Research On Target Detection Technology For Smart Water

Currently,pointer-type pressure gauges are widely used in various fields such as industrial and agricultural production,daily life,and transportation.The collection,monitoring,and recording of pressure information are of great significance to production and life.Therefore,the task of automatically identifying pointer-type pressure gauge readings has become a research hotspot and focus in the field of intelligent water management systems.Today,with the rapid development of deep learning technology and the widespread use of microcontroller chips,the application of artificial intelligence at the edge is becoming increasingly diverse.However,due to the design structure and performance limitations of microcontrollers,they have certain limitations compared to personal computers.By porting a pre-trained deep learning model to a microcontroller,various application scenarios can be intelligentized,thereby improving production efficiency and reducing labor costs.This article focuses on the SSD object detection algorithm,and carries out research on the target detection model algorithm from two aspects: lightweight model and improved recognition accuracy,while ensuring the accuracy of identifying water data(pressure gauge dial targets).The main research contents are as follows:(1)Lightweighting of the target detection network: Firstly,this article designs a pressure gauge dial object detection model SSD-Light based on the improvement of SSD.This model uses the size of the receptive field to design a lightweight backbone network structure and creates and annotates a dataset.It transforms the original SSD multi-layer and multi-scale detection task into a single-layer and single-scale detection task,significantly reducing the number of layers in the backbone network and detection layers of the original network model.Then,this article introduces depthwise separable convolution kernels to replace the ordinary convolution kernels in the original network,further lightening the model.The experimental results on the dataset show that while ensuring effective recognition of the target,the number of model parameters in the model is only 0.1% of the original SSD network,and the model capacity is only 0.2% of the original SSD network,greatly reducing the model parameter capacity.(2)Improving the detection accuracy of the lightweighted model: The above-mentioned lightweighting techniques may cause the model to perform poorly when processing image data.To address this issue,a lightweight object detection model with a fusion attention mechanism is proposed to improve the model’s accuracy and precision in identifying targets.A large number of repetitive and good results demonstrate that this model can ensure the accuracy of the model while only reducing the m AP by 5% compared to the original SSD,thereby ensuring the accuracy of the model.