Delta tributary network-An efficient alternate approach for bottleneck layers in CNN for plant disease classification

Currently, numerous research works have been proposed for diagnosing leaf diseases using state of the art convolutional neural networks. In this work, we propose a novel architecture called "Delta Tributary Network" that is built by stacking microarchitecture blocks called delta blocks specifically designed for leaf disease classification. These delta blocks utilize a novel channel split algorithm to reduce the number of channels given as input to 3 x 3 convolution layers. Unlike the existing bottleneck design which uses 1 x 1 convolution layers to decrease channel dimension space, Delta tributary network utilizes the a novel channel split algorithm to control the number of input channels togiven to 3 x 3 convolution layers, thereby preventing the linear stack of layers and henceforth avoiding over fitting and vanishing gradient problems. Delta tributary network when tested on plant village dataset gives an accuracy of 96% with just 0.3 million parameters on 133 MFLOP (Million Floating Point Operations) calculations. Further, delta tributary network tested on other bench mark datasets like CIFAR 10, CIFAR 100, MNIST and Fashion MNIST delivers higher accuracy than the other state of the art models with lesser trainable parameters, proving that delta blocks extract efficient potential features.