Research And Implementation Of Betel Nut Slice Grading System Combining High Resolution And Attention

As a common snack food,betel nut is widely loved by people.The market demand for betel nut is increasing year by year,and the automated processing and production of betel nut has a broad market space.The quality of betel nut slices has a decisive influence on the final betel nut product.The quality of betel nut slices mainly depends on the length and width of the inner and outer contours of the betel nut slices and whether they are clean.At present,the classification of betel nut tablets mainly uses manual classification,which is not only inefficient,the classification standard is subjectively affected by people,and the hygienic coefficient is low.In response to these problems,the convolutional neural network is used to segment the inner and outer contours of the betel nut slices,and the length and width are measured separately,and it is judged whether the core is clean,and the automatic classification is finally completed.After analyzing the production requirements,the Mask R-CNN network capable of segmentation,classification and detection is selected as the algorithm basis for the classification of betel nut slices,and improvements are made on this basis.The improved algorithm is applied to the betel nut slice grading system to achieve the expected effect.The main tasks are as follows:(1)A semi-automatic contour labeling method is designed to simplify data labeling.The training of the deep learning network model requires a large amount of data to support.The Mask R-CNN network is essentially supervised learning,so a large number of input image contours need to be annotated.In view of the low efficiency and heavy workload of manually labeling pictures in the past,a semi-automatic labeling method was designed to label the completed data.By comparing a variety of traditional image segmentation algorithms,MSER is selected to complete contour point extraction,designing and formulating a semi-automatic contour annotation process,and removing outliers to optimize the contour.The data is saved in JSON format to complete the annotation.Experiments show that this method greatly reduces the manual workload.(2)Design a Mask R-CNN semantic segmentation algorithm that combines highresolution network(HRNet)and attention mechanism.In order to improve the accuracy of the inner and outer contour segmentation of betel nut slices,increase the accuracy of nucleus and seedless judgments,reduce the amount and complexity of network parameters,and increase the calculation speed,the Mask R-CNN network is optimized and improved: by improving high-resolution feature extraction Network,reduce the loss of features in the process of multiple convolution pooling.Increase the training dimension of the low-resolution network branch to improve the accuracy of contour segmentation;by improving the convolution block attention module,using multi-level and different convolution kernels for pooling for effective feature weighting,strengthening the effective features of kernels,and suppressing Invalid and interfering features,enhance the accuracy of discrimination with seeds;use depth separable convolution to replace traditional convolution methods,reduce the amount of parameters,and increase the speed of calculation.The experimental results show that the improved model has certain advantages over the unimproved network in terms of segmentation accuracy,accuracy of kernel discrimination,parameter amount,and calculation amount.The improved network is applied to the test items of electrical insulator segmentation,and the effect is also good.(3)Complete the visual hardware selection and visual system construction,and design the grading process.The trained model was applied to the industrial trial production site,and the long-term non-stop test showed that the average pass rate of betel nut slice classification was 96.66%,and the average production efficiency was 75.57kg/h,which met the production requirements.