This paper presents the importance of transfer learning in a deep learning model for detection of trash of different size and shape. The detector is a real-time implementation to be incorporated for a specific robotic model developed to clean the floor autonomously. A model was trained using CNNs for the detector. The dataset, consisting of 2418 images, was divided as 65% (1572 images) for the training set, 20% (481 images) for the validation set, and 15% (365 images) for the test set. The model was assessed at 50 and 100 epochs. The results showed that transfer learning consistently improved the performance indicators. Compared to the non-transfer learning model, the transfer learning model showed a significant 26.8% improvement in mAP50-95 and a 16.8% increase in mAP50 after 50 epochs. This trend continued till the whole training period, i.e., 100 epochs, highlighting the long-term advantages of transfer learning. These results were further supported by evaluation on the testing dataset, which showed a 14.1% increase in mAP50-95 and a 7.5% increase in mAP50 for the transfer learning model. Transfer learning outperformed performance metrics by demonstrating a 32.21% decrease in box loss during training. The study also demonstrated how transfer learning shortens training times and reliably achieves target mAP in fewer epochs.