Lightweight Deep Learning Approaches For Internal Mango Defect Detection Using Image Segmentation

Abstract

Precision in identifying internal blemishes in mangoes is one of the most significant issues that cannot be overcome by automated quality evaluation of the products, especially export based grading whereby manual evaluation is inefficient, subjective, and destructive. This research will be conducted with the aim of coming up with and testing lightweight deep learning methods to achieve reliable and internal mango defect detection without compromising on accuracy or complexity of the computations. Internal image dataset of mangoes, comprising of both normal mangoes and defective ones, was created by self and images were preprocessed to remove noise, normalization, and resizing to segment defect-prone areas using the OpenCV. GLCM contrast, correlation, energy, homogeneity, local binary patterns, as well as texture features were extracted and data augmentation methods were used to enhance model generalization. An out-of-the-box custom CNN was used and tested against lightweight transfer learning-based architectures such as MobileNetV2-Lite, EfficientNetB0-Lite, NASNet-Lite, DenseNet121-Lite and ResNet50-Lite when trained under the same conditions. Experimental data indicates that segmentation is a performance boosting strategy in all the models. MobileNetV2-Lite obtained the highest results with a score of 98% accuracy, 97% precision, 99% recall, and a F1-score of 98% with segmentation and only 2.34 million parameters and model size of 8.93 MB. Other designs like NASNet-Lite and DenseNet121-Lite attained 93 and 85 percent accuracy respectively whereas the baseline CNN attained 85 percent. The results confirm that lightweight transfer learning models yielded with segmentation yield an efficient, scalable, and computationally efficient method of internal mango defect detection, thus is adequate and applicable to the real-time agricultural quality inspection systems.