Deep learning approach with image noise reduction to determine planting density and defected paddy seedlings

In recent years food security issues caused by climatic changes, human resources, and production costs require a strategic approach. The emergence of artificial intelligence due to the capability of recent technology in computer processing could become a new alternative to current solutions. Optimal planting density is a crucial factor for paddy cultivation as it will influence the quality and quantity of production. There have been several studies involving planting density using computer vision and remote sensing approach. While most of the studies have shown a good overall performance, they have disadvantages and room for improvement. Among the disadvantages is that the studies aim to detect and count all the paddy seedlings to determine planting density. The defective paddy seedling locations are not pointed out to help farmers in the sowing process. All the previous studies for plant detection used two-stage method object detectors like Faster R-CNN and Mask R-CNN combined with different types of feature extractor architecture such as VGG16, ResNet, Inception and MobileNet. Even though this type of object detector showed high accuracy it has a high inference time. Exploring a one stage method object detector such as Single Shot Detector could solve the high inference time for a real-time defective crop detection and may lead towards the development of an autonomous mobile transplanter. One stage method object detector however tends to show a slightly lower accuracy compared to two stage method. For this work we start with a preliminary study to evaluate the performance difference between machine learning that involve the selection of filters as feature extractor and deep learning technique that eliminates the requirement of hand-crafted features engineering on the classification of paddy fields health as our based line for the classification and detection of defected paddy seedlings. All pretrained Deep Learning models achieved better accuracy when compared to Machine Learning with filters for feature extraction and custom Deep Learning Model. We then continued our study to find the best method for paddy seedlings density classification task. All the pre-trained models showed comparable accuracy with the Base Model when trained to classify a new object with hyperparameter tuning and our proposed image pre-processing which reduces noise and enhanced input images features. Transfer Learning also allowed faster training processes by reducing the number of trainable parameters and still capable of achieving good overall performance. The next following study is on the detection of defected paddy seedling using both one stage and two stage method pre-trained object detector model. The objective of the study is to propose a method that can accurately detect and count defective paddy seedlings to determine the sowing location. Four combinations were used, the EfficienDet-D1 EfficientNet that utilizes Bi-directional Feature Pyramid Network and Compound Scaling Constant method outperforming all other models. Our image pre-processing technique managed to enhance the performance of all state-of-the-art pre-trained object detector models. Exploratory research was conducted to propose a robust computer vision approach to classify paddy seedlings density and defected paddy seedlings detection using pre-trained CNN model with transfer learning. Data augmentation, image pre-processing and hyperparameter tuning were applied to achieve the desirable performance. The experiment showed that pre-trained DCNN model works well for classification and detection of a new task with a good overall performance using transfer learning with fine tuning method.

Text 113137.pdf Download (872kB)

Actions (login required)

View Item