Hey there,

Is the 23th the last layer (fully connected) of AlexNet? Are you sure you replace the last fully connect, which contains the number of categories as output, with the number of objects you wanna to recognize?

You can obtain more info and details about segmentation coupled with object detection in Mask R-CNN paper (https://arxiv.org/abs/1703.06870)

Dear Chaitali

This paper can be useful for your question. you should extract suitable features similar to what was done in this paper for each original texture image. (a few features and good classification results)

M. M. Feraidooni, D. Gharavian, "A new approach for rotation-invariant and noise-resistant texture analysis and classification", Machine Vision and Applications, 2017.

DOI 10.1007/s00138-017-0899-2

best regards.

if your objective is fine tuning the AlexNet you could approach Transfer Learning paradigm allowing you to add some new classes or extend the knowledge of a pre-trained CNN with respect to some new image domains. This is not a straightforward process as well, you need to balance the number of new sample images (the images of the new classes) to avoid and combat the overfitting.

You could follow the interesting practical article as it follows:

https://it.mathworks.com/help/nnet/examples/transfer-learning-using-alexnet.html

Thank you all. I will go through ur suggestions first and again get back to you.