Algorithm Research Of Image Semantic Segmentation Based On Full Convolution Neural Network

In the age of information,computer vision has become more and more widely used in real life.The use of cameras and computers instead of human eyes to identify,classify and track objects has become a new research craze.Image semantic segmentation is the basis of image understanding,which directly affects the pros and cons of the final result.Therefore,the effectiveness of the semantic segmentation algorithm is very important.Traditional image segmentation is based on the characteristics of the image itself,the object must be embodied,and the application is greatly limited.For complex scenes,the traditional image segmentation accuracy and efficiency are not up to standard.Semantic segmentation based on full convolutional neural network can accept input of any size,realize pixel-level segmentation,and improve the precision of segmentation.However,the semantic segmentation of the full convolutional neural network has problems such as low feature resolution,weak context inference ability,and no consideration of foreground and background distribution imbalance.In response to the above problems,this paper aims to improve the accuracy of semantic segmentation and improve it based on the FCN model.The specific research is as follows:(1)By reducing the pooling step size,the problem of low resolution of the feature map is improved,and the hold convolution without increasing the number of parameters and the amount of calculation is used to increase the receptive field of the neuron node,so that it is easy to learn the feature with higher semantic level;(2)The multi-scale pooling and superposition of high-level semantic information are used to strengthen the association between pixels and pixels,integrate context information,and improve the network's ability to obtain global information;(3)The improved cost function is used to adjust the weight distribution to avoid the bias of the classifier and solve the problem of imbalance of sample categories.Experiments were carried out on the VOC 2012 dataset,and compared with the results of other algorithms.Experiments show that the proposed network structure improves the accuracy of semantic segmentation,and the pixel accuracy reaches 89.9%,which proves the effectiveness of the improved algorithm.Finally,the algorithm is applied to road scene recognition applications and achieves better results.