Fast and accurate convolution neural network for detecting manufacturing data

Abstract

This article introduces a technique known as clustering with particle for object detection (CPOD) for use in smart factories. CPOD builds on regional-based methods to identify smart object data using outlier detection, clustering, particle swarm optimization (PSO), and deep convolutional networks. The process starts by removing noise and errors from the images database by the local outlier factor (LOF) algorithm. Next, the algorithm studies different correlations from the set of images in the database. This creates homogeneous, and similar clusters using the well-known k-means algorithm, and the FastRCNN (fast region convolutional neural network) uses these clusters to design efficient and more focused models. PSO is used to optimize the different parameters including, the number of neighbors of LOF, the number of clusters of k-means, the number of epochs, and the error learning rate for FastRCNN. The inference process benefits from the knowledge provided by training. Instead of considering a complex single model of the whole images database, we consider a simple homogeneous model. To demonstrate the usefulness of our approach, intensive experiments have been carried out on standard images database, and real smart manufacturer data. Our results show that CPOD when compared to baseline object detection solutions is superior in terms of runtime and accuracy.