Mango Skin Defect Detection Based On Convolutional Neural Network

The detection of mango skin defects is an important prerequisite for intelligent mango picking and fruit quality grading.Computer vision technology based on convolutional neural network provides a feasible and effective method for defect detection,which is the most mainstream detection method at present.In the natural environment,the intensity of light,the complexity of the background,the mutual shielding of the fruit branches,leaves and stems and other constraints bring great challenges to the detection of mango skin defects.By using deep convolutional neural network,more features can be extracted and more real-time and accurate recognition effect can be achieved.Therefore,in this study,semantic segmentation and case segmentation were used to study the detection of mango skin defects in natural environment.To achieve a more rapid and effective detection.The main research conclusions and work are as follows:(1)Establishment of a data set of mango skin defects.The mango skin defect data set was established by collecting mango images in natural environment and manually labeling them.In order to further expand the data set,image enhancement methods,such as flipping,mirroring,rotation,scaling,contrast enhancement,etc.,are adopted.In order to reduce the impact of occlusion on images in natural environment,random occlusion was carried out in the preprocessing stage to enhance the recognition accuracy of occlusion targets.Morphological processing(closing operation)was used to eliminate the mango epidermis spots and reduce the false detection rate of the model.(2)Segmentation Detection Based on Real-time LinkNet.The LinkNet network is fast,lightweight and easy to deploy.However,the low precision detection effect does not meet the actual demand.Improved network structure from the original ResNet18 LinkNet ResNet34,main model of the fourth layer added expansion convolution operation,the output of the network expansion rate for,2,4,8 [1],and by adopting the combination of cascade and parallel model,on the basis of original increased the receptive field of each layer of network and prospect target and the background of complex which improved the accuracy of small object.The Mean pixel accuracy(MPa)of the improved model increased from 71.33% to83.73%,and the Mean intersection over union(MIOU)increased from 69.68% to 82.42%.(3)Segmentation Detection Based on Deep Convolutional Network Deeplabv3+.DeepLabV3+ algorithm has the characteristics of high precision,fast convergence,but the tiny flaws in the complex environment recognition effect is poorer,the improved model using Atrous-ResNet as feature extraction,network Joint pyramid is added in the encoder module on sampling(be pyramid upsampling,JPU)structure,increase the structure model of multi-scale characteristics and speed up the model convergence,at the same time in the decoder module integration more shallow characteristic,obtained better result on the test set.The experimental results show that,compared with the evaluation index of Deeplabv3+ algorithm,the average pixel accuracy of the improved algorithm increases from 90.69% to 94.48%,and the average crosscuture ratio increases from 89.56% to 94.13%.The results of the improved algorithm are superior to those of Linknet and Segnet.(4)Detection of mango epidermal defects based on case segmentation Mask R-CNN.Mask R-CNN adopts the form of target detection combined with semantic segmentation.However,the model has a high rate of false detection and missed detection.On the basis of the original FPN,the improved model adds a side connection,so that the input and output are on the same level,and the multi-level feature fusion is increased,and a bottom-up feature fusion method is adopted to enhance feature extraction.The accuracy rate of the modified model for mango category and the accuracy rate of defect category increased by 7.14 percentage points and 6.77 percentage points compared with the previous model.The recall rate of mango category and defect category of the improved model were 5.75 percentage points and 6.49 percentage points higher than that of the improved model.