An image classification method based on local area depth feature coding

Abstract

The invention relates to an image classification method based on local area depth feature coding, in particular to a fusion likelihood sampling technology, a convolutional neural network technology and a local aggregation descriptor coding technology, and the method comprises the steps: 1, carrying out the likelihood sampling of an image, and obtaining a group of candidate windows and window scores which may contain a target object; Step 2, performing feature extraction on the obtained candidate window by using a convolutional neural network; Step 3, coding the extracted features by adopting an improved VLAD coding technology based on a multi-neighbor distribution strategy and a window score; And step 4, using a linear SVM as a classifier to realize an image classification task. Accordingto the method, a method of combining physical sampling and a convolutional neural network technology is adopted to extract local region depth features of the image, and then an improved VLAD technology is adopted to encode the local features, so that the calculation efficiency is effectively improved, a better classification effect is achieved, and the method is suitable for various image classification tasks.

Description

technical field [0001] The present invention generally relates to computer vision and pattern recognition technology, and specifically belongs to an image classification method based on local area depth feature coding, which is applicable to various image data. Background technique [0002] Image classification is widely used in object recognition, image retrieval and other tasks, and it is one of the research hotspots in the fields of computer vision and so on. Image classification has become one of the most challenging tasks in the field of computer vision due to changes in scale and viewing angle, complex background, and illumination changes. [0003] Early image classification research mainly used global features such as image color, texture, and shape. Since global features lack local information of the image and are less robust to illumination changes, occlusions, etc., image classification based on local invariant features such as SIFT Algorithms are proposed. In re...

AI Technical Summary

Problems solved by technology

[0005] The existing image classification method based on convolutional neural network has two shortcomings: (1) the existing method of extracting local features based on sliding window and convolutional neural network makes the effective expression ability of the proposed feature weak, And the amount of calculation is too large

(2) The existing local aggregation descriptor technology assigns features to the nearest neighbor dictionary for encoding, which will cause encoding conflicts and cause effective neighborhood information to be lost

The disadvantage is that with the deepening of the neural network, the memory usage is very large, and the amount of calculation also increases. At the same time, it is still aimed at extracting the global deep feature representation of the entire image, and lacks the ability to extract the deep feature representation of the target object.

Images