The Shoeprint Segmentation Via Multi-fine Scale Loss Function

Shoeprint as one of common crime traces from the case scenes,the segmentation result is helpful to extract tracks of information,access to expose and confirm the criminal evidence,and improve the detection rate of criminal cases.However,due to the rich detail characteristics of the shoeprint itself and the incomplete shoeprint and structural damage caused by the complexity of the crime scene environment,the progress of related work is slow.In recent years,with the development of deep learning method,the segmentation algorithm of natural image has developed rapidly,and its segmentation result is constantly improving.Main purpose of this article is through a reference solution to natural image segmentation,the fully convolution network,combined with the unique shoeprint image,design a suitable shoeprint segmentation model,to solve the problem of segmentation of shoeprint.On the basis of full convolution network,Through will jump layer structure and loss function to the bottom of the network structure,which has high resolution,small receptive field level,to obtain the corresponding characteristics of small scale-here say such as low dimension layer,to extract features is called the low scale,thus constructed respectively based on jump layer structure and multi-scale damage of low dimension.And through the fusion of local features generated scale across multiple receptive field,further enrich the network in various scales of feature extraction ability,construct fine scales eventually loss function the shoeprint of segmentation model.On the test set,the average accuracy and average coverage rate of the low-scale jump structure model are 40.40% and 31.99% respectively.The multi-scale loss function improves the network multi-scale feature representation by introducing the loss function,and increases the average accuracy and average coverage to 86.18% and 32.12%.Finally,by changing the topological structure on each scale of the network,as a single scale feature,the fusion of multiple sensory field features improved the average accuracy and average coverage to 90.83% and 33.17%.The experimental results show that the algorithm is able to segment the precise shoeprints,retain the fine structure features of the shoeprints,and be robust in the larger data sets.