An Improved AlexNet Model with Multi-channel Input Images Processing for Human Face Feature Points Detection

We present an improved AlexNet model with multi-channel input images for face feature point detection. Faces are rich expressions, they are the most direct vehicle for expressing the diversity of human inner emotions, and they are also the most important tools for communication between people. We focus on the detection of face feature points based on convolutional neural networks. We use the Caffe framework to train and test the improved network structure. The model adopts the AlexNet, we divide the face image into 3 sub-images that overlap and each have a colour channel, The weighted average of the results obtained by inputting 3 channel sub-pictures into the model. Finally, we get the feature point coordinates of the human face. We verified our method on multiple face datasets and compare the differences between the traditional face feature detection methods and the current mainstream deep learning face feature point detection methods and their respective performance characteristics. Our network runs at the speed of 80 FPS (frame per second) which is faster than DCNN counterpart and our method gets better detection performance.