Yes in general it is true, but any image classification problem mostly follow this pipeline: 

1. having the right dataset annotated correctly. For example, if you want to classify if a plant has certain disease or not, you should have enough examples that for that plant having that disease and enough examples for healthy ones. 

2. you need to pre-process the image such as converting to gray, reduce the size, historam equalization etc

3. you need to extract the features 

4. you need to pick the right machine leanring model to train it on the extracted features such as SVM

5.  you need to test your results by passing new samples that you didn't use in the training. 

 I think Deep Learning Packages do not require feature engineering for images classifications. Example of these packages are Keras with Theano as in Deep Learning with Python Book https://machinelearningmastery.com/deep-learning-with-python/