Plants are a major and important food source for the world's population. Smart and sustainable agriculture should be capable of providing precise and timely diagnosis of plant diseases. This helps in preventing financial and other resource loss to the farmers. Since plant diseases show visible symptoms which a plant pathologist will be able to diagnose through optical observations. But this process is slow and requires continuous monitoring as well as the availability and successful diagnostic capability of the pathologist. To overcome this, in smart agriculture, computer-aided plant disease diagnostic/detection model is used to help increased crop yield production. Common diseases are found in tomatoes, potatoes and pepper plants, some of them are bacterial spot, early blight etc. If a farmer can detect these diseases early, and can apply an appropriate treatment then it will improve crop yield and prevent the economic loss. In this work, we train the dataset on three different deep convolution neural network architecture and found the best suitable model to detect tomato leaf diseases. In order to avoid overfitting of the mode, batch normalization layer and a drop out layer has been included. The proposed Deep CNN is trained with various dropout values and a suitable dropout value is identified to regularize the model. The experimental methodology tested on plant village dataset showed improved accuracy of 96%, even without performing pre-processing steps like noise removal. By introducing batch normalization and dropout layer training accuracy improved to 99% whereas validation and testing accuracy is found to be 98%.