Garbage Classification and Recognition System Based on Neural Network

Traditional garbage classification has many errors and consumes resources seriously, which no longer meets the current needs of garbage classification. In response to these problems, this research proposes a garbage classification model based on migration learning and the application of this model to the actually created garbage separation device. First, pre-train the ResNet50 deep network learning model on the ImageNet dataset; Secondly, transfer the underlying features shared by the edges, colors, and textures learned by the convolution module of the ResNet50 deep network model to the residual network layer of the network model for garbage classification as the initialization parameters; Then use the extracted feature map as input to train the garbage classification model; Finally, modify the fully connected layer to a four-classification problem to accurately classify the garbage. By using the improved Trashnet training set to compare the three pre-training networks, namely alexnet, googlenet and resnet50, the results show that resnet50 has a relatively good recognition accuracy. After fine-tuning the training parameters and the training set, the final verification rate is 91.42%. which basically meet the accuracy requirements of garbage classification. At the same time, this garbage sorting network is implanted into the STM32F4 single-chip microcomputer through the Raspberry Pi to obtain a rubbish sorting device that can identify garbage. The photos recorded by the camera in real time are transmitted to the single-chip microcomputer for processing to obtain the classification result. At this time, the motor starts to work so that the garbage can fall into the corresponding bucket. While reducing the consumption of human and material resources, it greatly improves the accuracy of garbage classification and provides a new method for garbage classification.