Research On Shoeprint Image Based Gender Prediction Method

With the advent of the era of artificial intelligence,biometric identification technology is widely used in areas such as social governance and business activities.Biometric recognition technology is a technology to identify human individuals based on their behavioral characteristics（gait,voice,footprint,etc.）or physiological characteristics（face,fingerprint,iris,etc.）,which can obtain one’s personal information such as identity,gender,height and age of individuals.Gender prediction is a kind of biometric technology,and it mainly includes face-based gender prediction,gait-based gender prediction and footprint-based gender prediction,among which,footprint-based gender prediction is widely used in the field of criminal investigation.Efficient and accurate gender prediction plays a crucial role in the detection of criminal cases.Footprint refers to the traces formed by the contact between human foot and the surface of the bearing body such as the ground during walking,which has the characteristics of high retention rate and rich behavioral and physiological information.Footprints can be divided into barefoot footprints,footprints in socks and footprints in shoes（shoeprints）.Because of the high retention rate of shoeprints at the crime scenes and the current gender prediction method based on shoeprints mainly relies on footprint experts to carry out,this thesis intends to study the automatic gender prediction method based on shoeprints,and the main work includes.（1）A shoeprint image dataset is constructed,and a footprint preprocessing method is proposed.A single shoeprint image dataset and a sequential shoeprint image dataset are constructed respectively.A shoeprint image centering preprocessing operation is performed on the single shoeprint image dataset to construct a centered single shoeprint image dataset（CSiSIS）;a shoeprint tread energy map generating operation is performed on the sequence shoeprint image dataset to construct a sequence shoeprint tread energy map dataset（SeSTEMS）.（2）Geometric features and Stacking model based gender prediction method is proposed.The method is divided into two input scenarios: single shoeprint and sequential shoeprint.For the scenario with only a single shoeprint image,shoe length,shoe width,centroid distance,forefoot length,forefoot width,backfoot length and backfoot width are extracted as geometric features for gender prediction;for the scenario with a sequence of shoeprint images,shoe length,shoe width,centroid distance,step length,step width and step angle are extracted as geometric features for gender prediction.Random forest,support vector machine,gradient boosting decision tree and logistic regression classifier cooperate to form a Stacking model for gender prediction.Gender prediction accuracies of 89.46% and 89.59% have been achieved on two types of shoeprint image geometric feature datasets,respectively,which are higher than other existing gender prediction methods based on geometric features.（3）Convolutional neural network and channel attention mechanism based gender prediction method is proposed.The proposed method first obtains the feature maps of the shoeprint images by the feature extraction module,then uses the channel attention module to reassign the feature weights of each feature map,and finally uses the gender prediction module to obtain the gender prediction results.Shoeprint forming mechanism-based data augmentation method is proposed to augment two kinds of datasets.The gender prediction accuracies of 91.80% and 99.25% have been achieved on the two types of shoeprint image datasets,respectively,which are higher than the existing similar methods and the geometric feature-based gender prediction methods.