Recognition of mango skin defect based on semantic segmentation

Objective  The semantic segmentation technology was used to automatically identify mango and its skin defects, to realize the quality evaluation and sorting of mango and provide a reference for the rapid and nondestructive testing of mango quality.

Method  Mango skin defect images in multi-scene of natural environment were collected for model training and testing. Atrous spatial pyramid pooling (ASPP) in DeepLabV3+ was replaced by joint pyramid upsampling (JPU) structure, and Xception model in DeepLabV3+ was replaced by Atrous-ResNet model. Class pixel accuracy (CPA), mean pixel accuracy (MPA) and mean intersection over union (MIoU) were used as the accuracy evaluation indexes of each model.

Result  JPU module was used to replace ASPP module, and Atrous convolution was applied to ResNet network which was conductive to increase the receptive field of the model. In general, the predicted boundary was smoother, and the identification of small defects was more accurate. The comparison with SegNet and LinkNet algorithms showed that Atrous-ResNet model had higher accuracy, with CPA slightly improved, MPA was up 3.79 percent point and MIoU was up 4.57 percent. Atrous-ResNet model had better identification effects.

Conclusion  The method based on semantic segmentation is feasible for mango skin defect recognition. Compared with SegNet and LinkNet algorithms, Atrous-ResNet model has higher recognition accuracy.