Research On Visual Identification System Of Underground Scene Based On Embedded Terminal

In order to realize the unmanned and automatic of underground locomotives and ensure the safe and orderly operation of locomotives,the research on the image recognizer as the eyes of locomotives is a key step of mine safety production monitoring.The image recognizer is mainly used to identify various objects and targets in the downhole scene,and provides scene analysis for the locomotive.It is the key guarantee for the safe and efficient production of the mine,as well as the strong guarantee for reducing the occurrence of underground transportation accidents,casualties and property losses.The continuous development and gradual maturity of deep learning technology has made the image recognition effect better and better.At present,many excellent target detection and recognition networks have emerged.However,most of the network models are based on the road surface,and there are few networks that can be directly used for downhole scene recognition.Moreover,most of the networks are too large for the practical mobile applications to be transplanted,and there is no formed hardware equipment for underground scene recognition.Therefore,it is of great research value and practical significance to study a visual identification system of underground scene based on embedded terminal.In view of the above situation,this thesis has made two parts of theoretical and engineering research work:1.The SSD network model based on VGG is studied and improved to identify more small target objects,and the average recognition accuracy of small target is improved by nearly 3 percentage points in the public data set and nearly 5 percentage points in the downhole data set.The compact single target classification network model is studied,the design principle of compact network module is discussed,the module structure is given,and the compact single target classification network structure is proposed.Compression algorithm of deep network is studied,we have compared the existing typical compression methods and improved the existing method.By pruning and quantitative parallel compression to improve the compression effect of the network,and the compression method is applied to the underground scene target detection and classification of the network.Under the premise that the recognition accuracy has little influence,the target detection network is compressed from 20.56 MB to 1.78 MB,the target classification network is compressed from 4.76 MB to 372.06 KB,and networks speed up 3 times on TX2 platform.2.This thesis studies the engineering application of object detection and recognition in downhole scene,presents the application requirements of image recognizer applied in visual recognition of downhole scene,and puts forward relevant technical indexes,main features and main functions.The hardware structure of the image recognizer is designed and improved based on the TX2 platform,the hardware design process is introduced,the related schematic diagram of the hardware design is given,the software development process of the image recognizer is introduced,and the actual detection and recognition effect of the light compression network is given.