Object semantics and deep appearance feature fusion-based scene identification method

Abstract

The invention discloses an object semantics and deep appearance feature fusion-based scene identification method. The method comprises the specific steps of obtaining a to-be-identified scene image; extracting object semantics information of the scene image to generate an object semantics feature for keeping spatial layout information; extracting appearance context information of the scene image to generate an appearance context feature; extracting global appearance information of the scene image to generate a global appearance feature; and according to the object semantics feature, the appearance context feature and the global appearance feature, obtaining an identification result of the scene image. By adopting a multi-class object detector algorithm, a key object, a class and layout information are accurately obtained; the object semantics feature of the indoor scene image is obtained through an SFV model; and convolutional layers and LSTM layers form an end-to-end trainable hybridDNN structure, so that the context information of the scene image can be effectively extracted. The method fuses the object semantics information, the global appearance information and the appearancecontext feature, so that the identification rate and robustness of the identification method are improved.

Description

technical field [0001] The invention belongs to the technical field of image processing, and in particular relates to a scene recognition method based on fusion of target semantics and deep appearance features. Background technique [0002] Scene recognition is one of the important topics of computer vision, and it is widely used in many fields, mainly including image information retrieval of large databases, mobile positioning of robots and interaction with the environment, event detection in the field of security monitoring, etc. [0003] Since 2006, deep learning theory has become a research hotspot in the field of machine learning and artificial intelligence. Deep learning establishes a deep network structure to simulate the cognitive mechanism of the human brain, and extracts input data layer by layer through a deep multi-layer network to form The mapping relationship from low-level signals to high-level semantics. Deep learning architecture has achieved great success ...

AI Technical Summary

Problems solved by technology

[0005] However, when the above method is applied to indoor scene recognition, the recognition effect declines. The reason is that indoor scenes mainly have intra-class differences, large inter-class differences, etc.

Images