Deep Learning Automatic Inspections of Mushroom Substrate Packaging for PP-Bag Cultivations

Abstract

The mushroom cultivation is an important smart agriculture in Taiwan. This study uses the deep learning object detection method to inspect the cap flaws or positional imperfection in the automatic production of the mushroom PP-bag packaging. This study uses the UR robotic arm and integrated 3D vision module, and uses the extra positioning axis to achieve the purpose of multi-positioning inspections by robot arm. Projecting the structured LED light sources to the object to be inspected has the advantages of a larger identification ranges and complex objects detection. A duallens CMOS industrial camera is used to capture images, and a 3D point cloud image of a basket of PP-bag packages is created by software calculation, which can obtain detailed information on the appearance of the whole basket of PP-bag packages. Deep learning is performed by the training set with labelling, and the image recognition such as the cap flaws in the PP-bag package or positional shift is performed after the training is completed. In this paper, the image data is divided into four sets of datasets, and the same training parameters are used for individual training. With images of dataset1 and the ambient illumination level of 200 lm to 800 lm, the matching score is up to 0.989. The clamping force and the opening degree are adjusted by the variable jaws. The clamping force of the jaws is maintained at 20 N to prevent the clamping force from damaging the dimensions of the PP-bag package and existing holes inside it, making the product unusable. Using the variable jaws and repeating 30 times of clamping experiments, the hole diameter inside the PP-bag package can still be maintained within around 25 mm, which can meet the needs of the mushroom PP-bag packaging.