A Study on Deep Learning Models for Automatic Species Identification from Novel Leather Images

This paper introduces a new set of large-scale leather image data. Unlike the existing dataset, it comprises 7600 images with varied non-ideal behavior. The aim is to develop a versatile identification model that can efficiently determine the species from complex/practically challenging leather images. Hence, this research proposed to train a convolutional neural network (CNN) on a large-scale dataset. It performs a comparative study on five CNNs: ResNet50, MobileNet, DenseNet201, InceptionNetV3, and InceptionResNetV2. The analysis reveals that InceptionNetV3 outperforms with 98.23% accuracy and 1.71% negligible error. It also evaluates the generalization power of the trained InceptionNetV3 on the existing small-scale dataset (1200 images). Although the model is trained on non-ideal leather images, it results in 94.07% accuracy. However, learning from present and existing datasets improves the prediction rate to 98.5% accuracy. Thus, this work efficiently models a deeper CNN to predict species from leather images with ideal and non-ideal behavior. Contrary to the previous machine learning-based species prediction methods, the present deep learning method designs a fully-automated model with accurate and robust results.