Image Recognition Based on the Depth-Wise Separable Convolution and Softpool

For the purpose of enabling the garbage classification to work accurately and efficiently, the image recognition method based on improved Inception-ResNet-V2 network is studied, and four types of daily domestic wastes are classified and identified. In the proposed network, the connection structure in the primary inception module is improved to achieve a dense connection, Softpool is applied to replace the traditional Maxpool pooling method, fine-grained feature information is retained, more intensive feature activations are enlarged, and the Depth-wise separable convolution is used to replace the common convolution method. The improved network not only reduces the quantity of calculation and expedites the training speed for the network, but also captures more image features fully, thereby the recognition accuracy is improved further. Compared with the ResNet50, AlexNet, and YOLOv5 network model, the results show that the recognition accuracy of the network model proposed in this paper comes up to 96.8%, which is 5% higher than that of the YOLOv5 network. The performance of the improved network is significantly enhanced comparing with the traditional network. It is proved that the algorithm is eligible to be successfully applied to the problem of garbage classification, and it greatly weakens the difficulty of municipal garbage recovery.