Multi-scale face detection based on convolutional neural network

To achieve faster and more accurate face detection, this paper proposes an end-to-end multi-scale face detection algorithm based on a convolutional neural network (CNN). First, a layer of data pre-processing was added between the convolutional layers, and the pixel values of the original image were inverted and then XORed with the original image value. Second, an improved inception structure was added between the first and second convolutional layers, thus reducing dimensionality during the extraction of richer and deeper features. Third, a spatial pyramid pooling layer was added between the convolutional layer and the fully connected layer, to solve the multi-scale problem. Fourth, the image was divided into multiple grids, and faces were detected by multiple prediction frames on each grid. Finally, the loss function commonly used in the detection algorithm was improved by modifying the loss function in the detected algorithm, to improve the accuracy of the model. The test results of public face datasets revealed that this method could extract meaningful facial features and demonstrate a superior performance to that of several competitive face detection algorithms.