I think you can use the transfer learning schema, I means you first pretrain the netwrok in the other source task which training images data is ample, and then fix the networks parameters but change the last few layers, and use your 1000 images to re-train the last few layers, it is something like fine-tuning style.
在 2014年10月9日星期四UTC+8下午4时19分07秒,salem ameen写道: