| As two tasks of computer vision,object detection and image segmentation have been gradually applied in number detection,face recognition,movie green screen,etc.,and have great application potential in virtual reality and interactive games.Accurately segmenting objects from images is an important part of other applications,such as the body size measurement of cattle based on image segmentation technology,and the perception of the surrounding environment according to the results of semantic segmentation in the automatic driving scene to further make decisions.In these scenes,the segmentation accuracy is higher.The work of this thesis is mainly divided into three parts,including the production of weak texture image data set.Considering the difficulty and practicability of image acquisition,subway security check image is selected as the original data set,and on this basis,the instance segmentation data set required in this thesis is generated.The object detection based instance segmentation method,namely two-stage instance segmentation method,is studied.The direct instance segmentation method,i.e.,single stage instance segmentation method,is studied to achieve effective segmentation of contraband in X-ray images.In view of the image segmentation method based on target detection in the positioning of the boundary is not accurate cause poor segmentation results,based on the feature extraction in the network joined CBAM attention mechanism,different weights in different spatial location,making objects appear larger regional weights,can produce a more accurate boundary box information;Aiming at the problem that the segmentation accuracy is not high due to the large change of target size,the features of different layers are fused in this thesis,and the precise location information in the bottom feature map is added to the high-level semantic features,which can improve the final segmentation accuracy of the model.In terms of running speed,in order to add network reasoning,part of the convolutional layer is replaced to accelerate the network operation.Aiming at the problem of low precision of overlapping object segmentation in direct instance segmentation algorithm,a feature fusion module based on empty convolution is proposed in this thesis to strengthen the semantic information of feature graph and achieve better segmentation effect.Because weak texture image is different from ordinary RGB image,its boundary is not clear,and the object is incomplete due to overlap,so the segmentation effect at the boundary is not good.Therefore,this thesis carries out postprocessing on the segmented boundary,optimizes the segmentation effect at the boundary,and improves the segmentation accuracy. |
