Fast and Compact Model for Palm Bunch Classification Using DT-CWT and LSTM on Hue Histogram

Abstract

Traditionally, the grading process of fresh palm bunches has used human observation at palm trading sites, which is not precise and unreliable due to human bias. Several studies have presented automatic grading methods based on image processing and machine learning. Unfortunately, these models are not sufficiently compact to be implemented on trading sites. Therefore, a compact grading model for automatic grading suitable for implementation at palm trading sites is presented. Two key elements make the model more compact and efficient. The first element is a reduction in the size of the input dataset. We achieved this by replacing the multi-dimensional RGB palm bunch image with a one-dimensional hue histogram. The second element was the core engine of the proposed classifier. It consists of a dual-tree complex wavelet transform (DT-CWT) connected to the LSTM back end as the front end. The proposed model was proven using real image datasets of 800 palm bunches of several varieties collected from trading sites. The robustness of the model was investigated by verifying its accuracy in several noisy environments. Based on the testing process, the proposed model achieves 91.67% accuracy at 6 dB signal-to-noise ratio (SNR).

Keywords

• Feature extraction
• deep learning
• deep neural networks
• modeling