Comparing deep-learning networks for apple fruit detection to classical hard-coded algorithms

During recent years, with the increase of production in agriculture, the need for more precise tools and practices has increased. One of those practices is the estimation of the fruit number in the tree. Computer vision techniques such as histogram of oriented gradients and edge/color detection have been used to extract features thus recognizing fruit based on shape and color. Existing methods usually rely heavily on computing multiple image features, making the whole system complex and computationally expensive. In this paper we compare those classical detection algorithms to new state-of-the-art convolution neural networks. Specifically, we compare two types of algorithms for apple detection in the tree. The first approach refereed as hard-coded uses commonly feature extraction filters (edge detector, color filtering, corners). On the other side are techniques using CNNs convolution neural networks like (residual networks, sliding window, regional dividers). More than thousand images of apple trees were taken during the season from flowering time to harvest. Same pictures have been processed through both techniques and based on results and the trade-offs of both techniques have been compared. For hard-coded algorithms, with few pictures we were able to see the performance of algorithm, while with CNNs, huge number of labeled pictures were needed for the algorithm to be more than 50% accurate. However, when a different picture from another date or completely new cultivar was used, the hard-coded algorithm failed to detect thus had to be rewritten to accommodate new changes. In other hand CNNs were very flexible and were able to detect apples even though the picture taken-date was changed or picture from another cultivar was used.