| With the expansion of tomato cultivation scale in facility agriculture in China,the workload of tomato picking and sorting is increasing,and intelligent identification and sorting have become the development demand of The Times.At present,the related technologies of crop maturity grading and size detection have been studied and optimized to some extent,but the field of crop size detection is still in the in-depth research stage,and most of its work is still completed by manual or mechanical equipment design.The maturity classification and intelligent detection of size can effectively improve the detection accuracy,reduce the labor force and reduce the loss rate and consumption rate,and lay a foundation for the sustainable development of facility agriculture.According to the tomato image feature recognition algorithm,this paper compared and analyzed deep learning algorithms such as VGG16,Res Net50,Res Net101 and Res Net152 in tomato maturity recognition,and proposed an improved Faster RCNN tomato maturity grading and size detection model based on the target values of precision and accuracy.The research contents and results of this topic are as follows:(1)The effect of four different tomato maturity grading detection models was compared.The tomato maturity recognition of four recognition models,Res Net50,Res Net101,Res Net152 and VGG16,was realized by using transfer learning thinking,and the recognition results were compared and analyzed.The results showed that the MAP(Mean Average Precision)values of tomato ripeness grading were 87.46%,88.06%,87.67% and 88.63%,respectively,and the single detection time was 0.517 s,0.599 s,0.677 s and 0.49 s,respectively.The results showed that VGG16 model with high accuracy and fast speed was more suitable for tomato maturity grading detection.(2)Optimize and improve the Faster RCNN VGG16 tomato maturity grading detection model.Firstly,by modifying the threshold range of IOU(intersect-over-union)in RPN(region-proposal Network),the convolution layer of VGG16 was optimized and reduced again,and some other parameters were adjusted.Finally,Gaussian filtering was carried out on tomato image data set,and the optimized recognition model was used to train the data set.The results showed that the MAP value of image detection reached 90.21%,the detection time was0.46 s,the detection accuracy was improved by 1.58%,and the detection time was accelerated by 0.03 s.(3)Tomato size detection model based on Faster RCNN framework.The Bounding-Box regression function in the Faster RCNN framework was used to obtain the diagonal coordinate position of the outer rectangular frame of the tomato,and the length and width of the outer rectangle were calculated by using the diagonal coordinate value.After scaling,the fruit axis and fruit length of the tomato after image processing were obtained,and the detected value was compared with the actual measured value of the vernier caliper.The results showed that the error range between the measured value and the measured value of tomato was0.1-3.4mm,the relative error of fruit axis reached 1.70%,and the relative error of fruit length reached 1.13%,which met the requirements of actual environmental detection. |
