Comparing Performances of CNN, BP, and SVM Algorithms for Differentiating Sweet Pepper Parts for Harvest Automation

Abstract

For harvest automation of sweet pepper, image recognition algorithms for differentiating each part of a sweet pepper plant were developed and performances of these algorithms were compared. An imaging system consisting of two cameras and six halogen lamps was built for sweet pepper image acquisition. For image analysis using the normalized difference vegetation index (NDVI), a band-pass filter in the range of 435 to 950 nm with a broad spectrum from visible light to infrared was used. K-means clustering and morphological skeletonization were used to classify sweet pepper parts to which the NDVI was applied. Scale-invariant feature transform (SIFT) and speeded-up robust features (SURFs) were used to figure out local features. Classification performances of a support vector machine (SVM) using the radial basis function kernel and backpropagation (BP) algorithm were compared to classify local SURFs of fruits, nodes, leaves, and suckers. Accuracies of the BP algorithm and the SVM for classifying local features were 95.96 and 63.75%, respectively. When the BP algorithm was used for classification of plant parts, the recognition success rate was 94.44% for fruits, 84.73% for nodes, 69.97% for leaves, and 84.34% for suckers. When CNN was used for classifying plant parts, the recognition success rate was 99.50% for fruits, 87.75% for nodes, 90.50% for leaves, and 87.25% for suckers.

Special Issue on the Engineering of Smart Agriculture

“…Using techniques such as normalized difference vegetation index (NDVI) and local feature analysis, this study compares the performance of different algorithms. The results demonstrate the capabilities of these algorithms, with significant successes achieved by the convolutional neural network (CNN) approach [26]. An architecture potentially capable of optimizing productivity has been proposed, as it uses simulation software to optimize (i) climate control strategies related to crop microclimate control and (ii) crop management treatments [27].…”

Detection of Material Degradation of a Composite Cylinder Using Mode Shapes and Convolutional Neural Networks

“…At this time, however, this drawback of CNNs is not crucial. The advantages of CNNs are even more obvious when image analysis is necessary, as shown, e.g., in [53][54][55], the performance of a CNN is superior to both shallow networks and support vector machines (SVMs).…”