"AUTO-ENCODER BASED CONVOLUTIONAL NEURAL NETWORK (AECNN) FOR MULTISPECIES FRUIT FLOWER DETECTION"

On trees, process of counting and detecting fruits count defines the task of crop estimation. At various locations, manual fruits and vegetables counting is performed in recent days. Excessive amount of labor requirement and consumption of more time are the major drawbacks of this manual counting. Bloom intensity which corresponds to flowers count present in orchard guides critical crop management decisions in the production of fruits. Even with high importance, human visual inspection is used for estimating bloom intensity. Hand engineered methods forms base for the automated computer vision system used for identifying fruits. Under specific conditions only, their performance is defined and it is a limited one. Major drawbacks of existing methods are, fruit flower images are selected without the reduction of dimensionality, in predictions, objects orientation and positions are not encoded by CNN.For detection of fruit flower, Auto-Encoder based Convolutional Neural Network (AECNN) and Animal Migration Optimization (AMO) algorithms are proposed for solving these issues. There are three main stages in this work. They are, training of network for deep Fully Convolutional Network (FCN), pre-processing, reduction of dimension and segmentation. On commercial GPU, high resolution image evaluation procedure are defined with deep FCN. GPU memory space is needed for fully convolutional computations and based on image resolution, there will be an exponential increase. From image, noises are reduced or removed using an Additive white Gaussian Noise (AWGN) technique which is introduced next. At last, small patches are formed by splitting high resolution image, AMO is used for reducing dimensionality, fine-tuned AECNN is used for evaluating every patch and final segmentation is performed by applying refinement algorithm. Three datasets of peach and apple flowers are used in experimentation. Metrics like Intersection-over-Union (IoU), F-Score(F1), Recall(R), Precision(P) are used for measuring the results.