOG feature and support vector machine(SVM)is adopted.The experimental results showed that the 100% of the separated strawberries can be detected,the average detection rate of overlapping strawberries was 97%,the average detection time of the 1024×1024 image in the 64×128 window was 1.1 seconds,and the strawberry average detection time was 0.238 seconds.Then,two recognition strategies for mature strawberries were designed:(1)SVM recognition algorithm based on HOG feature and H feature;(2)CaffeNet recognition method based on deep learning network.The results of comparative experiments show that both SVM and CaffeNet can quickly and accurately identify mature strawberries to be picked.But the average recognition rate of CaffeNet is 95.67%,which is obviously better than 85.12% of SVM,and the average recognition time of SVM is 0.0077 seconds,which is lower than 0.044 seconds of CaffeNet.Finally,three strategies were designed for the classification of mature strawberries during picking:(1)improved BP neural network;(2)CaffeNet grading network;(3)GoogleNet grading network.Comparing the experimental results shows that the GoogleNet's classification accuracy is the highest,with an average of 95.3%,which is only two percentage points higher than CaffeNet.However,GoogleNet average grading time is as long as 0.642 seconds,which is nearly 20 times higher than CaffeNet.Given the real-time requirements,CaffeNet grading performance is the best.