Target detection - Small Target Detection depth study (reason detection difficult and Tricks)

Small target difficult to detect the reason

main reason

(1) small size of the target picture is relatively small, the common object model detection, the general basis for the neural network backbone (VGG series and series Resnet) has several down-sampling processing, resulting in only small target size substantially characteristic of FIG. pixel size single digits, led to the design of object detection classifier to classify the effect of small goals difference.

(2) small target in the original graph size is relatively small, common target detection model, the general basis neural network backbone (VGG series and Resnet series) has several down-sampling process, if the classification and regression operate under several layers after sampling wherein layer processing is performed, the target feature small receptive fields mapped back to the original target may be larger than the small size of the original image, resulting in poor detection.

other reasons

Less (1) a small number of targets in the original image is less detector feature extraction, resulting in small target differential detection effect.

(2) neural network is leading in learning goals, small goals are ignored in the whole learning process, leading to results in the detection of small effects goal difference.

Tricks

. (1) data-augmentation simple and crude, such as image amplification, detection using multi-scale image pyramid, the final drawback is that the detection result of the fusion complex operation, large computation, the actual situation is not practical.;

(2) a method wherein a fusion: FPN these, multiscale feature map predicted, feature stride less from the start;

(3) a suitable training methods: CVPR2018 and the SNIP SNIPER;

(4) to set a smaller denser anchor, anchor match strategy design, etc., reference S3FD;

(5) using a small object GAN amplified detected again, CVPR2018 such papers;

(6) using the contact context information, the context object and resume, such relation network;

(7) with dense cover and how to do better location and Classification, reference IoU loss, repulsion loss and so on.

(8) as far as possible the design of convolutional neural network employed in steps of 1, as many targets retention characteristics.