000			03548nas a2200421 a 4500
005			20260818094231.0
008			240229s2022 th u m tt 000 a eng d
035			_a.b12421728
099			_aAIT Thesis no.DSAI-22-10
100	0		_aSarach Rujiranurak
245	1	0	_aDurian ripeness prediction using audio processing and deep learning
260			_aPathum Thani, Thailand : _bAsian Institute of Technology, _c2022
300			_a43 leaves : _bill.
490	1		_aThesis ; _vno. DSAI-22-10
500			_aA thesis submitted in partial fulfillment of the requirements for the degree of Master of Science in Data Science and Artificial Intelligence, School of Engineering and Technology
502			_aThesis (M. Sc.) - Asian Institute of Technology, 2022
520			_aDue to its taste and odor, durian is a high-value fruit product in Thailand. With their custard-like pulps, durians can be eaten at various ripeness stages, giving a variety of sweetness appreciated by consumers. However, determining the level of ripeness of a durian is complicated. The fruit is covered with hard spikes and a thick peel, making it challenging to observe its condition from the outside. In practice, experts knock a durian with a rubber stick and listen to the sound that is typical to distinguish ripe and unripe fruit. This technique requires a great deal of experience on the part of the listener to make accurate predictions. This study indicated the audio processing and deep learning techniques applied for durian ripeness detection from the knocking sound. The audio samples were recorded from 611 Mon-thong durians harvested from the south region of Thailand, Chumphon, Surat Thani, and Yala provinces. There were four classes of ripeness levels, unripe, mid-ripe, ripe, and overripe, for the target of the deep learning models. The sound from each hit was extracted and transformed into three forms of spectrograms, Short Time Fourier Transform (STFT), mel spectrogram, and Mel Frequency Cepstral Coefficients (MFCCs). Two deep learning models, ResNet50 and Audio Spectrogram Transformer (AST), were evaluated for their performance when trained with different data structures and learning procedures. This study found that ResNet-50 provided the best prediction at 74.5% accuracy when predicting in four levels and 89.4% accuracy when predicting two classes when grouping the targets into unripe and ripe conditions.
650		0	_aComputer sound processing
650		0	_aNeural networks (Computer science)
650		0	_aDeep learning (Machine learning)
650		0	_aDurian _zThailand
700	1		_aDailey, Matthew N., _eChairperson
700	0		_aChaklam Silpasuwanchai, _eExamination Committee
700	0		_aLoc, Thai Nguyen, _eExamination Committee
710	2		_aThe Royal Thai Government, _eScholarship Donor
810	2		_aAsian Institute of Technology. _tThesis ; _vno. DSAI-22-10
856	4	0	_3Full-Text _uhttp://203.159.5.9/ait-thesis/Viewer/viewer.php?id=B20423
907			_a.b12421728 _bmnait _ca
902			_a250307
998			_b0 _c240314 _dm _ea _fa _g0
945			_lmnarc
945			_lmnarc
942			_c67
942			_c40
909			_aBarcode : - _bCREATED : 2024-02-29 _cRECORD # : i13487930 _dLPATRON : 0 _eLCHKIN : - _f# RENEWALS : 0 _g# OVERDUE : 0 _hIUSE3 : 0 _iTOT CHKOUT : 0 _jTOT RENEW : 0
909			_aBarcode : 30050120899009 _bCREATED : 2025-07-03 _cRECORD # : i13537209 _dLPATRON : 0 _eLCHKIN : - _f# RENEWALS : 0 _g# OVERDUE : 0 _hIUSE3 : 0 _iTOT CHKOUT : 0 _jTOT RENEW : 0
999			_c31864 _d31864