Research On Image Semantic Segmentation Based On Embedded And Its Application In Coal Preparation Plant

Image semantic segmentation is to divide each pixel in the image into a certain semantic category,and the predicted target is accurate to the pixel level.At present,the cutting-edge high-precision image semantic segmentation algorithms often have complex model structures and are difficult to deploy to embedded and mobile terminals,and the real-time image semantic segmentation algorithms either have high segmentation accuracy which leads to slow model segmentation or the fast running speed of the model segmentation makes the segmentation accuracy low,so it is difficult to make the model achieve an ideal balance between the segmentation accuracy and speed.In addition,for the scraper conveyor in the coal preparation plant,when the scraper conveyor is working for a long time,it is very easy to cause the scraper to malfunction,which can reduce work efficiency and cause major machine accidents.At present,sensor technology or traditional image segmentation methods are mainly used to detect the scraper.This method has the problem of low accuracy of identifying the scraper and cannot be applied in a complex environment.Therefore,combining the image semantic segmentation technology to complete the video monitoring of the scraper conveyor is of great significance to ensuring the safe production of coal mining.This thesis studies the real-time image semantic segmentation algorithms.Firstly,it analyzes the shortcomings of the existing real-time image semantic segmentation algorithm ERFNet,improves and optimizes on the basis of ERFNet,and proposes a lightweight semantic segmentation algorithm based on improved ERFNet;then the algorithm Transplant to the embedded platform to realize the image semantic segmentation under the embedded platform;finally,the proposed segmentation algorithm is applied to the scraper conveyor of the coal preparation plant.The main work of this thesis is as follows:(1)Propose a lightweight semantic segmentation algorithm based on improved ERFNet.The algorithm is improved from ERFNet.In the encoder,the Non-bottleneck-1D module is replaced with a lightweight convolution module to reduce the amount of model parameters;in the decoder,a lightweight convolution module with Tripet attention mechanism is introduced to obtain rich contextual semantic information to complete finer image semantic segmentation;then use the Focal Loss function with weights to experiment on the common data set.The results show that the accuracy and speed of the segmentation algorithm proposed in this thesis are better than ERFNet,and it also has certain advantages with other real-time semantic segmentation algorithms.(2)The lightweight semantic segmentation algorithm proposed in this thesis is trained in the common data set to obtain the model,and the model is deployed on the embedded TX2 to realize the image semantic segmentation under the embedded.In TX2,TensorRT is used to merge the number of network layers of the segmentation model and perform low-precision calibration processing on the parameters of the model.The results show that the image semantic segmentation is realized under the embedded system,and the model can be accelerated to a certain extent.(3)The lightweight semantic segmentation algorithm proposed in this thesis is applied to the scraper conveyor of a coal preparation plant.Using the segmentation algorithm proposed in this thesis to segment the scraper will obtain a binarized scraper image;then use the Hough transform algorithm and the least square method to obtain the best scraper straight line equation,and calculate the scraper angle value,if the calculated scraper angle value If it is not within the set threshold,it means that the scraper is inclined,otherwise it is normal;finally,a scraper detection system is built on the TX2 to realize the monitoring of the video surveillance of the scraper conveyor and remind the abnormal scene of the scraper.The thesis has 42 pictures,11 tables,and 80 references.