**3. Methodology**

### *3.1. Study Setup*

The CNN in single shot multibox detector (SSD) method of deep learning required massive training images for learning. Firstly, image files of a construction site were collected and converted to matrices by regulating the size of images before data pre-processing, such as optimization. Next, features were extracted using CNN and fed into the fully connected neural network to predict and identify classes. Finally, the trained model was verified by feeding it the test data. The model's learning rate setting would affect the weight adjustment, so this study set the learning rate = 0.00002, epoch = 100, step per epoch = 320, and optimizer type = sgd. The model structure of this proposed method is presented in Figure 8.

**Figure 8.** Structure of job site image object detection model.
