An improved deep learning approach for localization and recognition of plant leaf diseases

Abstract

A nation's economic progress is significantly influenced by its percentage of crop yields. However, the major barrier to the quantity and quality of yield is crop disease. For quick and reliable recognition of various plant illnesses, it is mandatory to design a computer-aided system. Timely and accurate recognition of numerous crop leaf infections is a complicated job because of the presence of vast sample distortions like the prevalence of clutter, blur, texture, and luminance changes in samples. Moreover, the extreme resemblance between the normal and infected parts of visual samples also extends the difficulty of the identification procedure. Further, the massive differences in the size, structure, and orientation of crop leaves and infected areas also hinder the accurate recognition of various crop diseases. To deal with the listed issues, we have proposed an improved and effective deep-learning strategy namely the PlantRefineDet. Our approach comprises three steps. First, the sample annotations are created for defining the target object. Next, an improved RefineDet approach is presented that employs the ResNet-50 as its base network for extracting a set of deep features. Lastly, the one-step detector RefineDet is utilized to localize and classify numerous crop disorders. The PlantRefineDet approach improves the plant disease localization and categorization results because of its improved feature calculation ability which facilitates the reemployment of features from the previous layers and increases the recall power of the system. Also, the PlantRefineDet approach adopts an additional phase to eliminate the irrelevant anchors and better adjust the bounding box orientation to exactly locate the infected regions of plant leaves which result to improve the recognition performance of the introduced model. We have confirmed the effectiveness of our approach through extensive evaluation on a challenging PlantVillage data sample and obtained a remarkable accuracy of 99.994%.