Dermoscope image retrieval method based on Cauchy anti-rotation loss function

Abstract

The invention provides a dermatoscope image retrieval method based on a Cauchy anti-rotation loss function. The dermatoscope image retrieval method comprises the following steps: 1, establishing a dermatoscope image data set; 2, designing a Cauchy anti-rotation CAR loss function; 3, designing a convolutional neural network structure based on a Cauchy anti-rotation loss function; 4, performing network training; and 5, carrying out dermatoscope image retrieval. The loss function enables the network to directly learn the similarity between image pairs, so that the extracted hash code has the characteristics of intra-class compactness and inter-class separation. The designed rotation invariant loss item can restrain the output difference of the network learning different angle sample data so as to obtain certain rotation invariance. Through the steps, the similar image of the dermatoscope image to be retrieved can be accurately and quickly obtained from the database, the attached diagnosis report provides objective suggestions and references for dermatologists, and the diagnosis accuracy is improved.

Description

technical field [0001] The invention belongs to the field of dermoscopic image processing, in particular to a dermoscopic image retrieval method based on a Cauchy Anti-rotation (Cauchy Anti-rotation, CAR) loss function. Background technique [0002] As the largest organ of the human body, the skin covers the whole body and is an important barrier to prevent pathogens from invading and water loss in the body. Affected by multiple factors, various skin diseases are violating people's health. Dermoscopy is a non-invasive microscopic image analysis technique for observing submicroscopic structures and pigments on and below the surface of living skin, which is of great significance for clinical diagnosis in dermatology. At present, clinical dermatology mainly relies on doctors to use dermoscopy technology to diagnose the disease by observing the color, texture and other characteristics of the skin lesion area, combined with their own subjective experience. This method is subjec...

AI Technical Summary

Problems solved by technology

Dermoscopy image retrieval based on deep learning is mostly based on classification loss for network training, and the obtained high-level features lack the similarity between similar images, so the retrieval accuracy still does not meet the clinical requirements, and there is a lot of room for improvement

[0005] The characteristic of no main direction of the skin lesion target leads to the rotation of a large number of skin lesion targets in the dermoscopic image, and because the convolutional neural network does not have rotation invariance, it is more sensitive to the change of target rotation, which greatly affects the learning ability of the network

Images