In the crack defect detection project map 0.954 was improved to 0.979 , the increase point was obvious, and the bloggers tested multiple data sets effectively, achieving a violent increase point;
1.VanillaNet
Paper: https://arxiv.org/pdf/2305.12972.pdf
Researchers from Huawei Noah and the University of Sydney proposed a minimalist neural network model, VanillaNet, with a minimalist design as the concept. The network only contains the simplest convolution calculations, and removes residuals and attention. module, which has achieved good results in various tasks in computer vision.
VanillaNet, an elegantly designed neural network architecture. By avoiding complex operations such as high depth, shortcuts, and self-attention, VanillaNet is concise yet powerful.
- For the Stem part, 4×4 convolution is used for feature transformation;
- For each stage of the body part, first use MaxPool for feature downsampling, and then use a 1×1 for feature processing;
- For the head part, two nonlinear layers are used for classification processing