MobilenetV2-Based Network for Bamboo Classification with Tri-Classification Dataset and Fog Removal Training

The problem of low recognition accuracy of the front and back sides of bamboo slices has been a challenge in traditional bamboo processing, which has hindered the large-scale promotion of fully automated production equipment. In this paper, a classification recognition network based on MobilenetV2 was proposed to tackle this problem. The materials for the front and back sides of processed bamboo were collected on-site in the factory. The bamboo slices were initially recorded using a camera, and then the images were captured by capturing frames from the recorded videos. The classification accuracy reached 98%-99% after training on three different datasets in a three-classification experiment. Nevertheless, the recognition accuracy may decrease in actual production and processing environments, due to the presence of dust. To overcome this issue, two types of datasets were selected for screening: clear and unclear. The recognition accuracy of MobileNetV2 training on the unclear dataset showed a noticeable decline. However, the overall recognition accuracy can reach 96%-97% after training the dehazed MobileNetV2 network.