Improved SSD for Door Panel Missed Installations Inspection Using PASSD

摘要

Abstract In the context of made in china 2025 and industry 4.0, the application research of intelligent manufacturing and process design of intelligent workshop are in a full swing, traditional manual inspection methods are time consuming, inconsistent and prone to error, so this research proposes the use of PASSD for Automatic inspection of Missed installation on Car Door panels. The proposed custom inspection algorithm is a modified Single Shot Detector (SSD) with a ResNet50 backbone, then We used the Residual Convolutional Block Attention Module (RCBAM) to reduce the influence of background factors. Then added a multi-branch dilated convolution module (MDCM) to obtain information on multi-scale receptive fields and get more contextual information, and finally we added a Progressive Attention Module (PAM) to further refine the feature representations. The PAM module takes the output of the MDCM as input and progressively refines it through a series of a three stage attention mechanism. A focal loss is used as the loss function. These improvements are done to improve the detection algorithm detection rate and reduce false positives at modest speed. A Jaka robot arm is also used to move the camera across the panel to perform multiple inspections per-cycle. The experiments conducted show promising results, improving the quality inspection process in the car door panel production industry. This research provides a foundation for future studies in the application of Computer vision for car parts inspection.