Face detection method and apparatus, computer device, and storage medium

A face detection method for a computer device, comprising:

• constructing a cascaded regressive probability regression model according to a sample image set, wherein the probability regression model comprises multi-level random forests, and each level of the random forests comprise a plurality of regression trees;
• the constructing a cascaded regressive probability regression model according to a sample image set comprises:
• selecting a first pixel pair collection randomly from the sample image set when constructing each level of the random forests;
• dividing sample images in the sample image set according to the first pixel pair collection, splitting to generate each of nodes of a regression tree, configuring a probability of each of leaf nodes of a generated regression tree, and determining an error of each of the leaf nodes according to a true shape of a sample image allocated to each of the leaf nodes and an initially predicted shape of the generated regression tree;
• and
• calculating a predicted shape of the generated regression tree according to the initially predicted shape of the generated regression tree and the error of each of the leaf nodes, taking the predicted shape as an initially predicted shape of a neighboring next regression tree and iteratively generating the next regression tree until generation of a regression tree of a last level of the random forest in the probability regression model is finished;
• obtaining a to-be-detected image;
• obtaining an initial shape of a current regression tree in a pre-constructed the probability regression model;
• extracting image features from the to-be-detected image and respectively calculating a probability of each of leaf nodes of the current regression tree according to the image features;
• extracting an error of each of the leaf nodes from the current regression tree;
• determining a shape error of the current regression tree according to the probability and the error of each of the leaf nodes;
• calculating an estimated shape of the current regression tree according to the initial shape and the shape error;
• and
• performing iterative calculation by using the estimated shape as an initial shape of a neighboring next regression tree until a last regression tree in the probability regression model, to obtain an estimated shape of the last regression tree as a detected face shape from the to-be-detected image.