2nd ICAI 2022

International Conference on Automotive Industry 2022

Mladá Boleslav, Czech Republic

based on large-scale data. Transfer learning (TL) is an optimization stage that allows rapid progress or improved performance when modelling the second task. Transfer learning is a machine learning (ML) research problem that focuses on solving one problem, preserving the knowledge gained, and applying it to another related problem. Figure 3 presents the basic diagram of transfer learning (Tan et al ., 2018).

Figure 3: Transfer learning diagram

Source: Own elaboration The basic aim of the implementation was to minimize financial outlays in reduction of infrastructural and personnel costs. Therefore, the implementation activities were based on deep transfer learning methods from pretrained convolutional networks because of the sparse representation of objects that could be used in training the network from scratch In the documentation of the Keras library, we can find the basic definition of the next stages of the TL process. The stages are described below: 1. Obtaining the weights of the model from the network trained on correlated data. 2. Freezing weights for layers containing feature generalization to avoid weight recalculation in the process of training a new network. 3. Adding new layers on top aimed at correct representation of the model output data. 4. Training model in domain dataset. 5. Unfreezing of selected layers in order to achieve a better match of weights to new data(optional step) (Chollet, 2020). As described above, the use of TL methods is not a complicated task, and in tasks with correlated data it can significantly reduce the time of prototyping or implementing unique solutions. 2.3 YOLO3 Architecture Implementation The technique of detecting the recognizing, area, and class of at least one object within an image is known as object detection. Every individual object in the image is labelled as a class object by the detection method, which forms a bounding box around them. YOLO is an object detection approach that use a deep convolutional neural network (Menon, Omman & Asha 2021). YOLO is a deep learning-based single-shot object detection approach formed in 2016. The term “one-shot detection” refers to

111