Maize is Ethiopian most dominant cereal crops both in terms of area coverage and level of production. There are different varieties of maize in Ethiopia. Maize varieties are classified based on different morphological features such as shape and size. The shape is the most important attribute among the features used. Due to the nature of maize seed and its rotation variant, it is difficult to identify maize seed varieties. Identification of maize seed varieties is difficult even for expert eyes due to their similar morphological features and visual similarities. It is labor-intensive and time-consuming to identify and differentiate between maize seeds. To solve this problem, the researchers proposed a deep learning approach with hybrid features. This study evaluates the different maize seed verity to help famers and experts to realize maize variety awareness and transform the current agricultural status in to modern production system. In this study, we have performed maize seed classification using deep learning algorithms and tolls. 400 image data were collected from Adet in the west Gojjam agriculture and research center using a smart phone. As a feature extractor convolutional neural network and hybrid feature model have been used. The hybrid feature consists of a convolutional neural network, Gabor and histogram of oriented gradients features. For classification, softmax and multi-class support vector machine classifiers were used. In this work a new method based on combined maize seed shape and texture features using SVM as a multi-class classifier for the identification of maize seed varieties has been proposed. Shape and texture features are extracted from the seed image using HOG and Gabor filter feature extraction techniques respectively. These two features are combined to form a feature set that trained the SVM classifier. An experiment on convolutional neural network feature with softmax classifier is conducted and shown 89% accuracy. The researcher has also tested HOG with SVM classifier and obtained accuracy of 88%, again CNN using SVM classifier and obtained 95% accuracy, Gabor feature with CNN feature obtained test accuracy of 89%. The proposed hybrid CNN and HOG features trained with SVM classifier is 99% accurate which is the highest accuracy obtained in this work.