Research On Compression Method Of Target Detection Model For Intelligent Water Service

Artificial intelligence technology integrates information technology with traditional water management and promotes the intelligent implementation of water management.Intelligent recognition of water data is the basis and necessary link of smart water management.Currently,to meet the real-time and security requirements of water data reading,intelligent reading of water data at the edge has become the research focus.However,deep neural network models have complex network structures and massive computation,making it difficult to meet the portable needs of models at the edge.Therefore,model compression algorithms have become one of the key research topics in current deep learning algorithm productization research to meet the deployment conditions of resource-limited edge devices.This article focuses on the SSD object detection algorithm and sequentially carries out algorithm research on model compression from two aspects of network structure and parameter quantity,while ensuring the recognition accuracy of water data.The main research contents are as follows:Firstly,for the complex network structure of SSD,a structured pruning algorithm based on receptive field and channel pruning is proposed to streamline the network structure without compromising recognition performance.This algorithm starts from the detection layer of SSD,optimizes the detection layer of the SSD network through data analysis for training and detection,and compresses the network from depth;then prunes the backbone network used for feature extraction through convolution kernel pruning,compressing the network from width;and finally,uses knowledge distillation to further restore the accuracy of the pruned network.The experimental results on the water meter dataset collected in the field show that the algorithm can compress the model structure by about 73% without losing the recognition accuracy of the original model,which initially meets the need for model deployment at the edge.The structurally compressed SSD model still has a large number of redundant parameters and massive computation.To address these issues,a model compression algorithm based on streamlined network operation units and parameter quantization is proposed,further compressing network parameters of the SSD model through streamlined basic network operation units and low-bit quantization.The experimental results on the water meter dataset collected in the field show that the compressed SSD model can reduce the parameter quantity to about 19% and the computation to about 20% while ensuring accuracy,further meeting the portability of the model at the edge.