Barcodes are ubiquitous and have been used in most daily activities for decades. However, most traditional barcode scanners require well-founded barcode under standard conditions. While wilder conditioned barcodes such as underexposed, occluded, blurry, wrinkled, and rotated are commonly captured in reality, those traditional scanners show weaknesses of recognizing. This work aims to solve the detecting and decoding problem using a deep convolutional neural network with the possibility of running on portable devices. To be more specific, in the barcode decoding problem, we proposed a special modification of inference based on the attribute of having self-validation (checksum) in the prediction phase of a trained model; on the other hand, to also cover the full flow of detecting and decoding, we introduced the one-stage model based on the YOLO object detection model which is more efficient and faster than those of two-stages. The later method not only works for barcode but also similar applications such as license plate capturing and reading. To prepare for a reliable evaluation, we collected and annotated a large number of barcodes as benchmark datasets. Experimental results prove models' efficiency by outperforming industrial-standard tools in both detection and decoding rates with a smooth fps on portable devices.