Algorithm Research Of Convolutional Neural Network In Contour Detection

Contour detection is designed to extract the boundary between the target and the background in the image.It is one of the most basic problems in the field of computer vision research.As a low-level visual task,contour constraints play a significant role in advanced visual tasks such as target recognition,semantic segmentation and motion tracking.Identifying contours as an important function in the human primary visual cortex.It is important for deepening the understanding of the objects in the image and the processing of information such as semantic analysis and memory storage.Widely used in agriculture,transportation,medical,military and other fields.Deep learning technology,as an effective method to learn feature representation directly from data,has made breakthroughs in the field of inspiration contour detection in recent years.The supervised learning of target contours by constructing an end-to-end deep convolutional neural network is the main method of contour detection tasks.Based on the convolutional neural network model in deep learning,this paper deeply studies the related aspects of natural image contour detection.The main contributions include the following aspects:1)Aiming at the problem of traditional enhanced network in feature decoding and feature fusion,a deep enhancement network is proposed based on the idea of coding-decoding structure.The network consists of multiple layers containing multiple refinement modules.Compared to traditional reinforcement networks,deep refinement network have better feature representation and excellent contour "accuracy" performance.2)The traditional weighted cross-entropy and Dyce coefficient are poor in multi-person annotated samples.By transforming the multi-person annotated sample into the two-class weighting problem,this paper proposes an improved cross entropy and improved Dice cost function.This method solves the problem that the traditional cost function cannot fully utilize the labeling of multi-person annotation.Experiments show that the improved cross-entropy cost function can improve the "accuracy" of the network,and the improved Dice coefficient cost function can improve the contour "fineness".3)Aiming at the problem that the multi-cost function directly combines the imbalance of weight distribution,based on the idea of Nash equilibrium,an adaptive cost function is proposed.Make a relatively good trade-off between the balance of "accuracy" and "fineness".