Semantic Segmentation Based On Convolutional NeuralNetworks

In recent years, with the development of deep learning, computer vision is also rapidly developing. Depth convolution neural network (CNN for short) has been proven to be very effective in the field of computer vision. Meanwhile, in everyday life, whether to identify the detected objects, or to do video streaming data analysis both are inseparable from computer vision. Since image segmentation is considered as the most basic computer vision algorithm, because the result of semantic segmentation directly affects the effectiveness of subsequent image classification. Therefore, the implementation of an effective semantic segmentation has practical significance.Conventional method about semantic segmentation needs to conduct the following steps:First, it needs to search for spatial regions in the image. Second, it extracts features from all the regions and classifies the regions into different classes, finally get the result by merging the proper regions. The process seems complicated. As deep learning continues to boom up, recent practice of CNN on image classification has been proven efficient. So nowadays how to introduce CNN to improve semantic segmentation is a major research focus.After a deep investigation, here we implement a deep neural network by combining CNN layer and De-convolutional Neural Network layer to do pixel-wise semantic segmentation prediction. Then we optimize the model’s efficiency by two-stage training and perform controlled Trials with state-of-art algorithms. Finally, this model is applied to medical image analysis (for detecting heart lesions) and cloth parsing (for dressing mix), which all achieve impressive results.