Supervised anomaly detection by convolutional sparse representation

Abstract

Anomalies are known as patterns which are not conforming expected structures; therefore, they must be detected to inhibit future hazards. A common method for defect detection in images is background modeling in which characteristics of normal structures are captured and employed for detecting anomalous regions. In this paper, we utilize convolutional sparse representation method for capturing normal structures and feature extractions. Indeed, normal patterns are recorded as atoms of a dictionary in the phase of dictionary learning. Then for abnormality detection, new images are divided into patches, which encompass normal and/ or abnormal structures, and afterwards they are encoded into coefficient maps with respect to the learned dictionary atoms, in the phase of sparse coding technique. Subsequently, coefficient maps are utilized for the purpose of feature extraction and finally, anomaly detection will be done using supervised learning approaches. Simulation results demonstrate remarkable capabilities of the proposed approach. It is worth mentioning that the presented approach leads to higher True Positive Rate (TPR) and lower False Positive Rate (FPR) up to 0.95 and 0.005 respectively in comparison with reviewed methods.