As technology evolves, more components are integrated into printed circuit boards (PCBs) and the PCB layout increases. Because small defects on signal trace can cause significant damage to the system, PCB surface inspection is one of the most important quality control processes. Owing to the limitations of manual inspection, significant efforts have been made to automate the inspection by utilizing high resolution CCD or CMOS sensors. Despite the advanced sensor technology, setting the pass/fail criteria based on small failure samples has always been challenging in traditional machine vision approaches. To overcome these problems, we propose an advanced PCB inspection system based on a skip-connected convolutional autoencoder. The deep autoencoder model was trained to decode the original non-defect images from the defect images. The decoded images were then compared with the input image to identify the defect location. To overcome the small and imbalanced dataset in the early manufacturing stage, we applied appropriate image augmentation to improve the model training performance. The experimental results reveal that a simple unsupervised autoencoder model delivers promising performance, with a detection rate of up to 98% and a false pass rate below 1.7% for the test data, containing 3900 defect and non-defect images.

Voice-activated artificial intelligence (AI) technology has advanced rapidly and is being adopted in various devices such as smart speakers and display products, which enable users to multitask without touching the devices. However, most devices equipped with cameras and displays lack mobility; therefore, users cannot avoid touching them for face-to-face interactions, which contradicts the voice-activated AI philosophy. In this paper, we propose a deep neural network-based real-time sound source localization (SSL) model for low-power internet of things (IoT) devices based on microphone arrays and present a prototype implemented on actual IoT devices. The proposed SSL model delivers multi-channel acoustic data to parallel convolutional neural network layers in the form of multiple streams to capture the unique delay patterns for the low-, mid-, and high-frequency ranges, and estimates the fine and coarse location of voices. The model adapted in this study achieved an accuracy of 91.41% on fine location estimation and a direction of arrival error of 7.43° on noisy data. It achieved a processing time of 7.811 ms per 40 ms samples on the Raspberry Pi 4B. The proposed model can be applied to a camera-based humanoid robot that mimics the manner in which humans react to trigger voices in crowded environments.

The advent of voice assistance technology and its integration into smart devices has facilitated many useful services, such as texting and application execution. However, most assistive technologies lack the capability to enable the system to act as a human who can localize the speaker and selectively spot meaningful keywords. Because keyword spotting (KWS) and sound source localization (SSL) are essential and must operate in real time, the efficiency of a neural network model is crucial for memory and computation. In this paper, a single neural network model for KWS and SSL is proposed to overcome the limitations of sequential KWS and SSL, which require more memory and inference time. The proposed model uses multi-task learning to utilize the limited resources of the device efficiently. A shared encoder is used as the initial layer to extract common features from the multichannel audio data. Subsequently, the task-specific parallel layers utilize these features for KWS and SSL. The proposed model was evaluated on a synthetic dataset with multiple speakers, and a 7-module shared encoder structure was identified as optimal in terms of accuracy, direction of arrival (DOA) accuracy, DOA error, and latency. It achieved a KWS accuracy of 94.51%, DOA error of 12.397°, and DOA accuracy of 89.86%. Consequently, the proposed model requires significantly less memory owing to the shared network architecture, which enhances the inference time without compromising KWS accuracy, DOA error, and DOA accuracy.