An Image Contour Detection Method Based on Multi-level Feature Channel Optimal Coding

Abstract

The invention relates to an image contour detection method based on multi-level feature channel optimization coding. For the input image I(x,y), the present invention first obtains the optimal scale m of the Gabor filter based on the similarity index opt and direction θ opt , and m opt and θ opt As the frequency separation parameter of NSCT; then the contour subgraph obtained by NSCT is enhanced and fused with I(x,y) to realize the primary contour detection of I(x,y); finally, the fully convolutional neural network is designed specifically The network, including feature codecs composed of different scales of FCN‑32s, FCN‑16s, and FCN‑8s network units, uses the convolution and pooling modules of feature encoders to realize active learning of network parameters, and utilizes the feedback of feature decoders The convolution and upsampling module obtains the image contour mask map corresponding to I(x, y), realizes the optimized coding of multi-level feature channels, and completes the efficient and accurate detection of image contours.

Description

technical field [0001] The invention belongs to the field of machine learning and image processing, and in particular relates to an image contour detection method based on multi-level feature channel optimization coding. Background technique [0002] Contour information is of great significance for the segmentation and recognition of image data. It will realize the rapid outline of the image target area and help to analyze and understand the image in the limited feature dimension. Therefore, the automatic detection of image contour is an important aspect of machine learning and image processing One of the important research contents in this field. Traditional detection algorithms based on region gradient information usually consider linear filtering and local directional features of images, such as methods based on image local energy, but they generally do not involve important information such as active contours, texture edges, and region boundaries. At present, the contou...

AI Technical Summary

Problems solved by technology

At present, the contour detection method based on deep learning has attracted attention. It simulates the processing process of visual information by the human visual perception system through the deep network structure, and actively performs feature learning, which effectively simplifies the original complex feature extraction and data reconstruction process. However, this kind of The following problems generally exist in the method: (1) The image segmentation and fusion directly through the neural network will lead to imprecise segmentation results and generalization of feature information

(2) Failure to combine deep learning with traditional feature-based methods, the detection performance is heavily dependent on the quantity and quality of training samples, and the ability to filter redundant information including texture background is weak

(3) Although some methods consider the extraction of multi-source features, such as SAR image segmentation based on Gabor-NSCT and spiking neural network, it involves multi-source feature extraction of Gabor and NSCT at multiple scales, and then the extracted Gabor The feature and NSCT feature are trained as the input of two spiking neural networks respectively, so the segmentation performance will depend heavily on the perception ability of Gabor and NSCT for the image content, and it does not make full use of the multi-source feature signal fusion coding ability at multiple scales. In addition, The spiking neural network does not belong to the category of deep learning in terms of model hierarchy and structure.

For example, there is an image contour extraction method based on Gabor-NSCT and visual mechanism, which also involves multi-source feature extraction at different scales, but considering the computing power of the visual mechanism model, a simplified fusion coding method is usually used, essentially The learning process represented by convolutional neural network training is missing, so it does not really reflect the effectiveness of multi-source features in expressing contours

Images