Image Semantic Segmentation Algorithm Based On Deep Learning

Semantic segmentation is an important research topic in the field of computer vision.Semantic segmentation is widely used in medical imaging,automatic search,driverless,scene analysis and other fields.Before the emergence of deep learning algorithms,traditional segmentation algorithms have limited understanding ability in complex scenarios due to the limitation of accuracy and speed.With the development of artificial intelligence,semantic segmentation algorithm based on deep learning has become the mainstream.In this paper,a deep learning algorithm based on convolutional neural network is used to study semantic segmentation and its subclass instance segmentation.The essence of semantic segmentation is to determine the location of the target and identify the category of the target by classifying the image pixels.Convolutional neural network based semantic segmentation models may exist insufficient features extraction and detail information lossing while reconstruction in upsampling processes,and may reduce the accuracy of the segmentation.An improved algorithm based on fully convolutional network Deeplab v3 is proposed.Firstly,the Leakyrelu activation function is applied to enhance the ability of the network to extract weak features.Then after atrous spatial pyramid pooling,the dense upsampling convolution is used to reconstruct the prediction map with the same size of the input original image.Finally,in training stage,a Nadam optimization is adopted to improve the convergence speed and to enhance the robustness.The algorithm is verified and evaluated on the PASCAL VOC 2012 data set,compared with five classical algorithms.Experiments show that the average intersection ratio(m Io U)is the highest in this paper,and the effect of the algorithm is better than that of other algorithms.Instance segmentation combined with object detection technology is a subtype of semantic segmentation,which is a further research on the basis of semantic segmentation.It can segment each instance in the image with different pixel markers,which solves the problem that semantic segmentation cannot segment different individuals of similar objects.In order to solve the problem of poor segmentation of objects in overlapping,firstly this paper designs multi-path enhancement based on the Mask R-CNN algorithm,which shortens the transmission path of network information,so that the high-dimensional feature and the low-dimensional feature can be fullyintegrated.Secondly,the adaptive feature pooling technique is used to select the characteristics of the region of interest reasonably and avoid the wrong selection;thirdly in mask prediction,the joint use of full convolution and full connection network to obtain more adequate mask,improve mask quality;finally,to improve the generalization ability of the algorithm,training,verification and testing on the ADE20 k dataset,the test results prove that the algorithm segmentation effect is good.