Channel and space fusion perception-based deep learning target detection method

Abstract

The invention relates to a channel and space fusion perception-based deep learning target detection method, which comprises the following steps: constructing a channel and space fusion perception module, embedding the channel and space fusion perception module into a deep neural network architecture, and performing target detection on a target picture by using the reconstructed deep neural networkarchitecture, wherein the construction of the channel and space fusion sensing module specifically comprises the following steps: firstly, carrying out channel sensing on an originally input featuremap, and then, carrying out cascading of space sensing. According to the channel and space fusion perception-based deep learning target detection method, the depth or width of the network is not deepened, extra space vectors are not introduced, and meanwhile the real-time performance and precision are guaranteed.

Description

technical field [0001] The invention relates to the technical field of image recognition, in particular to a deep learning object detection method based on channel and space fusion perception. Background technique [0002] At present, the target detection framework based on deep learning is mainly divided into two categories: two-stage detectors and single-stage detectors; two-stage target detection is named for its two-stage processing of pictures, also known as region-based methods, which It is to abstract the detection into two processes. One is to use random selection based on the picture to propose several areas that may contain objects, that is, the local cropping of the picture, which is called the candidate area; the other is to generate the feature vector of the area through a deep convolutional neural network. After encoding, it is used to predict each category of the candidate area, so as to obtain the category of objects in each area. The two-stage detector algo...

AI Technical Summary

Problems solved by technology

[0004] In addition to simply relying on network depth, there are currently many methods to improve network performance by designing functional modules that improve feature representation capabilities: FPN combines deep features with shallow features, and strengthens spatially powerful shallow features through richer semantic information at a deeper level. Layer features; DSSD uses a deeper basic network ResNet-101 and deconvolution layer feature fusion on the basis of SSD, while skip connections give shallow feature maps better representation capabilities, but the performance is significantly improved while the speed is significantly reduced These methods no longer deepen the model to enhance the feature representation of the network, but directly strengthen the learning of deep features in the convolutional neural network by directly superimposing, sampling, and connecting the feature maps to improve performance.

Images