Research On Algorithms Of Transparent Object Detection Based On CNN In Real Scene

Transparent object detection plays an important role in intelligent robot,intelligent laboratory,intelligent experiment teaching and other fields,promoting the development of scientific and technological life.However,it is challenging to detect transparent objects.Because transparent objects do not have their own texture attributes,their appearance largely depends on the environmental background,and it is very difficult to accurately identify transparent objects in complex environments.At present,deep learning technology is widely used in object detection tasks,which greatly improves the performance of detection.Therefore,this thesis applies deep learning technology to the field of transparent object detection,and proposes a transparent object detection algorithm based on CNN in real scene.First of all,there are few datasets for transparent object detection at present,and some images in some large transparent object datasets are synthesized by certain technical means.They are not image data in real scenes,and the background is relatively single,which does not meet the research conditions of transparent object detection in complex scenes in this thesis.Thesefore,this thesis designs a screening processing method MC-TOPM for transparent object datasets under multiple constraints.Combined with the idea of production scheduling under multiple constraints,multiple constraints are formulated.Considering the priority of the task,multiple datasets containing transparent objects are filtered according to conditions,and the TODatasets is obtained for the research of transparent object detection algorithm.based on this,the algorithm models proposed in this thesis are all carried out on TODatasets dataset.Secondly,aiming at the research on improving the accuracy of object detection,this thesis proposes a transparent object detection method(IA-MLFF RCNN)based on improved attention architecture and multi-level feature fusion.Based on the Faster RCNN model,the improved attention architecture(IAA)is first designed,which is embedded into the backbone network,and the channel attention mechanism is used to selectively emphasize the channel containing more important information to improve the discriminant ability of the network.Then,aiming at the multi-scale problem in object detection algorithm,based on the idea of FPN,a multi-level feature fusion module(MLFFM)is designed to fully integrate the rich global semantic information of deep features and the local spatial location information of shallow features,and introduces weights for different levels of features to balance the feature information of different scales.Finally,the two modules are integrated into the Faster RCNN model to form the IA-MLFF RCNN model,and the IA-MLFF RCNN is applied to the detection research of transparent object dataset TODatasets.The experimental results show that the accuracy of the IA-MLFF RCNN is improved by about 1.41% compared with other detection algorithms.Finally,in order to reduce the model parameters and improve the detection speed,this thesis proposes a transparent objects detection model(PCS-YOLO)based on the strategy of parameter compression.Its lightweight and high efficiency is suitable for small devices,embedded devices or mobile terminals with limited computing power and storage,and alleviate the problems of complex network model structure,large amount of parameters and slow detection speed.PCS-YOLO includes three modules: mobilenet-V1 as feature extraction network module,enhanced feature extraction network module(DW-EFEN)composed of SPP and Dw-PANet,and prediction and prior anchor adjustment module(P-BAM).Based on the YOLOv4 model,the lightweight Mobilenet-v1 backbone network is firstly selected to extract feature information of objects.Then,DW-EFEN based on depth separable convolution is designed to integrate and strengthen the extracted feature information to obtain more effective feature information.Finally,the small P-BAM is used to predict the input image and adjust the prior box,and output the final detection result.In the study of PCS-YOLO for transparent object detection,the experimental results show that the number of parameters of PCS-YOLO is only 13.32 M,and the detection speed is increased by about 1.7 times.