The Feature Extraction Method in Power Spectrum Images for Connector Parts

Abstract

This paper proposes a method to improve classification performance for an automatic classification system of surface mount devices (SMDs) using power spectrum images. Automatic classification methods for the type of SMD are required as preprocessing for mounting SMDs in chip mounters. Due to the variations of connector parts, the previous method has a low accuracy rate. This paper presents a classification method for parts images categorized as odd form under the previous method where HLAC feature from the power spectrum of an SMD image was used to extract the periodicity of leads. In the comparative experiment using HLAC, LBP features, up to 98.2[%] of the discrimination performance of presence of leads was obtained in the case of using HLAC features with power spectrum images. In the evaluation of robustness to the pitch and width of the lead, the higher average of true positive rate (91.9[%]) to 3 classes clustered by k-means method was obtained when using HLAC, compared to LBP (78.6[%]). By incorporating the proposed method using HLAC into automatic classifier, it is possible to classify in accuracy rate of 93.2[%] for the parts they are classified as odd form by previous method.