Color correction method based on full convolutional neural network and feature pyramid
A convolutional neural network and feature pyramid technology, which is applied in image data processing, instrumentation, computing, etc., can solve problems such as unsatisfactory color correction effect, inability to accurately predict light source information, limited training, etc., to ensure the post-correction effect , good correction effect, good correction effect
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
Embodiment 1
[0067] Such as figure 1 As shown, this embodiment provides a color correction method based on a fully convolutional neural network and a feature pyramid, including the following steps:
[0068] Perform color correction, cropping, and data enhancement processing on the image in sequence to obtain the corresponding color cast pictures, and a training data set is composed of several color cast pictures;
[0069] Build a preset full convolutional neural network and feature pyramid model for estimating image light source information, use the training data set to train the preset full convolutional neural network and feature pyramid model, based on the preset loss function and optimizer, Optimize and adjust the network parameters of the preset full convolutional neural network and feature pyramid model until the network converges, and output the trained full convolutional neural network and feature pyramid model;
[0070] The trained fully convolutional neural network and feature p...
Embodiment 2
[0072] This embodiment is further optimized on the basis of Embodiment 1, specifically:
[0073] The images used for the training in this embodiment come from the public dataset——Shi's Re-processing of Gehler's Raw Dataset;
[0074] The color correction of the image is specifically: converting the image in the public data set from the RAW format to the 8bit image data format, using the real light source information ground_truth to process the 8bit image data to obtain a standard picture, and using the real light source information ground_truth to process the 8bit image data for processing, including:
[0075] Read the 8bit image data into RGB format [R, G, B], set the corresponding real light source information ground_truth to [real_R, real_G, real_B], then:
[0076] Gain_R=max(ground_truth) / real_R
[0077] Gain_G=max(ground_truth) / real_G
[0078] Gain_B=max(ground_truth) / real_B
[0079] R'=min(R*Gain_R,255)
[0080] G'=min(G*Gain_G,255)
[0081] B'=min(B*Gain_B,255)
...
Embodiment 3
[0105] This embodiment is further optimized on the basis of Embodiment 2, specifically:
[0106] Such as figure 2 As shown, the preset fully convolutional neural network and feature pyramid model constructed are seven layers, and its specific structure is:
[0107] The input layer connection includes the first layer of neural network with two convolutional layers, where the first convolutional layer has 32 convolutional kernels, each convolutional kernel has a size of 3*3, a step size of 2, and an activation function of Relu; the second convolutional layer has 64 convolutional kernels, each convolutional kernel has a size of 3*3, a step size of 1, and an activation function of Relu; the output of the second convolutional layer is connected to the second layer of nerves the input of the network;
[0108] The second layer of neural network includes two branches and adder A, one of which includes a convolution layer with 128 convolution kernels, each convolution kernel size is...
PUM
Login to View More Abstract
Description
Claims
Application Information
Login to View More