Research On Image Detection Technology Of Industrial Images Based On Deep Neural Network

Manufacturing industry accounts for about 30% of China's GDP and is the main body of the national economy.But the problems of large but not strong and low automation level exist objectively.In the manufacturing process of products,surface defect detection is an indispensable key link.The traditional task of surface defect detection is completed by human,which is faced with the problem of strong subjectivity and low efficiency.In recent years,driven by the computing power of equipment,deep learning technology promotes the rapid development of image processing and computer vision technology.Moreover,under the promotion of national strategy,deep learning technology is deeply integrated with manufacturing industry.Aiming at the requirement of automatic defect detection in industrial scene,this paper studies the algorithm of industrial image defect detection based on deep neural network.According to the difference of semantic information between the defect image and the natural scene image,the proposed feature extraction backbone network fully reuses the front layer features,and provides a fast up transfer channel for the bottom layer texture information.In addition,multi-layer feature fusion and reuse methods are used to solve the problem of large defect size difference and small size defect information annihilation in deep feature map.Aiming at the specific problem of the surface defect detection of engine piston ring,the defect detection algorithm of piston ring is designed based on the idea of fast RCNN two-stage detection algorithm.This algorithm adopts the idea of multi task learning.By adding a classification module to the backbone network,the classification task with or without defects is set as the main task of the network.The defect detection is set as the sub task of the network,and the strong supervision signal of defect location information and category information is used to guide the backbone network to learn the specified defect features at each stage.The setting of network parameters is based on the statistical information of data set,and completes the task of platform building and network training.Using the evaluation index to evaluate the model,the validity of the algorithm is verified,and the advanced algorithm is verified by comparing with the mainstream algorithm.The embedded deployment of the algorithm is also a common requirement of industrial scenarios.In this paper,a lightweight defect detection algorithm for embedded devices,embednet,is proposed.The main network of the detection algorithm introduces deep separable convolution,which greatly reduces the network parameters and improves the reasoning efficiency of the algorithm.The experiment focuses on the embedded device of NVIDIA Jetson nano,and uses the piston defect data set to verify the effectiveness of the embednet algorithm.The experimental results show that the reasoning speed of embednet algorithm in Jetson nano embedded device reaches 19 fps.If the model is further accelerated by tensorrt technology,the reasoning speed can be increased to 25 FPS,which can reach the real-time detection standard.