A Thermal Imaging Super-resolution Reconstruction Method Fused with Visible Light Image Gradient Information
A super-resolution reconstruction and image gradient technology, applied in the field of image processing, can solve the problems of image structure distortion, lack of detail information, loss, etc., to enhance the expression ability, reduce the amount of noise, and improve the quality.
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
Embodiment 1
[0044] A thermal imaging super-resolution reconstruction method that fuses gradient information of visible light images, comprising the following steps:
[0045] Step S100: collect data and construct training samples, the training samples are composed of gradient maps of low-resolution visible light images and low-resolution thermal imaging images, and the gradient maps of low-resolution visible light images and low-resolution thermal imaging images constitute training sample pair;
[0046] Step S200: using the backbone network of the deep neural network to extract the deep convolution features of the training samples, such as figure 1 As shown, the backbone network includes a generation network part and a discrimination network part; figure 1 G in is the generation network part, and DI is the discriminative network part;
[0047] Step S300: Input the training sample into the generation network part, and perform fusion on each intermediate layer of the gradient map branch ne...
Embodiment 2
[0051] This embodiment is optimized on the basis of embodiment 1, such as figure 2 As shown, the generation network part in step S200 is composed of a convolutional layer, an activation function layer, a batch normalization layer, a residual dense block, an upsampling layer, and a feature splicing layer arranged sequentially from front to back; image 3 As shown in , the residual dense block is composed of a convolutional layer, a feature weight multiplication layer, and a feature addition layer stacked sequentially from front to back. Figure 2-Figure 4 C1-C9 are convolutional layers, BN1-BN9 are batch normalization layers, RDB_Block1-RDB_Block6 are residual dense blocks, Concatenate is feature splicing fusion layer, LR1-LR9 are leakage correction linear unit layers, UP1-UP4 are The upsampling layer, Mul is the feature weight multiplication layer, and Add is the feature addition layer.
[0052] Further, the generating network part in step S200 adopts a parameter sharing mec...
Embodiment 3
[0056] This embodiment is optimized on the basis of Embodiment 1 or 2. The discriminant network part in step S200 adopts a VGG network structure, and is composed of a convolution layer, an activation function layer, and a batch normalization layer arranged in sequence from front to back. . Such as Figure 4 As shown, in order to reduce the phenomenon of neuron deactivation, the activation function layer used in the traditional VGG structure is modified. Its structure consists of convolutional layer, activation function layer, and batch normalization layer from front to back, and its input is to generate the network part output and ground truth labels.
[0057] Further, there are 23 residual dense blocks in the generating network part, and 9 convolutional layers in the discriminant network part.
[0058] Further, the activation function layers of the generation network part and the discriminant network part in step S200 are both leakage corrected linear unit layers.
[0059]...
PUM
Login to View More Abstract
Description
Claims
Application Information
Login to View More