Deep transfer learning based and MFCC based acoustic detector of rice weevils, Sitophilus oryzae (L.) in stored grains

• Main Author: Rabano, Stephenn L.
• Format: text
• Language: English
• Published: Animo Repository 2018
• Subjects: Acoustic imaging; Deep learning (Machine learning); Rice weevil—Detection; Electrical and Computer Engineering
• Online Access: https://animorepository.dlsu.edu.ph/etd_masteral/6345; https://animorepository.dlsu.edu.ph/context/etd_masteral/article/13418/viewcontent/Rabano_S_MS_ECE_Thesis2_Redacted.pdf
• Institution: De La Salle University

Rice postharvest losses affects the sustainability in human consumption of rice. About one quarter to one third of the world grain crop is lost every year during storage, and most of this is caused by insect infestation. Even though there are grains that remain after the feeding of these insects, grain quality is also reduced by damages from the insects. One of these rampant insects is the rice weevil. Rice fumigation and other measures are done to avoid further infestation once the presence of rice weevils has manifested. Immediate recognition of the presence of these insects will prompt for immediate fumigation so that losses in rice postharvest are reduced. This study compared the accuracy of using an MFCC-based model with the accuracy of using deep transfer learning in the detection of rice weevil. An STM32 NUCLEO-F401RE development board was used along with an X-NUCLEO-CCA02M1 expansion board for audio acquisition. The recording was done by the open source audio software Audacity 2.2.2. MFCCs were extracted from the dataset of audio recordings with two clusters, one with rice weevil sound (positive cluster) and another without rice weevil sound (negative cluster). Classification was done using logistic regression on the MFCCs of the audio files. For deep transfer learning, the same dataset was used to generate spectrogram images of the audio files. Some Keras pre-trained models (Xception, VGG16, VGG19, ResNet50, InceptionV3, InceptionResNetV2, and MobileNet), trained on ImageNet, were used on those spectrogram images for feature extraction and prediction. The programming platform that will be used in both methods is Python 3.6.5. Considering preprocessed, unprocessed, and attenuated data, InceptionResNetV2 and MobileNet had the fastest prediction time at 171.8 ms. VGG19 had the highest average prediction at 99%, the highest average recall at 99%, and the highest average F1 score also at 99%. Using the same audio data, the MFCC-based model got an average precision of 97%. The same percentage was also the value of its average recall and average F1 score. This model performed at par with the Keras pre-trained models.